TITLE SLOPE MOVEMENT CLASSIFICATION AND FAILURE WARNING SYSTEM FIELD OF THE INVENTION [001] The present invention relates to the field of mine safety and in particular to a system for predicting slope failure and providing early warning. BACKGROUND TO THE INVENTION [002] Slope stability monitoring using radar interferometry has become well- established in a number of applications including open cut mining, underground mining, dam wall monitoring, and similar situations. The basis of the technique is well described in international patent publication number WO2002046790 titled Slope Monitoring System and assigned to GroundProbe Pty Ltd. [003] The basic technique has been employed over the last 20 years and various extensions and associated technologies have been developed. A Method and System for Determining Alarm Conditions is described in international patent publication number WO2007012112 and techniques for Interferometric Signal Processing are described in international patent publication number WO2007009175. A useful technique for setting alarms is described in international patent publication number WO2012100288 titled Slope Stability Alarm. This publication describes setting alarms based on the ratio of the time taken for a slope to move between selected points compared the ratio of the distances between these points. The publication also includes a useful background on different types of slope movement including: regressive movements leading to stability; progressive movements leading to collapse; transitional movements which combine regressive movements followed by progressive movements; and stick slip which is a number of regressive/transgressive movements normally induced by an external influence such as rainfall, blasting or mining. The understanding of slope movement has since advanced. All of these applications have matured to patents in the name of GroundProbe Pty Ltd in various jurisdictions. [004] In addition to methodologies for processing the interferometric data, various techniques have been developed for visualising the data to assist geotechnical engineers to make decisions about the stability of the monitored slope. Reference may be had to international patent publication number WO2019119041 which describes the production of slope deformation maps. A Method and System for Displaying an Area is described in international patent publication number WO2015081386 and a geo-positioning technique is described in international patent publication number WO2016094958. These applications have also matured to patents for GroundProbe Pty Ltd. [005] It is also known to use LiDAR to provide the range data for slope stability monitoring. Many of the techniques described in the patents listed above are also useful with LiDAR systems. The application of LiDAR to slope monitoring is described in international patent publication number WO2017063033, issued in various jurisdictions to GroundProbe Pty Ltd. [006] At present, the visualisation of data and generation of alarms serve as an aid to geotechnical engineers in determining when action needs to be taken. The geotechnical engineer must decide the likely time to slope failure and when to evacuate an area at risk of failure. In a mining context, this is a very important decision since the cost of an unproductive mine is significant. There is therefore a tension between the need to evacuate an area due to the risk of injury to people and damage to equipment, and the economic loss of shutting down production earlier than is necessary. [007] The volume of radar data has increased to such a degree that it is virtually impossible for a geotechnical engineer to monitor slopes manually. Even with the assistance of various automated alarms the geotechnical engineer may still be overwhelmed. As mentioned in WO2019119041 (Production of Slope Deformation Maps) the geotechnical engineer does not need to just note movement, but needs to note the difference between fast moving small areas, slow moving small areas, fast moving large areas and slow moving small areas. A fast moving small area may present a greater risk than a slow moving large area. Thus a simple binary indication of movement is not sufficient. [008] The usefulness of neural networks has been recognised in a recent publication, “Contador Villegas, N, Huenchulao Catalán, H, Oliva Miranda, JM, Dubournais Donoso, F & Gallardo Arriagada, M 2021, 'Artificial intelligence applied to the detection and early warning of geotechnical instabilities in mining slopes', in PM Dight (ed.), SSIM 2021: Second International Slope Stability in Mining, Australian Centre for Geomechanics, Perth, pp. 227-23”. In this publication a technique is described in which a neural network is used to compare data against a pre-determined model of slope behaviour so as to classify the data as being recognisable behaviour or non-recognisable behaviour. The publication presents a binary classification system that is not predictive. [009] There is therefore a need for a system that can provide a timely warning. SUMMARY OF THE INVENTION [0010] In one form, although it need not be the only or indeed the broadest form, the invention resides in a slope movement classification system comprising: a slope stability monitoring radar or lidar to scan a slope and provide raw interferometric data of the slope pixel by pixel and scan by scan; a processor that calculates movement data including at least deformation from the raw interferometric data; and a neural network device that is trained with the movement data to establish a library that classifies the movement data into one of: (a) no deformation; (b) linear