Technical Field

The product of the invention proposes a solution to the problem of detecting the cannabis plant whose cultivation is subject to permission, from the aerial images. It is a method for detecting cannabis from these images and presenting the detection to the user thanks to a sequence of processes to be applied on the images.

Brief Description of the Invention

The method of the invention is designed to detect the presence of a cannabis plant by using images from cameras that take images in various bands and runs on a computer. It informs the user by detecting whether there is a cannabis mark on the image, starting with the reading of the camera image and with various processing steps.

Detailed Description of the Invention

The method of the invention is a method for detecting cannabis from images from cameras that take images in various bands. Accordingly, images taken in multiple spectral ranges (for example, infrared regions called VNIR and SWIR) are primarily taken on a server. Here, although essentially from which source the images come, when they are taken and other variables are not important for the application of the method of the invention, the method of the invention allows the images taken from an aircraft to be processed instantly and converted into intelligence. Furthermore, it allows this process to be carried out conveniently within an aircraft. However, depending on user preferences, it is possible to archive the images taken by an aircraft and process them on the method of the invention after a while. Moreover, the fact that the images were taken by aircraft is not a necessity and is an example of ease of operation, for example, no problems were observed in the processing of the images taken by land vehicles.

Of course, capturing by plane is a matter of preference, as it offers certain advantages. Some of these advantages are to enable the detection of cannabis that is hidden in settlements, surrounded by other plants, and cultivated in remote areas. For this reason, air images obtained from UAV, manned aircraft and similar sources are preferably used in the use of the product of the invention. Thus, the product of the invention stands out in terms of its financial efficiency compared to other methods. Another issue is that terrorist organizations constitute sources of financial income. For this reason, as cannabis scans are frequently performed and it is necessary to detect cannabis planted on leave and cannabis planted illegally, this determination provides a cost-effective and effective screening opportunity at the desired frequency by air.

A two-camera system is described as an example throughout the description. This is divided into VNIR (Visible and Near Infrared) and SWIR (Short Wave Infrared). The product of the invention can work with images taken in this form, and this should not be perceived as a limitation. VNIR generally refers to images taken in the wavelength range of 400-1100 nanometers, while SWIR generally refers to images taken in the wavelength range of 900-1700 nanometers. These two cameras can be used in coordination or the same separation can be obtained by processing the images taken with a single camera as two different clusters. Due to the ease of expression and economic advantage throughout the description, the images at different wavelengths will be mentioned with different cameras. If a wavelength of 400-1700 nanometers is drawn with a single camera, the division of this image according to the wavelengths mentioned is a process that will be easily performed by a person skilled in the art, and this division can be easily performed if necessary.

To achieve this, the method of the invention first starts with the reading of the hyperspectral image. Here, the data received in various spectral ranges are taken as conjugate with other information on a server, after the radiometric corrections are made (101), the process of reading the images and removing the atmospheric bands (103) is carried out. Accordingly, the bands in which atmospheric bands are effective are removed from the VNIR camera image as they cause noise. Since these bands vary according to atmospheric conditions, they may correspond to different intervals in each sample. The bands corresponding to the wavelength of 760-780 nanometers in the open air are removed within this scope and the process sequence is continued with the remaining ones.

Then proceed with the spectral normalization (104) process. Accordingly, the vector normalization process is applied to eliminate the amplitude difference between the pixel signatures of the VNIR data and the reference library cannabis signatures. After spectral normalization (104), target determination is performed over spectral signatures. Reflection is obtained according to wavelengths with a hyperspectral image. Errors are observed in reflection spectra due to translocation and illumination that are not evenly distributed. It is aimed to eliminate the differences caused by the normalization process to a certain extent. Vector normalization is performed to eliminate the amplitude difference between the low-light pixel and the multi -light pixel in the data.

Here, the ||y ||i symbol represents the L-l norm and y represents the normalized spectral signature. The L-l norm of an N-band spectral signature is calculated as follows. lyli = Sr=ilyrl (2)

After normalization, target determination is performed over spectral signatures. As can be seen in the figure below, the amplitude difference between the non-normalized signatures is eliminated by the normalization process.

Instead of the L-l norm, normalization can be performed using the L-2 norm.

