The invention relates to determining a change in an object or class of objects in image data, preferably remote sensing data; and, in particular, though not exclusively, to methods and systems for determining a change in an object or class of objects in image data, preferably remote sensing data and a computer program product enabling a computer system to perform such methods.

Background of the invention

Remote sensing data, such as satellite data and aerial image data, may be used for a wide variety of purposes, such as creating and updating maps, monitoring land cover and land use, water management, et cetera. Any monitored entity, e.g. a building, field, or road, may be considered an‘object’. For many purposes, detecting changes in such objects, e.g. new buildings, cut down trees, or additional lanes on a road, is especially relevant, as they may indicate a need for action, such as updating a map, or checking building permits or logging concessions. Detecting, categorising, and registering changes is typically an at least partially manual process.

However, the number of satellites and drones providing remote sensing data keeps growing, and they are equipped with increasingly powerful sensors, acquiring images at very high resolutions. This results in an increasing amount of (high-resolution) remote sensing data, necessitating automated tools for image processing. Consequently, automated change detection, i.e. automated detection of changes in geographical objects based on changes in image data, is a rapidly evolving field. Detection of an object in an image by an automated system may be referred to as‘object signal’, detection of a change between images may be referred to as a‘change signal’. An aim of such automated systems may be to provide a system that is at least comparable to a human regarding accuracy.

One of the difficulties in automated change detection is avoiding a high rate of false positives, which may lead to unneeded reactions. Change detection methods should be able to differentiate between image changes due to changes in objects of interest, and other image changes, for instance due to different circumstances, e.g. clouds, changes in illumination or shadows, vegetation changes, et cetera. Whether a change signal is a‘true’ signal or a‘false’ signal, may depend on the object of interest (e.g. changes in tree foliage may lead to a true change signal when studying trees) or changes in the interest of the object, but to a false change signal when studying the road under the trees. Similarly, weather applications may be interested in clouds, while clouds may be considered noise for applications interested in land use. Typically, there is a balance between specificity and sensitivity. For many applications, especially in the context of longer time series, specificity is more important than sensitivity (i.e. false positives are worse than false negatives).

An example of a change detection method for remote sensing data can be found in A. Song et ai,‘Change detection in hyperspectral images using recurrent 3D fully convolutional networks, Remote Sensing, Vol. 10, No. 11 (2018) art. 1827. Song et ai describe the use of an end-to-end trained Recurrent three-dimensional (3D) Fully

Convolutional Network (Re3FCN) for multitemporal data analysis. The input data are patches of hyperspectral remote sensing images, i.e. images that have pixels in rows and columns, where each pixel has an array of values for each of a large number of spectral bands (light sensed at different wavelengths). An input image may therefore be considered a 3D data set, with two spatial dimensions (rows and columns) and one spectral dimension.

The Re3FCN includes two main modules: a spectral-spatial module comprising 3D convolutional layers and a temporal module comprising a recurrent network with a single-layer Convolutional Long Short-Term Memory (ConvLSTM). The spectral- spatial module uses 3D convolutional layers to encode so-called spectral and spatial features, such as edges or textures. The spectral-spatial module does not have a‘memory’, i.e. each input is analysed independently of the previous inputs. The temporal module uses the output of the spectral-spatial module as input and models the temporal dependency of the spectral and spatial features in multitemporal images, i.e. performs the change detection proper. The output of the temporal module may be binary (unchanged or changed), or subdivided in a limited number of classes (unchanged, or type of change), based on the spectral and spatial features determined by the 3D convolutional layers.

However, the method of Song et al. also has various drawbacks. For example, the method does not discriminate well between relevant and irrelevant changes, and may therefore yield a high number of false positive change detections. Additionally, the method is sensitive to misclassification of pixels, and is not suitable for comparing images from different image sources (e.g. sensors operating at different wavelengths). Although reference is made to multitemporal images, the examples and embodiments in the text are limited to

comparisons of only two time instances. Change detection over longer time series, which may have different requirements regarding e.g. data interpretation or training, are not explicitly disclosed by Song et al.

There is therefore a need in the art for a method to reliably detect changes to (physical) objects in remote sensing data that removes, or at least reduces one or more of the preceding drawbacks associated with object-based and /or pixel-based change detection methods. Summary of the invention

As will be appreciated by one skilled in the art, aspects of the present invention may be embodied as a system, method or computer program product. Accordingly, aspects of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.) or an embodiment combining software and hardware aspects that may all generally be referred to herein as a "circuit," "module" or "system”. Functions described in this disclosure may be implemented as an algorithm executed by a microprocessor of a computer. Furthermore, aspects of the present invention may take the form of a computer program product embodied in one or more computer readable medium(s) having computer readable program code embodied, e.g., stored, thereon.

Any combination of one or more computer readable medium(s) may be utilized. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non- exhaustive list) of the computer readable storage medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

A computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber, cable, RF, etc., or any suitable combination of the foregoing. Computer program code for carrying out operations for aspects of the present invention may be written in any

combination of one or more programming languages, including a functional or an object oriented programming language such as Java(TM), Scala, C++, Python or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer, server or virtualized server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider).

Aspects of the present invention are described below with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor, in particular a microprocessor or central processing unit (CPU), or graphics processing unit (GPU), of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer, other programmable data processing apparatus, or other devices create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. For example, and without limitation, illustrative types of hardware logic components that may be used include Field-Programmable Gate Arrays (FPGAs), Application-specific Integrated Circuits (ASICs), Application-Specific Standard Products (ASSPs), System-on-a-Chip systems (SOCs), Complex Programmable Logic Devices (CPLDs), etc.

These computer program instructions may also be stored in a computer readable medium that can direct a computer, other programmable data processing apparatus, or other devices to function in a particular manner, such that the instructions stored in the computer readable medium produce an article of manufacture including instructions which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

It is an objective of the embodiments in this disclosure to provide a computer- implemented method for determining a changed object or class of objects in image data, preferably remote sensing data. In an embodiment, the method may comprise: receiving a first image data set of a geographical region associated with a first time instance and receiving a second image data set of the geographical region associated with a second time instance; determining a first object probability map on the basis of the first image data set and a second object probability map on the basis of the second image data set, a pixel in the first and second object probability maps having a pixel value, the pixel value representing a probability that the pixel is associated with the object or class of objects; providing the first object probability map and the second object probability map to an input of a neural network, preferably a recurrent neural network, the neural network being trained to determine a probability of a change in the object or class of objects, based on the pixel values in the first object probability map and in the second object probability map; receiving an output probability map from an output of the neural network, a pixel in the output probability map having a pixel value, the pixel value representing a probability of a change in the object or class of objects; and determining changes in the object or class of objects in the

geographical region, based on the output probability map.

