Selection of Satellite Image Series for the Determination of Forest Pathology Dynamics Taking Into Account Cloud Coverage and Image Distortions Based on the Data Obtained from the Key Point Detector E.O. Trubakov1, A.O. Trubakov1, D.A. Korostelyov1, D.V. Titarev1 trubakoveo@gmail.com|trubakovao@gmail.com|nigm85@mail.ru|titaryovdv@mail.ru 1 Bryansk State Technical University, Bryansk, Russia Remote sensing of the earth and monitoring of various phenomena have been and still remain an important task for solving various problems. One of them is the forest pathology dynamics determining. Assuming its dependence on various factors forest pathology can be either short-term or long-term. Sometimes it is necessary to analyze satellite images within a period of several years in order to determine the dynamics of forest pathology. So it is connected with some special aspects and makes such analysis in manual mode impossible. At the same time automated methods face the problem of identifying a series of suitable images even though they are not covered by clouds, shadows, turbulence and other distortions. Classical methods of nebulosity determination based either on neural network or decision functions do not always give an acceptable result, because the cloud coverage by itself can be either of cirrus intortus type or insignificant within the image, but in case of cloudiness it can be the reason for wrong analysis of the area under examination. The article proposes a new approach for the analysis and selection of images based on key point detectors connected neither with cloudiness determination nor distorted area identification, but with the extraction of suitable images eliminating those that by their characteristics are unfit for forest pathology determination. Experiments have shown that the accuracy of this approach is higher than of currently used method in GIS, which is based on cloud detector. Keywords: remote sensing of the earth, monitoring of forest pathology, image processing, descriptors of key points regions, millions of hectares of sound timber were lost because 1. Introduction of the harmful activity of the insects. The time of gnawing of the tree crown spent by a silkworm depends on its population. For As provided for in the item 1 and 2, article 60.5 of the example, a population of one hundred species gnaw the needles Forestry code of RF the state forest pathology monitoring is of a tree within a period of about a day, if the population increases induced into the state ecological monitoring and represents a tenfold, the time is reduced to a few hours. system of supervision by applying of both land, and remote In Bryansk region, the main needle-eating pests or leaf methods in regard to sanitary state and forest pathologies beetles are sawflies and timber beetles. The population of the (including testing, analysis and projected changes) [1]. timber beetle on one tree can be up to several tens of thousands. The procedure for exercising the state forest pathology With so many species it is able to destroy the tree within a month. monitoring is enacted into law by the order of Ministry of Natural That’s why the European Union considers the beetle to be one of Resources of RF of 5 April 2017 N 156 «On approval of the the most dangerous among secondary insects. Procedure for state forest pathology monitoring» [2]. According Thus, it can be concluded that, depending on the type of pest to this the task of remote observations concerning sanitary state destroying trees, the development cycle of pathology varies from and forest pathology is to identify changes in the sanitary state of several years to several weeks. For this very reason a constant forest and find forest pathologies alongside with preallotment forest monitoring is indispensable. and determination of the forest range boundaries and areas Continuous monitoring of the forest implies examination of having these changes. Besides it, according to [2] it is determined great number of satellite images taken for large areas. In this case that remote observation of sanitary state of the forest and forest taking into account the amount of data, manual image processing pathology should be carried out through the interpretation of is neither suitable, nor effective, i.e. semi-automatic or automatic space images and aerial photographs obtained owing to the use processing is required. However, the region of interest may be of aircraft and unmanned aerial vehicles. This factor makes it covered by clouds or other interference that make the dynamics necessary to obtain a set of images that can identify forest of forest pathology erroneous. pathologies and monitor their changes. Popular systems that provide unclosed series of satellite Detection of forest pathologies by remote method is based on imagery [5], also contain information regarding areas covered by the fact that the stressed tree dries out in a vegetative way. Thus, clouds (cloud mask). However, as shown, for example, in [4] the we can say that the period of monitoring coincides with the cloud mask Level 1C of Sentinel-2 images often omits the period of vegetation in the area under observation. In the Central presence of clouds (the average error is 37.4%). For the cases Federal District this period is approximately from 15 may to 15 when satellite images contain either opaque clouds with a large October (about 5 months per year). transition zone (between the core of the cloud and the clean In this short period of time it is necessary to identify areas) or Cirrus clouds this error can be more than 50%. The pathologies remotely and as quickly as possible, besides, it is same situation takes place when using other systems that provide important by using ground-based