deformation; (c) regressive deformation; or (d) progressive deformation; wherein the neural network device matches further movement data against the library to classify the further movement data for every pixel and every scan into one of: (a) no deformation; (b) linear deformation; (c) regressive deformation; or (d) progressive deformation. [0011] The neural network device may further calculate a time to collapse if the further movement data is classified as (d) progressive deformation. [0012] Preferably the movement data also includes velocity. [0013] The deformation is preferably atmospheric-corrected deformation. The slope movement classification system may further comprise a weather station that provides the refractive index of the atmosphere for atmospheric correction. [0014] In addition to deformation the movement data may include one or more of amplitude, range and refractivity. [0015] The movement data may also include coherence. By coherence is meant a normalised complex cross-correlation function that indicates the quality of data at any point in time. [0016] The time windows for calculation of velocity may suitably be 30 minutes, 60 minutes, 360 minutes, 720 minutes, 1440 minutes or other similar time periods found to be suitable for the specific circumstances. [0017] In a further form, although again it need not be the only or indeed the broadest form, the invention resides in a method of classifying slope movement including the steps of: recording pixel by pixel and scan by scan raw interferometric data using a slope stability monitoring radar or lidar to scan a slope and produce movement data including at least deformation; training a neural network device with the movement data to establish a library that classifies the movement data into one of: (a) no deformation; (b) linear deformation; (c) regressive deformation; or (d) progressive deformation; recording further movement data using the slope stability monitoring radar or lidar, the further movement data including at least deformation; using the neural network device to compare the further movement data against the library to classify the further movement data for every pixel and every scan into one of: (a) no deformation; (b) linear deformation; (c) regressive deformation; or (d) progressive deformation. [0018] The method may further include the step of the neural network device calculating a time to collapse if the further movement data is classified as (d) progressive deformation. [0019] The movement data preferably includes velocity. [0020] The deformation is preferably atmospheric-corrected deformation. [0021] In addition to deformation the movement data may include one or more of coherence, amplitude, range and refractivity. [0022] The time windows for calculating velocity may suitably be 30 minutes, 60 minutes, 360 minutes, 720 minutes, 1440 minutes or other similar time periods found to be suitable for the specific circumstances. [0023] Further features and advantages of the present invention will become apparent from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS [0024] To assist in understanding the invention and to enable a person skilled in the art to put the invention into practical effect, preferred embodiments of the invention will be described by way of example only with reference to the accompanying drawings, in which: [0025] FIG 1 is block diagram of the elements of a slope movement classification system; [0026] FIG 2 shows detail of the neural network device of the invention; [0027] FIG 3 is a flow chart of the operation of the slope movement classification system of FIG 1; [0028] FIG 4 shows an example of a possible user interface displaying classification output from the invention; [0029] FIG 5 shows atmospheric deformation on the user interface of FIG 4; [0030] FIG 6 shows an example of a pixel over time exhibiting linear deformation changing to regressive deformation and then progressive deformation; [0031] FIG 7 shows an example of a pixel over time exhibiting linear deformation changing directly to progressive deformation; [0032] FIG 8 shows another example of a pixel over time exhibiting linear deformation changing directly to progressive deformation; [0033] FIG 9 shows an example of a pixel over time exhibiting regressive deformation changing to linear deformation and then progressive deformation; [0034] FIG 10 is a photograph of a neural network device; and [0035] FIG 11 is a photograph of the neural network device installed for operation. DETAILED DESCRIPTION OF THE INVENTION [0036] Embodiments of the present invention reside primarily in a slope movement classification and failure warning system. Accordingly, the method steps and elements have been illustrated in concise schematic form in the drawings, showing only those specific details that are necessary for understanding the embodiments of the present invention, but so as not to obscure the disclosure with excessive detail that will be readily apparent to those of ordinary skill in the art having the benefit of the present description. [0037] In this specification, adjectives such as first and second, left and right, and the like may be used solely to distinguish one element or action from another element or action without necessarily requiring or implying any actual such relationship or order. Words such as “comprises” or “includes” are intended to define a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed, including elements that are inherent to such a process, method, article, or apparatus. [0038] To assist in the understanding of the invention, there is shown in FIG 1 a block diagram of the elements