It is a further objective of the embodiments in this disclosure to provide a computer-implemented method for determining changes in a plurality of objects or classes of objects in image data, preferably remote sensing data, wherein the method may comprise: receiving a first image data set of a geographical region associated with a first time instance and receiving a second image data set of the geographical region associated with a second time instance; determining one or more first object probability maps on the basis of the first image data set and one or more second object probability maps on the basis of the second image data set, wherein a pixel in each of the one or more first object probability maps and the one or more second object probability maps having a pixel value, the pixel value representing a probability that the pixel is associated with one of the plurality of objects or classes of objects; providing the one or more first object probability maps and the one or more second object probability maps to an input of a neural network, preferably a recurrent neural network, the neural network being trained to determine a probability of one or more changes in the plurality of objects or classes of objects, based on the pixel values in the one or more first and second object probability maps; receiving one or more output probability maps from an output of the neural network, a pixel in the one or more output probability maps having a pixel value, the pixel value representing a probability of a change in one of the plurality of objects or classes of objects; and, determining one or more changes in the plurality of objects or classes of objects in the geographical region, based on the one or more output probability maps.

The first and second object probability maps may be generated by dedicated object detectors, potentially based on neural networks, that are specialised in detecting the object or class of objects a user may be interested in and returning, for each pixel, the probability that it belongs to that object or class of objects. These object probability maps are provided as input to a neural network configured to determine a change in said object or class of objects. The use of dedicated object detectors results in high quality input to the actual change detector specific to the object or class of objects, reduces sensitivity to noise and/or pixel misclassification, and reduces the rate of false positives.

In general, methods for change detection can be divided into two broad categories: pixel based and object based. Pixel-based change detection seeks to determine whether a pixel value has changed meaningfully, and optionally, to classify a changed pixel. An advantage of pixel-based methods is that pixels are relatively easy to process (compared to objects of, potentially, arbitrary shape and size), and may e.g. be processed by advanced neural networks. Object-based change detection seeks to identify one or more objects in an image, and then determines changes in these identified objects. Object-based methods tend to be less sensitive to noise than pixel-based methods, but are dependent on the quality of the object detection.

The method according to this embodiment may be considered a‘hybrid’ change-detection method, combining advantages from both object-based and pixel-based change-detection methods. The object detectors provide the object-based element, reducing noise and providing a high specificity. The comparison of the object probability maps is essentially a pixel-based method, allowing to use the advantages of, preferably recurrent, convolutional deep neural networks.

Object detection in remote sensing data may be based on image segmentation. Image segmentation is, essentially, identifying groups of pixels that, in some sense, belong together. Typically, an image segmentation method determines a discrete label for each pixel in an image, such that pixels with the same label are grouped together. A label may be binary or multi-valued. Segmentation methods can be broadly divided into semantic segmentation methods, aiming to classify each pixel as belonging to one of several classes of objects or even individual objects, and non-semantic segmentation methods, aiming to create contiguous patches by either grouping pixels that are similar in some way, or by creating edges at pixel discontinuities; in other words, non-semantic segmentation is based only on pixel properties, while semantic segmentation is based on the objects the pixels represent.

A system that uses a general image segmentation method to create objects in an image is typically less accurate in detecting predetermined objects than a dedicated object detector providing object probability maps, as it is less tailored to that end and may be more easily confused, misclassifying a pixel when e.g. two classes receive similar scores. As in many cases, a user is primarily interested in detecting a change in one or a few pre determined (classes of) objects, a high specificity in a limited range of objects may be preferred. Additionally, an automated non-semantic segmentation-based system may be less suitable to detect heterogeneous objects; for example, a segmentation algorithm detecting objects by creating relatively homogeneous pixel groups might have trouble identifying a parking lot, where the parked cars may have a high contrast with the ground, as a single object.

An additional advantage of a method using dedicated object detectors is that such a method allows for comparison of dissimilar remote sensing data (e.g. from different sensor types), provided the object of interest can be detected in both data sets. This may require using different object detectors for each input type. For example, a typical change detection algorithm cannot reliably detect changes between images from sensors operating in different wave bands (e.g. infra-red versus visible light), because the pixel values are typically very different, but with the method according to the invention, a comparison is possible because both images are first converted to probabilities using dedicated tools, and the resulting object probability maps can be reliably and meaningfully compared.

A neural network may be trained to distinguish between meaningful and meaningless changes to an object, resulting in a reliable interpretation of a comparison between two object probability maps. In some cases, such an interpretation may depend on more information than only the object probability maps; e.g. a tree carrying leaves in one image data set and not in another may be a meaningful change depending on the state of other trees in the images, or on the time of the year and geographic location, if such information is provided or derivable from the image. Change detection networks may be trained specifically for each object in which a change is to be detected, in order to maximize performance. Output may be binary, e.g. limited to just‘changed’ or‘unchanged’ or a probability of a change, but may also comprise information on the type of change, e.g. an object appeared, disappeared, or remained but was otherwise changed.

In an embodiment, the neural network is a recurrent neural network and the method further comprises: receiving one or more additional image data sets of the geographical region, each additional image data set being associated with an additional time instance; for each of the one or more additional image data sets, determining an additional object probability map on the basis of the additional image data set, a pixel in the additional object probability map having a pixel value, the pixel value representing a probability that the pixel is associated with the object or class of objects; and providing the one or more additional object probability maps to an input of the neural network, wherein the first object probability map, the second object probability map, and the one or more additional object probability maps are provided in an order based on a time ordering of the time instances associated with the first, second, and one or more additional image data sets.

In many use cases, it is preferable to monitor a geographical region over a longer period of time, repeatedly detecting changes to objects or classes of objects. It may also be necessary to provide multiple images to detect a change in an object, for instance because the quality of the images was insufficient to determine a change, or because part of the image may have been obscured in one or more of the image data sets. Usually, such a plurality of input image data sets should be provided to the neural network in chronological order or reverse chronological order. In a typical embodiment, each time a new image data set is acquired, this image data set may be provided to an object detector, and the resulting object probability map may be provided to the recurrent neural network; this procedure automatically takes care of the time-ordering. By (preferably automatically) providing all acquired images to the network, there is no need for human oversight, e.g. to handpick cloudless images, as the method is capable of distinguishing between true and false change signals.

An advantage of using a recurrent neural network is that it is capable of receiving and processing an undetermined number of images, whereas non-recurrent neural networks usually require a pre-determined amount of images; typically two, but larger numbers are also possible. Additionally, recurrent neural networks may provide an additional output probability map for each additional input image data set, whereas non-recurrent networks typically provide a single output probability map for each collection of input image data sets.

In an embodiment, the method further comprises: receiving an output probability map for each time instance after the first time instance in the time ordered set; and determining changes in the object or class of objects in the geographical region, based on each of the received output probability maps.

Instead of only receiving an output probability map after a number of input images have been analysed, it may be advantageous to create an output probability map for each time step (possibly except the first) and analyse this output to determine whether a change has occurred. In other embodiments, detecting a change may only occur at a selection of time instances, possibly in dependence of characteristics of the input images.

In an embodiment, the neural network is a deep recurrent neural network comprising at least two layers and at least one of the layers comprises a convolutional long short-term memory, ConvLSTM, cell, and the method further comprises the step of initialising the neural network. This network architecture was found to give particular accurate results.

In an embodiment, the method further comprises: pre-processing the image data sets and/or the object probability maps, the pre-processing comprising one or more of: rotating, scaling, resampling, cropping, and padding.