means to confirm them and to satellite imagery (Landsat, Wordview, etc.). Therefore, various identify the causes. After that, if applicable, it’s essential to get approaches and algorithms in the field of cloud detection are tough with the prevention of pathologies enlargement. being actively developed. Vegetation dying off of the forest can occur due to various Most of modern methods of cloud zone detection in satellite factors. Some of them can’t be fought against (for example, cases images can be divided into three groups. The algorithms of the of insufficient moisture). However, there are some other factors first group are based on calculations connected with certain destroying forest, such as diseases (root rot). This disease follows circuit groups of multispectral images devoted to indices and long while lasting for years (about a decade). One more cause of characteristics and also decision functions applied to them [9, 10, forest dying is plant pests. 14, 20]. They include the construction and analysis of A striking example of needle-eating insects and leaf beetles histograms, threshold determination, as well as the analysis of is the Siberian silkworm (dendrolimus sibiricus), which is the deviations, etc. The second group includes algorithms of most rampant in the forests of Siberia and the Far East. In these machine and deep machine learning based on artificial neural Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). networks, classification trees, Kohonen maps, etc. [7, 8, 16, 17, satellites, types of images and so on, can be specified. The 19, 21]. following parameters were used to study the proposed and Algorithms of these two groups give quite good results when standard methods: detecting clouds in certain classes of satellite images, but often 1. The coordinates of the study area: (N53.44835°, responds fallacious in more complex cases, when in satellite E34.26086°) – (N53.40170°, E35.37905°). images there are different types of clouds and snow of complex 2. Period: 2018-05-15 – 2018-10-15 (the vegetation period in geometric configuration. Therefore, in real-life situations, the 2018). algorithms of the third group, combining both approaches and 3. Satellite Sentinel-2. detecting not only clouds, but also snow and shadow [9, 20] are 4. Cloudiness: 0%. applied. At the same time it is possible to achieve the recognition With these search parameters 61 images are displayed in the accuracy of about 90% [14]. results. However, it is worth noting that in addition to clouds there Similar results are obtained when using the OData API when are other problems in satellite images that interfere with the prompted. successful detection of the forest pathology dynamics. They After analyzing the results visually, you can see that some include defocusing, foreign objects and their shadows, images have noise (including elements of clouds). Due to such turbulence, noise, etc. Consequently, such image may be defects (noise, clouds), problems may arise in the following unsuitable for analysis in spite of nil clouds conditions. analysis of the images. If images with defects are immediately Thus, research in the field of methods and algorithms to used for forest pathologies recognizing, the probability of automate the implementation of a satellite image series over a erroneous results increases for both: monitoring and training long period of time for forest pathology determination is tasks. Therefore, it is necessary to carry out additional processing required. with a series of received images to filter out noisy and defective ones. Often 1-2 images with noise or clouds in the analyzed area 2. Finding of forest pathologies in image series can lead to an erroneous forecast of forest pathology dynamics. For automated determination of forest pathology dynamics it 3. Detection of images suitable for determining is necessary to have a sufficiently large image series concerning the region of interest. Images from the Copernicus open system the forest pathology dynamics are often used for this purpose [5]. From almost all satellite Experiments have proved that the currently used approaches systems images of medium and high resolution are presented in to determine images without clouds and distortion, used in the form of multispectral images. This feature of these images Copernicus or in a number of modern GIS (eg QGIS) make a makes the selection of channels giving more information about number of errors. For example, for the conditions described in typical objects under investigation possible, because it can the previous paragraph, the Semi-Automatic Classification exclude unnecessary information concerning foreign objects and Plugin (SAC) from the QGIS package ( that is used very often to emphasize the data which is important for the task. solve similar problems) has identified 6 "ideal" images over the Transformation of the image by the principal component past 5 months (with the criterion of nil clouds). Besides, the method will allow allocating of the most significant reflected frequency of images is irregular (having intervals in images of spectra, excluding chartjunks but without data loss. In this 54 days). These indicators in some types of pathologies are method, the first component emphasizes the spectral contrast as insufficient. much as possible, while the second one reflects the boundaries of It is worth noting that among the images removed by the the contrast change. This approach allows improving the results module there are those that could be used to analyze the of manual (visual) interpretation and automatic (semi-automatic) dynamics of forest pathology. Most often these are the images classification [3]. which are partially covered by clouds, but the