of a slope movement classification system 100. One or more slope stability radars (SSR) 101 or slope stability lidars (SSL) collect data by monitoring a slope and using interferometric processing techniques to obtain deformation measurements of the monitored slope. The operation of an SSR or SSL is described in detail in the patents mentioned above, as well as in various scientific and trade publications. An SSR or SSL is a high resolution imaging sensor that works on collecting reflections of transmitted signals from small areas on a monitored wall called pixels (for a 2D device) or voxels (for a 3D device). Interferometric processing of the reflected signals detects phase changes that indicate wall movement. The techniques for slope monitoring using interferometric signal processing will be well known to persons skilled in the art. For clarity, the following description will refer to an SSR 101, although persons skilled in the art will readily understand that an SSL could also be employed. [0039] Processing of the raw interferometric data into movement data, such as deformation, velocity or inverse velocity occurs in a processor 102, which may be incorporated as part of the SSR 101, incorporated with a neural network device 103, or a separate stand-alone processor. Communication between the SSR 101, processor 102 and neural network device 103 may be wired, but due to site implementation requirements may include at least one wireless link 104. In most cases the neural network device 103 will be in a less harsh environment than the SSR 101, which will be deployed in the field. [0040] The neural network device 103 may produce a number of outputs including a user interface 105 and various alarms 106. The specific form of the user interface 105 does not form part of the invention, but for completeness a suitable user interface 105 is described below. Similarly, the alarms 106 may take various forms, as described later. [0041] The primary elements of the neural network device 103 are shown in greater detail in FIG 2. The neural network device 103 comprises in conventional form a central processing unit (CPU) 1031, RAM 1032, clock 1033 and library 1034. There is also provided an input interface 1035 for communication with the SSR 101 and/or processor 102, depending on whether the processor 102 is located with the SSR 101 or incorporated into the neural network device 103. An output interface 1036 communicates with the user interface 105 and alarms 106. [0042] The operation of the neural network device is described below, but in brief, the neural network device 103 uses machine learning for pattern matching between the data received from the processor 102 and the library of data 1034. To maximise operational speed a separate graphic processing unit (GPU) 1037 and associated GPU RAM 1038 may be used. Although the pattern matching could be performed in the CPU 1031, the dedicated graphics processor 1037 allows the machine learning function to run seamlessly with the operation of the SSR 101. [0043] Turning now to FIG 3, there is a shown a flowchart that sets out in broad terms the operation of the slope movement classification invention. The process commences with data collected from a SSR (or SSL as mentioned above). As is known, the raw radar (or LiDAR) data is interferometrically processed to provide movement data of a slope being monitored. [0044] As mentioned above, the processing of the SSR movement data may be done by a processor at the SSR, by a processor in the neural network device, or by a separate processor in communication with the SSR and neural network device. The processing of the movement data from the SSR generates a deformation measure, as is well known. In the preferred embodiment it also calculates a coherence measure, velocity and one or more of amplitude and range. [0045] Amplitude and range are obtained directly from the SSR. Amplitude describes the radar cross-section or reflectivity of target wall surfaces, which provides a useful metric for deformation signal quality. Similarly, the range can also be used as a quality metric, where a varying range for the same observed region indicates less confidence. [0046] The movement data provided to the neural network device is suitably corrected for any distortion due to atmospheric changes. SSR or phase-based interferometric radars are very sensitive to two-way propagation path length changes that can be introduced by the atmosphere, and these changes can look just like wall movement, especially to an untrained user. The radar propagation path variability is primarily determined by the amount of water vapour held in the atmosphere. In radar terms, the refractivity or dielectric properties of the air are modified by the amount of water, which is affected by changes in temperature, pressure and humidity. This issue is described in WO2007009175, mentioned above. [0047] Data analysis is often assisted by climate data estimations collected from a weather station (not shown), where temperature, pressure and humidity measurements can be used to estimate the refractive index of the atmosphere at a location. The use of a weather station is optional since refractive index can also be estimated directly from the radar data. The specific techniques employed are not part of the invention but examples and explanations are found in the patents referenced in the background section above. Use of a weather station is preferred. [0048] Coherence is described in international patent publication number WO20070091 (mentioned above), including examples of how it is an important and useful metric for slope monitoring, with