In a typical embodiment, the neural network requires object probability maps with a fixed number of pixels in the horizontal and vertical dimensions. Consequently, it may be necessary to crop and/or pad the image in order to obtain the required number of pixels. Depending on the architecture, this step may occur before and/or after the object detection step. Additionally, the input object probability maps should typically cover the same area in the same orientation and at the same resolution, which may necessitate rotating, scaling, and/or resampling of one or more of the input images or of the object probability maps. In some embodiments, pre-processing may also scale and/or crop the pixel values of the image data sets and/or the object probability maps.

In an embodiment, the method further comprises: receiving non-image data associated with the object or class of objects; converting the non-image data to pixel data; and concatenating the one or more of the object probability maps with the pixel data; and wherein providing an object probability map to an input of a neural network comprises providing an object probability map concatenated with the pixel data to the input of the neural network.

In an embodiment, the method further comprises: receiving non-image data associated with the object or class of objects; converting the non-image data to pixel data; and concatenating the one or more output probability maps with the pixel data; and wherein determining changes in the object or class of objects in the geographical region comprises determining changes in the object or class of objects in the geographical region based on a convolution of the output probability map and the pixel data.

Using additional data, i.e. data not comprised in the image data sets or resulting object probability maps, may increase the accuracy of the neural network or of the interpretation of the output probability map(s). Examples of such data are cadastre data such as building outlines and/or building types (typically stored as vector data), and governmental (e.g. municipal) data such as zoning plans or permits (e.g. building permits or logging permits).

In an embodiment, determining an object probability map on the basis of an image data set comprises: determining an auxiliary probability map on the basis of the first image data set, a pixel in the auxiliary probability map having a pixel value, the pixel value representing a probability that the pixel is associated with an auxiliary object or an auxiliary class of objects; and determining the first object probability map on the basis of at least the auxiliary probability map. Using a plurality of object detectors in series may increase the accuracy and reduce the need for training and the memory footprint. For example, an object detector detecting dormers may benefit from output of an object detector detecting buildings. In this example, the dormer detector only needs to be trained with buildings, and does not need to take into account non-building parts of the image, which reduces training time and memory footprint. Similarly, a cloud detector may indicate which parts of an image may have a lower reliability because of the presence of clouds; this may help the interpretation of a changed pixel value. An additional advantage of using modularised object detectors is that it provides an interpretation of the results, whereas a single end-to-end neural network typically functions as a black box, giving little insight why e.g. the network detected an object or failed to detect an object.

In a further aspect, the invention may relate to a computer system adapted for determining a change in an object or class of objects in image data, preferably remote sensing data, comprising: a computer readable storage medium having computer readable program code embodied therewith, the program code including at least one trained 3D deep neural network, and at least one processor, preferably a microprocessor, coupled to the computer readable storage medium, wherein responsive to executing the computer readable program code, the at least one processor is configured to perform executable operations comprising: receiving a first image data set of a geographical region associated with a first time instance and receiving a second image data set of the geographical region associated with a second time instance; determining a first object probability map on the basis of the first image data set and a second object probability map on the basis of the second image data set, a pixel in the first and second object probability maps having a pixel value, the pixel value representing a probability that the pixel is associated with the object or class of objects; providing the first object probability map and the second object probability map to an input of a neural network, preferably a recurrent neural network, the neural network being trained to determine a probability of a change in the object or class of objects, based on the pixel values in the first object probability map and in the second object probability map; receiving an output probability map from an output of the neural network, a pixel in the output probability map having a pixel value, the pixel value representing a probability of a change in the object or class of objects; and, determining a change in the object or class of objects in the geographical region, based on the output probability map.

In an embodiment, the neural network may be a recurrent neural network, preferably a deep recurrent neural network. In an embodiment, the deep recurrent neural network may comprise at least two layers and at least one of the layers comprising a convolutional long short-term memory, ConvLSTM, cell.

In an embodiment, the executable operations may further comprise: receiving one or more additional image data sets of the geographical region, each additional image data set being associated with an additional time instance; for each of the one or more additional image data sets, determining an additional object probability map on the basis of the additional image data set, a pixel in the additional object probability map having a pixel value, the pixel value representing a probability that the pixel is associated with the object or class of objects; and, providing the one or more additional object probability maps to an input of the neural network, wherein the first object probability map, the second object probability map, and the one or more additional object probability maps are provided in an order based on a time ordering of the time instances associated with the first, second, and one or more additional image data sets.

In an embodiment, the executable operations may further comprise: receiving an output probability map for each time instance after the first time instance in the time ordered set; and, determining changes in the object or class of objects in the geographical region, based on each of the received output probability maps.

In an embodiemt, the executable operations may further comprise: receiving non-image data associated with the object or class of objects; converting the non-image data to pixel data; and, concatenating the one or more of the object probability maps with the pixel data; and wherein providing an object probability map to an input of a neural network comprises providing an object probability map concatenated with the pixel data to the input of the neural network.

In an embodiment, the executable operations may further comprise: receiving non-image data associated with the object or class of objects; converting the non-image data to pixel data; and, concatenating the one or more output probability maps with the pixel data; and wherein determining changes in the object or class of objects in the geographical region comprises determining changes in the object or class of objects in the geographical region based on a convolution of the output probability map and the pixel data.

In an embodiment, determining an object probability map on the basis of an image data set may comprise: determining an auxiliary probability map on the basis of the first image data set, a pixel in the auxiliary probability map having a pixel value, the pixel value representing a probability that the pixel is associated with an auxiliary object or an auxiliary class of objects; and, determining the first object probability map on the basis of at least the auxiliary probability map.

In a further aspect, the invention may also relate to a computer program product comprising software code portions configured for, when run in the memory of a computer, executing the method steps according to any of the process steps described above.

Brief description of the drawings

Fig. 1 depicts a system for reliable object-based change detection in remote sensing data according to an embodiment of the invention; Fig. 2 depicts a flow diagram for change detection using a (recurrent) neural network according to an embodiment of the invention;

Fig. 3 depicts data flow diagram for change detection over a plurality of time steps using a recurrent neural network according to an embodiment of the invention;

Fig. 4A and 4B depict a schematic view of a convolutional long short-term memory (ConvLSTM) cell as may be used in an embodiment of the invention;

Fig. 5A depicts a schematic example of modular object detection as may be used in an embodiment of the invention, while Fig. 5B depicts a flow diagram of modular object detection as may be used in an embodiment of the invention;

Fig. 6 depicts a schematic example of object-based change detection in a time series of images according to an embodiment of the invention,

Fig. 7 depicts a flow diagram for training a neural network for reliable object- based change detection in remote sensing data according to an embodiment of the invention; and

Fig. 8A and 8B depict flow diagrams for reliable object-based change detection in remote sensing data using additional object-related information according to an embodiment of the invention.

Fig. 9 is a block diagram illustrating an exemplary data processing system that may be used for executing methods and software products described in this application.

Detailed description

In this disclosure embodiments are described of methods and systems to determine a change in an object or class of objects based on image data, preferably remote sensing data. The methods and systems will be described hereunder in more detail. An objective of the embodiments described in this disclosure is to determine changes in pre determined objects or classes of objects in a geographical region.