After the spectral normalization (104) process, the band reduction (105) process is applied. Accordingly, the increase in the number of bands in the classification creates the problem of analysis and processing of data sets containing complex and intensive information, although it provides detailed information to the objects on the stage. Although theoretically, as the spectral resolution and the number of bands used in the classification process increase, it is predicted that the pixels based on the classification can be more easily distinguished from each other, our studies show that this is not always the case. The main reason for this is that data sets with high spectral resolution are correlated with each other and have bands containing noisy, repetitive and unnecessary information. The dimensionality problem and the Hughes phenomenon are known as important problem that arises with the increase in the data size.

For these reasons, the images obtained from hyperspectral cameras in cannabis detection are reduced to 8 bands as a result of band reduction. It has been observed that dimension reduction has positive contributions to target detection algorithms. It was observed that the decrease in the number of bands increased the distinguishability of the cannabis target from other plant species. It has been observed that the contrast in the results of the target detection algorithms is higher. In addition, since dimension reduction from a large number of bands to 8 bands significantly reduces the data size, it increases the harness speed of the algorithms and provides results in a shorter time.

After this stage, signature similarity methods are applied (106). Within the scope of the method of the invention, Spectral Angle Mapper [1] was preferred as a signature similarity method. The Spectral Angle Mapper (SAM) is one of the most commonly used methods in the classification of hyperspectral images. This method is based on the calculation of the angle between the reference spectrum and the test spectrum. The angle between our reference spectrum and the unknown spectrum is calculated using the formula in the equation, x represents the spectrum of the test pixel, and y represents the reference spectrum that we already know. The smaller the calculated angle, the more similar the test spectrum is to our reference signature. SAM classification is performed by comparing the result of the algorithm with the predetermined threshold value. In the SAM algorithm, the angle is calculated using the following formula:

<x,y> ( ) (3) llxll llyll

The first N pixel with the lowest score of the SAM algorithm result is taken. Whether each pixel with a high score is cannabis is compared with the following criteria.

• Is the SAM result of the pixel higher than the predetermined threshold value?

• Application of Signature Similarity Methods for Certain Bands: If high, is the SAM result of the 3rd and 7th band interval of the pixel signature higher than the predetermined threshold value?

• Application of Signature Similarity Methods for Derivative Signatures: If high, is the SAM result between the derivative of the pixel signature and the derivative of the library signature higher than the predetermined threshold?

• Segment Pixel Number Analysis: Is the spatial segment of the pixel large enough? This is decided by comparing the total number of pixels in the region with the predetermined threshold value, which varies depending on the height. Here, the threshold value is a fixed value previously recorded on the server and can be changed by the user.

The application of signature similarity methods (106) can be applied to the whole, part or derived parts of the image as described above. What is envisaged here is the application of the same process, first for derivative signatures, then for certain bands, and finally for derivative signatures, respectively.

Then, window-based confidence interval analysis (107) is applied to the VNIR data. Accordingly, first of all, a binary map is created in which we can navigate the window we have determined. Signature similarity methods are applied to the image using the library signature to create this map. The SAM algorithm was used here as a signature similarity method. The SAM result obtained is sorted from the lowest value to the highest value. Then, depending on the height, the first N pixel with the highest score is taken according to the predetermined threshold value. This threshold value varies depending on the image height. A binary map is obtained as a result of the thresholded SAM. This map obtained will be used in window navigation. Then, starting from the pixel with the highest score, a square is drawn around each pixel so that its position is in the center. If the inside of the drawn square contains a dense amount of thresholded pixels on the binary map, that pixel is considered to be a reliable pixel. If the pixel is in a field, density is expected in the square. However, if the pixel is not a target but is in a small area (tree, etc.), it is expected that a small part of the square will be filled, not the density. Pixels with this feature are not considered reliable pixels.