figures that describe how the metric is used to identify wall movements from clutter signals. [0049] Coherence is a normalised complex cross-correlation function across different scans. The auto-correlation functions for both images are used to normalise the cross-correlation to produce a coherence value between 0 and 1, representing “bad” and “good” quality measurement areas on the wall surface. It indicates how much the wall surface around and within each pixel has changed between scans. It indicates the data quality for a pixel (or voxel for a 3D radar) at any point in time. [0050] The processor 102 may calculate velocity in mm/day (or mm/hour) from the movement data with different time windows. Suitably time windows include 30 minutes, 60 minutes, 360 minutes, 720 minutes and 1440 minutes. The invention is not limited to these specific time windows, but it has been found that these times are effective. Other time windows such as 45 minutes, 240 minutes, 1200 minutes and any other selected time may also be suitable so the invention should not be understood as being limited to these specific times. The preferred time window is selected to achieve acceptable signal to noise. For instance, a longer time window will have reduced sensitivity to atmospheric clutter but also delay detection of new movement. Conversely, a shorter time window will detect movement sooner, but at the expense of greater noise. In the preferred embodiment the time windows are user selected, however it is envisaged that the neural network device could be trained to automate the window selection for optimum signal to noise ratio, or some other desired outcome. [0051] The processor 102 may also calculate inverse velocity which is expressed as hours/mm rather than mm/hour. It has been found that the inverse velocity may be a useful measure to predict an imminent collapse when 1/v=0. In one embodiment of the invention an alarm 106 is generated when 1/v=0. [0052] The movement data is provided to the neural network device 103 for pattern matching against the library 1034. The library 1034 is compiled by training the neural network device with real movement data including at least deformation and possibly velocity. It will be appreciated that the neural network can be trained with deformation only, but a time base is inferred by the sequential nature of the data so that the neural network essentially uses velocity to classify the deformation. With sufficient training data a reliable library is established against which new data is compared. [0053] The library contains movement data classified into four possible classifications: a. No Deformation. This may include data which is identified as noise, static rocks, trees, grass, water, clouds, machinery, atmospheric contamination, etc; b. Linear Deformation. This applies when the ground is moving at a relatively constant rate that is not accelerating or decelerating to a significant degree; c. Regressive Deformation. This applies when the ground is decelerating and approaching a stable rock mass equilibrium; d. Progressive Deformation. This applies when the ground is accelerating towards a possible collapse and also ground movement when a collapse is in progress. [0054] The process of compiling the library requires various movements to be classified into the four classifications. The process may involve a labelling step and a training step. [0055] The labelling step requires a selection of historic data with different rock movement types to be collated and labelled correctly by a human, preferably a trained geotechnical engineer. The engineer considers deformation, amplitude, coherence, refractivity and a range of velocity window data to determine for which period of time each pixel or group of pixels in the historic data belong to which of the four output states – no deformation, linear deformation, regressive deformation or progressive deformation. This process is repeated until a sufficient amount of data has been labelled. As little as one dataset per class is required to train a machine learning model, however the greater the number of examples the more robust the model will be. Generally there must be hundreds of samples to generate a high quality model. Furthermore, the model may be trained with deformation alone, but the resilience of the model is greater if more parameters are included, such as amplitude, coherence, refractivity and range. For example, coherence is useful for learning to identify and ignore vegetation, and refractivity is useful to recognise weather conditions, such as snow. [0056] The training step requires that the labelled data is then passed to a neural network to learn the patterns and encode the patterns into a library of patterns. [0057] There exist two types of Machine Learning (ML) algorithms that may be useful to classify the time-series data obtained from the radar/lidar. The first type is the algorithms that can process static data with a defined temporal window such as random forest, support vector machines, feed-forward neural networks and the like, and the second type is the algorithms that can directly process time- series data with any variable length such as recurrent neural networks. The main difference is that the first type cannot directly deal with time-variable data and data must be pre-processed first. [0058] Windowing is a suitable method to transfer the time-series data into static data. In this method, a window size is chosen to divide the time-series data into multiple chunks, and then the chunks can be given to ML algorithms directly or it is possible to pre-process them with algorithms such fast Fourier