Fig. 1 schematically depicts a system for reliable object-based change detection in remote sensing data according to an embodiment of the invention. When a new image 102, typically an aerial image or satellite image, is received by the image processing and storage system 100, the image may be georeferenced 104, i.e. the internal coordinates of the image may be related to a ground system of geographic coordinates. Georeferencing may be performed based on image metadata, information obtained from external providers such as Web Feature Service, and/or matching to images with known geographic

coordinates. The image may additionally be pre-processed by a pre-processor 106, e.g. the pixel values may be normalised to a predefined range, or the image may be sliced, cropped, and/or padded to a predefined size. The (optionally georeferenced and pre-processed) image may then be stored in an image storage 120. Alternatively or additionally, the raw image may be stored in an image storage. The image is subsequently provided to one or more object detectors H O1-3. The object detectors may operate on the image in series and/or in parallel. The object detectors may be implemented as neural networks, preferably convolutional neural networks, as analytic feature detectors, or in any other way or combination of methods. The object detectors may receive additional input from an object data storage 128, for example data from a municipal building database, (publicly) available GIS data, et cetera. The object detectors may receive additional input from stored images or stored object signals, i.e.

detected objects or detection probability information related to objects. The object detectors are described in more detail with reference to Fig. 5A,B. The object detectors may output one or more object probability maps 112, which may be stored in an object signal storage 122. Additionally or alternatively, a thresholded or otherwise processed image comprising detected objects may be stored. Preferably, each object probability map encodes the probability that a pixel or group of pixels belongs to a single object or class of objects. For example, a first object probability map may encode the probability of a pixel showing a solar panel and a second object probability map encoding the probability of a pixel showing a roof dormer. A pixel may be associated with one or more objects, e.g.‘building’, and‘flat roof top’ and‘solar panel (roof mounted)’.

If there are earlier images of the same geographical region in the image storage and/or object signal storage, the change detector 130 may be triggered. Depending on the resolution and size of the object probability map, and the amount of geographical overlap with an earlier, stored object probability map or change signal, the object probability map may be resampled, cropped, and/or padded or otherwise processed by a resampler 114; the earlier, stored object probability map or change probability map may be treated similarly by a resampler 116. The object probability map and one or more stored object signals and/or change signals are provided to the change detector 130. The change detector may also receive additional information from an external data storage 128, for instance municipal data regarding concessions, (changes to) zoning plans, or news reports. The change detector may be a conventional CNN or, preferably, a Recurrent CNN, and is described in more detail with reference to Figs. 2-4. The change detector outputs a change probability map 132, which may be stored in change signal storage 124. Based on this change probability map, a change map 140 may be determined, e.g. by thresholding.

Optionally, the change map may comprise additional information from the input image data or external sources. For example, the change map may be overlayed on the input image to visually show the changed objects, and/or the change map may be combined with a map of expected changes based on e.g. requested concessions. In some embodiments, a binary change map (i.e. representing changed / unchanged) may be combined with e.g. input data or object probability maps to determine a (possibly multi-class) type of change.

Fig. 2 depicts a flow diagram for reliable object-based change detection in remote sensing data using a neural network according to an embodiment of the invention. In an embodiment, the neural network may be a recurrent neural network, preferably a deep recurrent neural network (deep RNN), At a first time instance t = to, a first image data set 202 of a geographical region is obtained and provided to the change detection system. As was described before with reference to Fig. 1 , the first image data set may be georeferenced and pre-processed as needed. Subsequently, the first image data set is provided to an object detector 204 or ensemble of object detectors, that determine at least a first object probability map 206. The first object probability map may comprise pixels, a pixel having a pixel value representing the probability that the pixel is associated with the object or class of objects the object detector is intended to detect. The object probability map may optionally be resampled, leading to a resampled object probability map 208, which should be in a format that can be provided to the neural network 220, which is preferably a recurrent neural network.

At a second time instance t = , different from the first time instance, a second image data set 212 of the same geographical region is obtained and provided to the change detection system. The geographical region depicted in the second image data set may fully or partially overlap the geographical region depicted in the first image data set. The second image data set may be similar to the first image data set, e.g. a satellite image acquired with the same satellite as the first image data set, so that both image data sets have the same resolution and the same colour channels. Alternatively, the image data sets may be (very) different, e.g. one image data set may be acquired with a satellite using red and infrared colour channels, while the other image data set is acquired with a drone using only visible light, resulting in possibly different resolutions, colour channels, and pixel encodings

(including data format, e.g. floats or unsigned shorts, and data ordering, e.g. numbering rows from the top or the bottom). The geographical regions of the first and second image data sets should at least partially overlap. For example, the second image data set may cover the same geographical region as the first image data set, or only a part thereof, and vice versa.

The second image data set may be georeferenced and pre-processed as needed. Subsequently, the second image data set is provided to an object detector 214 or ensemble of object detectors. The object detector 214 may be the same as the object detector 204, or a different object detector. When the first image data set and the second image data set have the same or a similar data source, it may be preferable to use the same object detector, whereas when the first and second image data sets have different data sources, it may be preferable to use different object detectors, each specialised for detecting objects in a different data source. The object detector(s) 214 may determine at least a second object probability map 216. The second object probability map may comprise pixels, a pixel having a pixel value representing the probability that the pixel is associated with the object or class of objects. The object probability map may optionally be resampled, leading to a resampled object probability map 218, which should be in a format that can be provided to the neural network 220. This may result in first and second resampled object probability maps 208,218 that have the same size in pixels, the same resolution, cover the same geographical area, and have the same pixel encoding. In some embodiments, there may be no resampling step. In other embodiments, resampling is only done as part of a pre processing step prior to the object detection step.

The neural network 220 subsequently receives the, optionally resampled, object probability maps and determines as output, based on the pixel values in the first probability map and in the second probability map, a change probability map 222, in which a pixel represents a probability of a change in the object or class of objects. Based on the change probability map 222, changed objects 224 may be determined, e.g. objects that appeared, disappeared, grew, shrunk, or otherwise changed.

Fig. 3 depicts a data flow diagram for change detection over a plurality of time steps using a recurrent neural network according to an embodiment of the invention. A deep recurrent neural network (RNN) 300 may have an input, an internal state, and an output. The blocks 301 o- _n refer to one block of convolutional layers, with the same trained parameters, but different (time step dependent) stored values in the memory cell. A number of input data sets may consecutively be provided to the input. The internal state and the output may depend on the input and previous internal states. The internal state may act as a sort of memory. Before a time instance t = to, the RNN may be initialised with initialisation data 312, typically all zeroes, but other values are also possible. At a time instance t = to, a first object probability map 314o is provided to the input of the RNN (step 302), each pixel in the first object probability map representing the probability that a pixel is associated with a predefined object or class of objects in a geographic region. As shown in the figure, in a first layer, the object probability map may be convoluted with a number ci of convolution masks, resulting in Ci feature maps. These feature maps may be combined to one feature map, which is passed to the next layer (step 306i) and a block of layers associated with the next time instance (step 308i). In the embodiment depicted in Fig. 3, the input object probability map has 512 ^c 512 pixels, but other embodiments may use input maps of different sizes.