Then, the target detection (108) process is run by statistical methods on the VNIR region image. Accordingly, the selected reliable cannabis pixel signatures are then searched in the image using target detection algorithms. There are various statistical-based detection algorithms based on sign detection and prediction theory used for target detection in the field of remote sensing. In other words, statistical detection methods such as Adaptive Cosine Estimator (ACE) and Generalized Likelihood Test Rate (GLRT) can be used as target detection algorithms. In the developed algorithm, the GLRT method was used in target determination. Generalized Likelihood Ratio Test (GLRT) algorithms are commonly used statistical detection algorithms. [2] There are two situations at this point: Only noise H O where the target is in the main beam, H O where the stirrer is not in the side beam, H l where the stirrer is in the side beam and where the target is not in the main beam.

H _o : noise (No target), noise ~ N _p (0, ^) _b ) (4)

In this equation, Np (0,£b) shows the receiver noises in the main and auxiliary channels, respectively. Here, the circular symmetrical complex Gauss with a mean of Np (0,£b) of zero and variance shows the random variable.

To distinguish the HO and HO hypotheses, the likelihood ratio test (LRT) can be written as follows:

Here, under the hypothesis Hi, the probability density function of X (7):

And the Ci matrix is calculated with the following formula (8):

C _t = E[XX ^H ; Hil i = {0 , 1} (8)

Then, the GLRT result obtained for each reliable cannabis pixel is thresholded with a previously recorded constant and the cannabis areas in the image are determined. A possible cannabis VNIR result detection map (109) is created by combining these determined areas. For this purpose, signatures that comply with the conditions specified in the image and are similar to the library signature are searched again with target detection algorithms in the image. A single result map is obtained by combining the result detection maps obtained. The result map obtained is compared with the determined threshold value and the VNIR result detection map (109) is obtained. At this stage, the method of the invention returns to the image from the band reduction (105) process. It calculates the ND VI index for the image by taking this data and taking it into the process of ND VI mapping (201). There are possible target areas by thresholding the calculated ND VI scores with a previously recorded constant.

The ND VI index is calculated according to the formula given in Equation (9). From the 8-band VNIR data obtained, the index was calculated using the 6th band and the 1st band. ND VI result map is obtained by comparing the calculated ND VI values with the determined threshold value.

At this stage, the special band selection (202) process is carried out within the SWIR region image and the bands containing different spectral information from the other plant species of the cannabis plant are selected and used in the detection. The 3 bands at the specified wavelengths are selected from the entire SWIR image. Target determination is made over the energy values in these bands.

After the special band selection (202) process, the energy analysis (203) of the bands is performed. Accordingly, the bands obtained from the SWIR hyperspectral data are compared with the determined threshold values. By multiplying the obtained result with the ND VI result map, SWIR camera detection results are obtained.

In the last stage, the results of the transactions made in the VNIR region and the results of the transactions made in the SWIR region are combined with the merging of the result maps (204). Accordingly, the final result map is obtained by multiplying the target detection result obtained from the VNIR camera with the result map obtained from the SWIR camera. Thus, it is seen that most of the false alarms obtained by using only the VNIR camera are eliminated by using VNIR and SWIR camera information together. Consolidated result maps can be presented to the user on the server on which the processes are performed or transferred to another device. Thus, the hyperspectral camera images are divided into two and then the process sequences carried out in two different spectral ranges are evaluated together and cannabis detection is performed quickly, effectively and reliably and presented to the user.

A method for detecting cannabis from infrared camera images for achieving the objects of the present invention is shown in the accompanying figure.

In this figure;

Figure 1: A schematic illustration of the method of the invention.

The explanations of the reference signs used in the drawings given in the annex of the description are as follows;

101- Radiometric corrections:

102: Reading images:

103: Removal of atmospheric gasses:

104: Spectral normalization:

105: Band reduction

106: Application of signature similarity methods

107: Window-based confidence interval analysis

108: Target determination with statistical methods

109: Creating a VNIR result detection map

201 : ND VI mapping

202: Custom band selection

203: Energy analysis of bands

204: Combining result maps

The non-patent references cited are as follows;

[1]: Demirel, B., Ozdil, O., & Esin, Y.E. (2016, May). Hyperspectral image segmentation based on spatial model. In 2016 24th Signal Processing and Communication Application Conference (SIU) (pp. 1249-1252). IEEE [2], D. Manolakis and G. A. Shaw, “Detection algorithms for hyperspectral imaging applications,” IEEE Signal Processing Magazine, vol. 19, no. 1, pp. 29 43, 2002.