transforms and give the frequency domain data to the ML models. Multilayer perceptron neural networks and support vector machines have been known to perform well for classification of those types of static data. [0059] Convolutional Neural Networks (CNNs) have become the most popular and dominate machine leaning in computer vision tasks due to their exceptional ability to capture spatial dependencies in images. Therefore, they have been used for time-series data as well. The time-series data must be first transferred into chunks by windowing and then chunks must be converted into images by using techniques such as Gramian angular field. [0060] The main shortcoming of the aforementioned techniques is choosing the proper window size that can properly capture the temporal dependencies of data. The problem exacerbates when dealing with data with variable-length temporal dependencies such as data captured with radar/lidar. [0061] Recurrent Neural Networks (RNNs) are popular and powerful machine learning algorithms that can capture temporal dependencies of time-series data without any pre-processing. In other words, raw time-series data are given to RNNs and they are able to capture variable-length temporal dependencies because of their internal state (memory). There exist several types of RNNs such as Elman networks and Jordan networks, simple RNNs without any gated connections, Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs). [0062] Gated recurrent neural networks are most widely used compared to other RNNs because they can deal with the vanishing gradient problem. In other words, they can deal with long temporal dependencies in data due to their gated connections. GRUs have a smaller number of parameters and simpler architecture than LSTMs, and therefore are computationally more efficient. However, it has been shown that their performance depends on the task and datasets, which means that although LSTMs are more complex and have more parameters, they do not necessarily perform better than GRUs. [0063] Once the library is established new movement data received from the SSR is processed and passed to the neural network device for classification. Classification occurs by comparing the current data to the library of data to derive a scan-by-scan, pixel-by-pixel classification of the new data. The output of the pattern matching process is a classification of the movement represented by the movement data into one of four classifications being: (a) no deformation; (b) linear deformation; (c) regressive deformation; or (d) progressive deformation. [0064] To exemplify the classification process, a series of user interface snapshots are shown in FIGs 4-9. The user interface 40 shows a chart 41 of deformation over time of an individual pixel. Immediately below is a chart 42 showing the classification over time of the same pixel. The bottom chart 43 shows the coherence measure that is input to the neural network device 103. On the right hand side of the user interface 40 is a visual image of the area being monitored, in this case an open cut mine wall at a mine in Central Queensland, Australia. [0065] The upper frame 44 shows a photograph of the mine wall with an overlay of classification for each pixel at a given time. The legend 45 for the overlay is on the right side. The classification is visualised on the classification map as shown in FIG 4 with progressive deformation in red 451, regressive deformation in yellow 452 and linear deformation in orange 453. No deformation is a clear lack of overlay 454. The same colours are used in the pixel classification chart 42. [0066] The lower frame 46 shows an overlay of the amount of deformation for each pixel at the same point in time. The legend 47 shows the deformation from -20mm to 20mm, with no movement being clear or no overlay. [0067] The charts 41, 42, 43 are for a single pixel which may be selected by the user. The selected pixel in each case is shown by a dotted circle, such as 48. The location of the dotted circle 48 is obviously the same in upper frame 44 and lower frame 46. The full circle visible on the charts is merely a mouse cursor location and can be ignored. [0068] The user is able to zoom to a relevant time period on the charts 41, 42, 43 by selecting a start and end time. The start time is indicated by a gray line 401 and the end time by blue line 402. A slider 49 provides the same indication. Various other tools and peripheral information are shown around the user interface, but these are not relevant for the present description. [0069] As mentioned above, it is useful to apply an atmospheric correction to the deformation measurements. By way of example, FIG 5 shows how atmospheric clutter is suppressed. Although the deformation image 56 and deformation time sequence 51 are noisy, the classification of pixel 58 is no deformation, as shown on the image 54 and on the pixel classification chart 52. Note that the scale for the deformation in this case is -5mm to 5mm. [0070] Turning to FIG 6, the highlighted pixel 68 displays interesting movement over time as indicated on the pixel classification chart. In the selected time period the initial deformation is linear but it becomes regressive at 621 and then progressive at 622, as indicated on pixel classification chart 62. The deformation image 66 and classification image 64 show deforming areas with complex spatial gradients and areas with different rate behaviours. As can be seen in the deformation chart 61, the change in the nature of the deformation is not readily apparent, but is identified in the classification chart 62. [0071] FIG 7 shows an example in chart 