In the embodiment depicted in Fig. 3, depicts a so-called deep CRNN wherein the block includes a stack of Convolutional Long Short-Term Memory (ConvLSTM) cells, which are described in more detail with respect to Fig. 4. As shown in Fig. 3, each ConvSTM cell may be configured to receive the hidden state h _t of a previous layer as input x _t in the current layer as denoted by the arrow 306i. An example of a deep non-convolutional RNN using LSTM cells is given by Graves in the article‘Generating sequences with recurrent neural networks’, arXiv:1308.0850v5 [cs.NE] (5 Jun. 2014). Apart from the use of non- convolutional LSTM cells, the example given by Graves differs from the example depicted in Fig. 3 in that in Graves, there are additional direct connections from the input to each layer, and from each layer to the output. Other embodiments of the invention may also comprise one or more of such additional direct connections. Some embodiments may use a RNN with a single layer, i.e. a so-called shallow CRNN. Other embodiments may have m, m > 1 layers. In such an embodiment, each layer / may have its own set of Q feature maps, corresponding to Q convolutions with Q convolution masks. An advantage of such a so-called stacked deep RNN is that each layer may operate on different time scales. Another advantage is that lower layers may have a greater spatial awareness, as is explained with reference to Fig. 4B.

At a time instance t = ti, a second object probability map 314i , associated with the same geographic region is provided to the RNN 300. This input is processed by the same RNN, which compares the received input data with the data stored in the internal and/or hidden states, and determines new internal and hidden states. The RNN repeats this step for each layer, combining information from the previous time instance at the same depth and the immediately superior layer at the same time instance. The output of the last layer results in a change probability map 316i . In some embodiments there may be one or more additional layers between the last ConvLSTM layer and the output, e.g. a convolutional layer that reduces the c _m layers to a single layer. This process may be repeated during n time instances. In an embodiment, the network could be trained to output at each time instance the cumulative detected changes in the inputted object probability maps with respect to the first object probability map of t = to. In a different embodiment, the network could be trained to output a detected change only at its first occurrence.

Fig. 4A depicts a schematic view of a convolutional long short-term memory (ConvLSTM) cell as may be used in an embodiment of the invention. A Long Short-Term Memory (LSTM) is a type of Recurrent Neural Network (RNN) that is particularly useful for analysing time series through a‘recurrent hidden state’ that acts as a memory for earlier input. A typical application for RNNs is video analysis, where knowledge of previous video frames may help analysing a subsequent video frame, as sequential video frames are usually similar to each other. An example of an, optionally multi-layer, non-convolutional LSTM is given by A. Graves,‘Generating sequences with recurrent neural networks’, arXiv:1308.0850v5 [cs.NE] (5 Jun. 2014). Graves does not refer to change detection. In a ConvLSTM, the matrix multiplications of the weights and inputs in a non-convolutional LSTM are replaced by convolutions, increasing the spatial awareness of the network. Such a ConvLSTM is described in e.g. Xingjian Shi et ai,‘Convolutional LSTM network: a machine learning approach for precipitation nowcasting’, Adv. Neural Inf. Process. Syst. 2015:1 (2015) pp. 802-810. Xingjian Shi et ai describe the application of a stacked multi-layer ConvLSTM to encode and predict, based on a consecutive time series of (RADAR echo) precipitation maps of an area, the precipitation maps (and hence precipitation) of the following time frames. They do not describe (object-based) change detection as such.

In an LSTM cell, the data flow is controlled by so-called gates. An LSTM cell comprises an input gate /, a forget gate f, and an output gate o, which control the data in a memory cell c, and a hidden state h. The forget gate determines to what extent the old memory value is retained, the input gate determines to what extent the input and the old hidden state are stored in the memory, and the output gate determines to what extent the memory value is output as a new hidden state. The gates typically have continuous values between 0 and 1.

A typical ConvLSTM cell may be described by the following equations:

where * is a convolution operator, juxtaposition denotes matrix multiplication, o is a sigmoid function, and Q is a tanh-like function; is an input gate, is a forget gate, c _t is a memory cell, O _t is an output gate, h _t is a hidden state, and x _t is the input data in the first layer, respectively the hidden state of the immediately higher layer in subsequent layers; t is a time label, b. are bias vectors and IA¾ are weight matrices, where the indices have the obvious meaning; for example, b is the bias of the input gate and IA is the hidden state-input gate weight matrix.

In this implementation, the third equation, defining c _t is theoretically unbounded. However, in practice the plus sign between the first and second terms reduces the rate of change and limits vanishing and/or exploding gradient problems, whilst the forget gate also tends to prevent unlimited grow. Nevertheless, in some embodiments the value of c _t may be limited between predefined boundaries. Before the first input is offered, Co and ho are initialised with initialisation data, which may consist of all zeroes.

A ConvLSTM cell 400 in a first layer of a RNN may combine inputs, in this case four inputs, at its input gate. These inputs may include input data x _t 402, the hidden state of the previous time step h ¹ _t~i 404, the value stored in the memory cell at the previous time step c ¹ _t -i 410, and a bias input (not shown). The input data and the hidden state are convolved with a weight matrix, while the memory cell is weighed with a (pointwise) matrix multiplication or Hadamard product. The forget gate uses the same inputs, but with independently optimised weights. The memory cell may be updated, wherein the forget gate determines to what extent the stored memory remains, and the input gate determines to what extent the input data and the hidden state of the previous time step are stored. The memory cell may e.g. retain information from previous probability maps that has not been replaced by new information. In the case of e.g. video analysis, the forget gate may erase the contents of the memory cell when a new scene is detected, and knowledge of previous frames that is not relevant for understanding a current frame.

The value of the output gate may depend on the input data x _t , the previous hidden state h ¹ _t~i , and the current (i.e. updated) value of the memory cell c _t (in contrast to the input and forget gates, which use the value of the memory cell associated with the previous timestep). The output gate controls the value of the output hidden state h ¹ _t 408. In some embodiments, the RNN may have more than one layer. In some of such embodiments, a ConvLSTM in a layer m, m > 1 , may receive the hidden state h ^m~1 of the preceding layer instead of the input data x. In other embodiments, a ConvLSTM in a layer m, m > 1 , may receive both the input data x and the hidden state h ^m~1 of the preceding layer as inputs.

Fig. 4B schematically depicts a number of convolutional operations, in this case 3 x 3 convolutions. The aim of the convolutions is to encode spatial information from a previous time step in the current time step or from a higher layer in the current layer. The value of a (dark grey) pixel 450 _I-4 in a feature map depends the 3 ^c 3 pixel area surrounding the pixel in the same location in the previous time step, respectively in the input data or preceding layer, with potentially a different weight for each pixel. For example, the pixel value 450i may depend on pixel values 452i (in this case nine pixel values) in the input data x _t and on pixel values 454i (in this case nine pixel values) from the hidden state of the first network layer at the previous time instance h ¹ _t -i. Other embodiments may use different convolution schemes, e.g. 5 ^c 5, or dilated 3 ^c 3, or any other convolution scheme. Performing a plurality of convolutions on subsequent layers may increase the spatial awareness of the deeper layers.

In some embodiments, multiple convolutions, each with different weights and possibly different sizes, may be performed, resulting in a plurality of feature maps. These may be combined to detect more complex objects. The input of the previous time step may also be convolved with one or more weight matrices, in a similar manner. This results in a combination of temporal and spatial information, and may hence detect e.g. moving or growing / shrinking objects.

Fig. 5A depicts a schematic example of determining an object probability map based on a remote sensing image of a geographical region. Image 500 depicts two houses 506 _I-2 , two trees 504 _I-2 , a fence 505, and a single lane road 508i that is under construction, being expanded with a second lane 508 ₂ . Part of the image 500 has been occluded by a cloud 502, obscuring at least part of one house and part of the road. Another part of the road is covered by a tree. The input image 500 may be pre-processed, which may include e.g. rescaling, resampling, cropping, and/or padding; in the depicted embodiment, the input image is rescaled to half its original size for further processing.

In a first step, the input image 500 may be provided to a first object detector, in this example a cloud detector. The cloud detector may output a cloud probability map 510; in this example a white colour represents a low probability that a pixel is associated with a cloud, and a dark grey or black colour represents a high probability that a pixel is associated with a cloud. In this example, the cloud detector has correctly assigned a high probability to the pixels 512 depicting the cloud. In a different embodiment the first object detector may be a different object detector or a group of detectors, e.g. a cloud shadow detector, a haze detector, and/or a snow detector. In some embodiments, the cloud probability map 510 may be used to determine the pixels that are associated with the cloud, e.g. to segment the cloud; this may e.g. be done by thresholding the cloud probability map, associating all pixels with a pixel value above the threshold value with a cloud label, and associating all pixels with a pixel value below the threshold value with a non-cloud label.

In a second step, the input image 500 and the cloud probability map 510 may both be provided to a plurality of second object detectors, in this example a building detector, a tree detector, and a road detector. The building detector may output a building probability map 520i, the tree detector may output a tree probability map 520 _å , and the road detector may output a road probability map 520 ₃ . In these probability maps 520I_ ₃ , dark colours again denote a high probability, and light colours a low probability. The building detector may detect the buildings 506I, ₂ , and consequently assign a high probability to the corresponding pixel regions 526I, ₂ . Similarly, the tree detector may assign a high probability to the pixel regions 524i _,å corresponding to the detected trees 504i _,å , and the road detector may assign a high probability to the pixels 528i corresponding to the single-lane road 508i.

As part of building 526 ₂ is covered by the cloud, the building detector may assign a lower probability to the cloud-covered region 526 ₃ than to the not cloud-covered part of the building. Nevertheless, based on the parts of the building that are visible, the building detector may be trained to infer the cloud-covered part. Consequently, the cloud-covered part of the building 526 ₃ may be assigned a probability that is lower than that of the visible parts, but still relatively high. Similarly, the road detector may assign a lower, but still high probability to the part of the road covered by the cloud 528 ₃ , and also to the part of the road covered by one of the trees 528 ₄ .

In the part of the image that is covered by the cloud 522I_3, the building, tree, and road detectors may be trained to assign an intermediate probability, as the building, tree, and road detectors have insufficient information to determine whether or not, respectively, a building, tree, or road is present in the cloud-covered region apart from the parts belonging to a building or road segment that is partly visible. The road detector may also assign an intermediate probability to the new part of the road that is being constructed 528 ₂ ; for example, it may exhibit certain properties of a road, such as shape, and it being immediately adjacent to another part of a road, but not other properties, such a surface material.

In a third step, the input image 500, the cloud probability map 510, and the building probability map 520i may be provided to a third object detector, in this example a chimney detector. The chimney detector may output a chimney probability map 530. Similarly to the previously discussed examples, the chimney detector may assign a high probability to the pixels corresponding to the detected chimneys 536I, ₂ . An advantage of first detecting building and subsequently detecting chimneys, which may assumed to be placed on top of buildings, is an increased precision of the chimney detector. For example, first detecting the building may help to differentiate between the chimneys and the fence posts 505. In other embodiments, different object detectors may be used, and/or the object detectors may be ordered in a different way. For example, in some embodiments it may be advantageous to provide the input image 500 and the cloud probability map 510 first to the tree detector, and subsequently provide the input image 500, the cloud probability map 510, and the tree probability map 5202 to the road detector. This may help the road detector in deducing road parts that are covered by a tree, and increase the probability assigned to e.g. patch 5284.

In a typical embodiment, the object detectors may be based on deep neural networks, preferably convolutional deep neural networks. Other objects may be detected using more conventional image-analysis methods, such as analytical edge detection or feature extraction methods. Object detectors that are based on a neural network, may all be trained and optimised independently. This may greatly reduce the number of examples required for training. In the example of Fig. 5A, the chimney detector only needs to learn that chimneys are placed on top of buildings, and then be provided with images of buildings with and without chimneys. Training time and effort is greatly reduced, as the network does not need to learn to differentiate between chimneys and chimney-like objects that are not placed on buildings, such as the fence posts in Fig. 5A. An object detection system combining a plurality of object detectors that may be combined in parallel and/or in series is also known as a modular change detection system.

Fig. 5B depicts a flow diagram of modular object detection as may be used in an embodiment of the invention. In general, a modular object detection system may comprise a plurality of object detectors, each object detector configured to provide as output an object probability map, and configured to receive as input an input image and zero or more object probability maps. The modular object detection system may be run on one or more computers, typically on a server system. In this example, an input image 550 is provided to a first object detector, cloud detector 552. The cloud detector provides as output a cloud probability map 554, which may be stored in a memory of a computer. The input image 550 is also provided to a cloud shadow detector 556 which is, in this example, independent from the cloud detector 552. The cloud shadow detector may be activated prior to the cloud detector, after the cloud detector, and/or concurrent with the cloud detector. The cloud shadow detector provides as output a cloud shadow probability map 558.

Subsequently, the input image 550 is provided to an input of a building detector 560, together with the cloud probability map 554 and the cloud shadow probability map 558. The building detector provides as output a building probability map 562. As was explained before, adding extra information as provided by the object probability maps may increase the sensitivity and/or specificity of an object detector, in this case the building detector. In other embodiments, the building detector might use more, less, or different object probability maps as input. In some embodiments, one or more of the object probability maps may be used as input for a plurality of object detectors, e.g. in the embodiment depicted in Fig. 5A, the cloud probability maps are used as input for a building detector, a tree detector, and a road detector.

Next, the input image 550 is provided to an input of a chimney detector 564, together with the cloud probability map 554, the cloud shadow probability map 558, and the building probability map 562. The chimney detector provides as output a chimney probability map 566. In other embodiments, the chimney detector might use different inputs, e.g. only the input image and the building probability map, or the input image, the cloud probability map and the building probability map.

In other embodiments, other object detectors may be used, and they may be connected differently. Any of the object detectors may be implemented as a neural network, e.g. a deep convolutional neural network, and/or analytic or algebraic image analysis software. In a typical embodiment, each object probability map is stored in a database and may be reused for any number of other object detectors and/or change detectors.