72 of linear deformation transitioning directly to progressive deformation at 721. As can be seen in the coherence chart, the slope fails (as evident by the loss of coherence). However, the deformation chart 71 does not give a clear indication of the impending failure until the rapid change in slope at 711. Clearly, the invention provides an alert of impending slope failure at 721 whereas prior art alerts would not be given until 711. This is a time difference of about 8 hours. [0072] FIG 8 shows another example in chart 82 of linear deformation transitioning directly to progressive deformation at 821. The example shows complex geomechanic behaviour of the rock mass and depicts the effectiveness of the invention in real-world situations. [0073] FIG 9 indicates how the deformation can have complicated changes that are not readily apparent in the deformation chart. In FIG 9 its can be seen that the pixel 98 is classified in chart 92 as displaying regressive deformation that transitions to linear deformation at 921 but then transitions to progressive deformation at 922. The slope subsequently fails as indicated by 933 in the coherence chart 93, about 5 hours after the indication by the classification. [0074] While the above description has focused on the classification system, the invention also includes the extension of the classification system to reliably predict time to failure. As indicated in the examples described above, the pixel classification chart identifies and displays transitions to progressive deformation, which is an indicator of future slope failure. Over time, the neural network device will build a library of indicators and time to failure which facilitates reliable prediction of time to failure. The invention then provides various alerts to initiate action at appropriate times so as to minimize risk to people and equipment but to maximize mine productivity (or other economic measures appropriate to the application). [0075] The neural network device 103 embodying the invention maybe a standalone device as shown in FIG 10. The device 103 may connect to the processor 102 by any of the available data transfer protocols, whether wired or wireless. In the embodiment of FIG 10 the connection is via Ethernet cable 1031. A series of coloured lights around the periphery 1032 provide a visual alert of the pixel classifications. The threshold for the visual alert may be based on the total number of pixels exhibiting, for instance, progressive deformation, or the percentage of pixels. If the threshold is reached the lights will glow red to indication a warning of impending progressive deformation, and therefore a risk of failure. The lights may then flash as an indication of the time to failure, the faster the flashing the closer to failure. Other forms of alarm, such as voice alarms and sirens are also envisaged. [0076] The neural network device 103 will most likely be connected to the processor 102 at a central monitoring site, as depicted in FIG 11. In this way a person monitoring a slope will have the benefit of the visual displays of FIGs 4-9, as well as the alerts described above from the neural network device 103. [0077] The significant benefit of the invention is that it provides a more discerning analysis of movement and displays it to a user in an intuitively useful manner. It will be appreciated that a small cumulative amount of ground movement that is progressive is more dangerous than a large cumulative movement that is linear or regressive. The same consideration applies to the size of a moving area. A big moving area does not necessarily mean a dangerous movement and a small moving area does not mean it is not dangerous. Prior art techniques do not adequately distinguish these differences. [0078] It should be recognized that classifications b, c and d above may all be significant deformation but, in most cases, b has insignificant risk, c has acceptable risk and only d has unacceptable risk. Prior art binary classification schemes would identify classifications b, c and d as significant deformation and therefore dangerous, whereas the invention takes risk into account and therefore only classifies d as significant high risk deformation. [0079] It is also important that the classification library is established using real data rather than pattern matching against synthetic data constructed from theoretical algorithms. [0080] The above description of various embodiments of the present invention is provided for purposes of description to one of ordinary skill in the related art. It is not intended to be exhaustive or to limit the invention to a single disclosed embodiment. As mentioned above, numerous alternatives and variations to the present invention will be apparent to those skilled in the art of the above teaching. Accordingly, while some alternative embodiments have been discussed specifically, other embodiments will be apparent or relatively easily developed by those of ordinary skill in the art. Accordingly, this invention is intended to embrace all alternatives, modifications and variations of the present invention that have been discussed herein, and other embodiments that fall within the spirit and scope of the above described invention. [0081] The reference in this specification to any prior publication (or information derived from it), or to any matter which is known, is not, and should not be taken as an acknowledgment or admission or any form of suggestion that the prior publication (or information derived from it) or known matter forms part of the common general knowledge in the field of endeavour to which this specification relates.