Fig. 6 depicts a schematic example of object-based change detection in a time series of images. Images 602o- _å of a geographical region, are acquired at time instances and fe, respectively. As the image at time instance t = to is the earliest available image, image 602o may be considered the reference image, depicting three houses, three trees, and a single lane road. In image 602i, part of the image has been occluded by a cloud, obscuring one house and one tree, and part of the road. One of the houses has now a chimney, and there are roadworks underway. The object detection steps 630 at time instance t = ti were discussed in more detail with reference to Fig. 5A; the object detection steps at time instances t = to and t = t _å are the same, in this embodiment, but may lead to different object probability maps, as discussed in more detail hereunder. In image 602 _å , the cloud is gone, the road works have finished resulting in a two-lane road, and one of the trees that was previously occluded by the cloud, has been felled and replaced by a house.

In this embodiment, the images are pre-processed, which comprises rescaling the image to half its original size; other embodiments may comprise, more, less, or different pre-processing steps. The rescaled images are then provided to a group of object detectors. The first object detector detects clouds, resulting in cloud presence information. This information may be used by subsequent object detectors to determine whether a pixel is likely to be associated with, respectively, a building, a tree, or a road. At time instance t = to, the cloud detector does not detect a cloud in input image 602o. The building detector detects three houses, the tree detector detects three trees, and the road detector detects one single lane road segment. Part of this road segment is covered by a tree, but its presence is assumed based on the detector’s knowledge of road segments. In this embodiment, the output of the building detector is provided to a chimney detector, which detects one chimney on a building. Having such prior information, training of the chimney detector can be greatly reduced, as the network does not need to learn to discriminate chimney-like objects that are not on top of buildings, such as e.g. fence poles. At time instance f = ft, the cloud detector detects a cloud. Subsequently, the building detector detects one complete building (black), and one partially obscured building (black); based on the detector’s knowledge of buildings, it may guess at the remainder of the partly obscured building (dark grey). The building detector also detects a number of pixels that probably do not belong to a building (white). The building detector has insufficient information to determine whether or not the pixels where a cloud was detected, belong to a building and ascribes them an intermediate probability (shown in grey). The chimney detector detects two chimneys, on two buildings. The tree detector detects two trees, and the road detector detects a road segment, part of which it is unsure about (shown in grey). These steps have been discussed in more detail with reference to Fig. 5A. Subsequently, the chimney probability map 612o, based on the image data set 602o acquired at t = to and the chimney probability map 612i, based on the image data set 602i acquired at f = ft, are provided to the chimney change detection network, which detects one changed (appeared) chimney. Tree probability maps 608o and 608i are provided to the input of a tree change detection network, which does not detect a change. Road probability maps 61 Oo and 61 (ft are provided to the input of a road change detection network, which detects that the road may have changed, but is insufficiently sure. Finally, the detected changes may be highlighted in the input image, resulting in image 620i.

At time instance t = fe, the cloud detector does not detect a cloud. The building detector detects four buildings, but still only two chimneys are detected by the chimney detector. Consequently, the chimney change detector does not detect any change, relative to the latest known chimneys. The tree detector again detects two trees. The tree change detection network does not have sufficient information to be sure this is a change with respect to the situation at time instance t = ft, but the memory property of the recurrent network results in the tree change detection network deciding that there has at least been a change (disappearance) in trees with respect to time instance t = to. The road detector detects a road segment (still partially obscured by a tree), but now it is a two-lane road. While both the change from to to ft and from ft to fe may not be clear enough to determine with certainty that there has been a change, the memory property of the network may help to determine that the overall change from to to fe was large enough to detect a positive change. Finally, the detected changes may be highlighted in the input image, resulting in image 620 ₂ . Note that the appearance of a building has, in this example, not been detected, because no building change detection network was employed.

In some embodiments, more than one kind of object probability map may be provided to the change detection network; the change detection network may e.g. use both the cloud probability map and the tree probability map as inputs, to detect a change in trees. Such a configuration is especially advantageous when dealing with features that are preferably detected at multiple scale levels. For example, cloudiness and cloud opacity is best judged on a satellite image at large scales. Outlines of clouds and tracks of haze can be distinguished at a scale of hundreds of meters, while detection of individual trees may require a scale of two orders of magnitude lower. A model detecting disappeared trees without knowing where haze is present might incorrectly interpret a small flake of haziness as the absence of a tree, while a model that includes knowledge of haze will not report a

disappeared tree here as it knows it lacks the required visibility at this location; provided the model is properly trained, allowing false negatives under cloudy conditions.

In an experiment performing change detection using a Convolutional

Recurrent Neural Network with a time series of 5 images, inclusion of an additional input layer with the probability for cloud presence (in this experiment output from a separate Neural Network for cloud detection) improved the Jaccard index (a measure of similarity between label and model prediction) on a validation set for tree detection from 90.0% to 93.5%. For the specific case of cloud presence, the benefit may greatly depend on cloudiness of the images in question.

Fig. 7 depicts a flow diagram for training a neural network for reliable object- based change detection in remote sensing data according to an embodiment of the invention. The diagram depicts a deep recurrent neural network (RNN) 702, that is initialised with initialisation data 712. It should be noted that while the RNN 702 is drawn three times in this diagram, it is thrice the same CRNN, with a single set of internal parameters, that receives a plurality different object probability maps 714o- _n at a plurality of time instances t = to, ti, ... , t _n . The amount of time instances offered during training, may influence the maximum amount of time information that may be stored in the network’s memory.

In order to train the change detection network, the network may comprise a training module. The system may be provided with training data and associated target data. The training data may include a plurality of probability maps 714o- _n of the object or class of objects in which a change is to be detected, while the target data may comprise at least one ground truth map preferably one ground truth map for each time instance except the first. A ground truth map may comprise information on actual (relevant) changes, as may be determined on the ground. The ground truth map is preferably a binary map, wherein each pixel may have a first value, e.g. 1 , indicating that a change in the object or class of objects has occurred since the previous time instance, and a second value, e.g. 0, that no such change has occurred. Alternatively, the ground truth maps may indicate the cumulative changes with respect to the first or reference time instance t = to.

During training, the change detection network may predict 706 one or more change probability maps, i.e. the recurrent neural network may predict for each pixel of a training data set a probability of a change in the object or class of objects. During training, the target data may be provided to a different input of the neural network. The internal parameters of the change detection network may then be optimised by minimising a loss function related to the difference between the predicted change probability map and the ground truth map associated with that time instance. Such a function may be related to the sum over all pixels of the absolute value of the difference per pixel between the predicted change probability map and the ground truth map. In other embodiments, false positive and/or false negatives may additionally be penalised. Information about the error may be backpropagated 708 through the network, e.g. using a gradient descent method.

The accuracy of the trained network may depend on e.g. the training time, training sample variance, and amount of training examples. Additional training samples may be created from existing training samples by e.g. shifting and/or rotating data, or by applying noise to the data.

The change detection network may need to be trained separately for each type of object in which a change is to be detected. In an embodiment, the training data and target data may be enhanced by adjusting, e.g. rotating and/or shifting, the probability maps from one moment onwards in the time series. This way, a ground truth change map for the adjustments (rotations and/or shifts) for different time instances in the time series can be generated. These synthetic changes may enhance the training data and improve the training process of the neural networks.

Fig. 8A and 8B depict flow diagrams for reliable object-based change detection in remote sensing data using additional object-related information according to an embodiment of the invention. Additional object-related information may be any kind of data related to the object or class of objects other than the pixel values from remote sensing images. The additional object-related information may also be referred to as non-image data. A typical example of such additional object-related information is Geographic Information System (GIS) data, such as infrastructure databases that are used for maps. Such information may be used to increase the quality of object detection. Using such information for object detection has been described by e.g. Weija Li et ai,‘Semantic segmentation- based building footprint extraction using very high-resolution satellite image and multi-source GIS-data’, Remote sensing, Vol. 11 , No. 4 (2019) art. 403. Weija Li et ai describe a method based on a U-Net (a type of deep convolutional neural network developed for binary segmentation) for segmentation of buildings in satellite data integrated with GIS map data. This method applies a binary segmentation to the satellite data, i.e. it identifies pixel regions according to whether or not they belong to a specific object type, i.c. buildings. Using GIS data increases the accuracy of the building detection. However, Weija Li et ai. do not refer to change detection.

Changes to the environment, especially by private parties, often require a permit, such as planning permission or a building or logging permit. These permits and request for permits are often stored in databases. Other changes may be (publicly) notified, such as road construction works. Such information may be used to increase the quality of change detection: for example, the information that a building permit has been granted at a certain address may increase the probability of detecting a meaningful change. Using such information may require different steps; for example, if a permit only refers to an address, a cadastral database may be used to obtain an outline of the affected building or parcel, typically in vector format, which may need to be converted to pixel format before being provided to the change detection network.

In an embodiment, in a time step, an image data set 802 is provided to an object detector 804 or ensemble of object detectors, that determine an object probability map 806. The object probability map may optionally be resampled, leading to a resampled object probability map 808. Other object information 810, typically non-image data associated with the object or class of objects or with changes in or to the object or class of objects may be obtained from e.g. a database or a different source. Examples of such data are cadastre data such as building outlines and/or building types (typically stored as vector data), and governmental (e.g. municipal) data such as zoning plans or permits (e.g. building permits or logging permits). In some embodiments, the object information may be converted to pixel data 812. For example, a building permit may be linked to an address, and the address may be linked to an outline of a building or plot in cadastre data; in such an example, a pixel map may be created with preferably the same size and resolution as the resampled probability map, and the pixels within the outline defined by the cadastre data may be given a first value, associated with a requested building permit, and the pixels outside the outline may be given a second value.

In the embodiment depicted in Fig. 8A, the resampled object probability map and the pixelized object information are concatenated 814, and provided as a single input to the neural network 820, which is preferably a recurrent neural network. The neural network may then determine changes in the object or class of objects in the geographical region based on the joint resampled object probability map and the pixelized object information, by comparing these data to similar data obtained at a different time instance (not shown in this figure).

In the embodiment depicted in Fig. 8B, the same steps of acquiring an image data set 852, providing the image data set to an object detector 854 or ensemble of object detectors, determining an object probability map 856, and optionally resampling the object probability map are taken, leading to a resampled object probability map 858. Similarly, object information 860 is converted to pixel data, resulting in pixelized object information 862. In this embodiment, however, the pixelized object information is not provided to the neural network 870, but is concatenated 864 with the output of the change detection neural network. This concatenated data set is provided to a convolution operation 866, in order to obtain a change probability map. An advantage of determining changes in the object or class of objects in the geographical region based on a convolution of a concatenation of the output probability map and the pixelized object information is that it can be added to an existing change detection network without having to retrain the change detection network.

Additionally, it may reduce the memory footprint of the change detection network, compared to the change detection network of Fig. 8A, and may require less training time. On the other hand, the change detection network from Fig. 8A may be expected to be more accurate, as the additional data can be remembered by the network as needed, and can be included in the optimisation.

Fig. 9 is a block diagram illustrating an exemplary data processing system that may be used in embodiments as described in this disclosure. Data processing system 900 may include at least one processor 902 coupled to memory elements 904 through a system bus 906. As such, the data processing system may store program code within memory elements 904. Furthermore, processor 902 may execute the program code accessed from memory elements 904 via system bus 906. In one aspect, data processing system may be implemented as a computer that is suitable for storing and/or executing program code. It should be appreciated, however, that data processing system 900 may be implemented in the form of any system including a processor and memory that is capable of performing the functions described within this specification.

Memory elements 904 may include one or more physical memory devices such as, for example, local memory 908 and one or more bulk storage devices 910. Local memory may refer to random access memory or other non-persistent memory device(s) generally used during actual execution of the program code. A bulk storage device may be implemented as a hard drive or other persistent data storage device. The processing system 900 may also include one or more cache memories (not shown) that provide temporary storage of at least some program code in order to reduce the number of times program code must be retrieved from bulk storage device 910 during execution.

Input/output (I/O) devices depicted as input device 912 and output device 914 optionally can be coupled to the data processing system. Examples of input device may include, but are not limited to, for example, a keyboard, a pointing device such as a mouse, or the like. Examples of output device may include, but are not limited to, for example, a monitor or display, speakers, or the like. Input device and/or output device may be coupled to data processing system either directly or through intervening I/O controllers. A network adapter 916 may also be coupled to data processing system to enable it to become coupled to other systems, computer systems, remote network devices, and/or remote storage devices through intervening private or public networks. The network adapter may comprise a data receiver for receiving data that is transmitted by said systems, devices and/or networks to said data and a data transmitter for transmitting data to said systems, devices and/or networks. Modems, cable modems, and Ethernet cards are examples of different types of network adapter that may be used with data processing system 900.

As pictured in Fig. 9, memory elements 904 may store an application 918. It should be appreciated that data processing system 900 may further execute an operating system (not shown) that can facilitate execution of the application. Application, being implemented in the form of executable program code, can be executed by data processing system 900, e.g., by processor 902. Responsive to executing application, data processing system may be configured to perform one or more operations to be described herein in further detail.

In one aspect, for example, data processing system 900 may represent a client data processing system. In that case, application 918 may represent a client application that, when executed, configures data processing system 900 to perform the various functions described herein with reference to a "client". Examples of a client can include, but are not limited to, a personal computer, a portable computer, a mobile phone, or the like. In another aspect, data processing system may represent a server. For example, data processing system may represent a server, a cloud server or a system of (cloud) servers.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.

The corresponding structures, materials, acts, and equivalents of all means or step plus function elements in the claims below are intended to include any structure, material, or act for performing the function in combination with other claimed elements as specifically claimed. The description of the present invention has been presented for purposes of illustration and description, but is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the invention. The embodiment was chosen and described in order to best explain the principles of the invention and the practical application, and to enable others of ordinary skill in the art to understand the invention for various embodiments with various modifications as are suited to the particular use contemplated.