area of interest in For the study of vegetation most indices are based on the them is not distorted. It should be mentioned that the manual difference between the spectral profile of healthy organics and selection of a series of images in practice is time-consuming and the profile oppressed by diseases or pests. The most actively is not popular. vegetating plant communities absorb more radiation in the red The paper proposes an alternative approach to the selection zone of the spectrum and reflect more in the near infrared scope of a series of images. The main idea of the approach is that it compared to other objects [18]. The most commonly used involves proceeding not from the parameters of clouds or standardized index NDVI [11]: distortions, but from the parameters of the area of interest and its 𝑁𝐼𝑅 − 𝑅𝐸𝐷 specific points and parameters. To do this, it is proposed to use 𝑁𝐷𝑉𝐼 = , 𝑁𝐼𝑅 + 𝑅𝐸𝐷 key points based on one of the known descriptors. In this case, where, 𝑁𝐼𝑅 is the reflection in the nearest infrared zone of the the image is considered suitable for a series of images if and only spectrum, 𝑅𝐸𝐷 is the reflection in the red region of the spectrum. if a predetermined percentage of key points of interest, calculated It should be borne in mind that the NDVI index is quite rough, from the reference image can be found in the candidate image. especially with high and low canopy density. Therefore, it has The proposed algorithm consists of two stages. been replaced by the less sensitive soil index MSAVI2 (Modified Stage 1-pre-training in the initial image: Soil Adjusted Vegetation Index 2) [12]: 1. Manual selection of the single reference image that shows the 𝑁𝐼𝑅 − 𝑅𝐸𝐷 area of interest. 𝑀𝑆𝐴𝑉𝐼2 = ∗ (1 + 𝐿), 𝑁𝐼𝑅 + 𝑅𝐸𝐷 − 𝐿 2. Marking of the area of interest. It is important to take into where, 𝑁𝐼𝑅 is the reflection in the nearest infrared region of the account that the area should not be too small (otherwise the spectrum; 𝑅𝐸𝐷 is the reflection in the red region of the spectrum, number of key points will not be sufficient) for the stable 𝐿 is the coefficient, which is by the formula: operation of the algorithm. 2 ∗ 𝑁𝐼𝑅 + 1 − √(2 ∗ 𝑁𝐼𝑅 + 1)2 − 8 ∗ (𝑁𝐼𝑅 ∗ 𝑅𝐸𝐷) 3. Obtaining reflection layers in the near infrared and red 𝐿 =1− . spectral region (𝑁𝐼𝑅, 𝑅𝐸𝐷): {B04, B8A}. 2 The official resource of the Copernicus project was used to 4. Formation of a monochrome image by the algorithm obtain images in the conducted studies [5]. The selection of 𝑀𝑆𝐴𝑉𝐼2: image series from this portal can be done interactively through 𝑁𝐼𝑅 − 𝑅𝐸𝐷 𝑀𝑆𝐴𝑉𝐼2 = ∗ (1 + 𝐿). search tools or using public API. In interactive mode, the search 𝑁𝐼𝑅 + 𝑅𝐸𝐷 − 𝐿 area of interest, dates of images, cloud parameters, types of 5. Calculation of key points of interest, computation of their more than the threshold, the point is considered an angle, in other descriptors. words, a special point. After the first stage we get a reference vector of descriptors The above mentioned methods determine the key points regarding key points of interest working with the pixels of the image. There is an alternative Stage 2– automated selection of a series of images with approach, which involves the application of machine learning neither distortions nor clouds covering of the area of interest: algorithms. For illustrative purpose of such methods the FAST 1. Getting a snapshot from Copernicus system via internal API. method is chosen, the principle of which lies in building decision 2. Selection of reflection layers from the obtained image in the trees to classify pixels [13]. nearest infrared and red spectral regions (𝑁𝐼𝑅, 𝑅𝐸𝐷). The method is based on the following: for each pixel p we 3. Selecting an area by applying a mask to the image using the consider a circle with a radius of 4 pixels, inscribed in a square logical « AND » operation by means of a single mask of the area with a side of 7 pixels. On the basis of the selected region, area of interest: 𝑅𝑒𝑠 = 𝐼𝑚𝑔 & 𝑀𝑎𝑠𝑘. the importance of the point is concluded. Each of the pixels of 4. Getting of layers in the spectra 𝑅𝐸𝐷, 𝑁𝐼𝑅: {B04, B8A} the circle (the circle includes 16 pixels) relative to the pixel p can 5. Formation of a monochrome image according to the be in one of three States: transformation algorithm 𝑀𝑆𝐴𝑉𝐼2. 𝑑, 𝐼𝑥 ≤ 𝐼𝑝 − 𝑡 (𝑑𝑎𝑟𝑘𝑒𝑟) 6. Calculation of key points and their descriptors. 𝑆𝑝→𝑥 = { 𝑠, 𝐼𝑝 − 𝑡 ≤ 𝐼𝑥 ≤ 𝐼𝑝 + 𝑡 (𝑠𝑖𝑚𝑖𝑙𝑎𝑟). 7. Comparison of the obtained descriptors and those obtained 𝑏, 𝐼𝑝 + 𝑡 ≤ 𝐼𝑥 (𝑙𝑖𝑔ℎ𝑡𝑒𝑟) from the reference image. For each 𝑥 and found 𝑆𝑝→𝑥 for each 𝑝 ∈ 𝑃 (𝑃 is the variety (𝑓 ∗ 𝑔)𝑖 ≝ ∑ 𝑓𝑗∗ 𝑔𝑖+𝑗 , of all pixels of the training image set) the set is divided into 3 𝑗 subsets of points 𝑃𝑑 (darker), 𝑃𝑠 (similar), 𝑃𝑏 (lighter). Then the where 𝑖 is dragging between the sequences relatively to each decision tree is built. According to the results of this decision other and the superscript in the form of an asterisk means tree, the angles on the test images are determined. complex conjugation. If the number of similarities is more than A key drawback of this approach connected with identifying some threshold value N, add the image to the series. If there are special points is the order in which the points are selected and some images in the Copernicus system obtained within the which influences the effectiveness of the work. It is also decided time period, proceed to step 1. necessary to take into account the fact that there may be other key points in the environment of the initial point and in this case 4. Key point detectors and descriptors the method may be fallible. Image distortion is a significant obstacle in the operation of At the moment, there are many well-proven methods of detectors. This is because the algorithm may fail to detect key identification (detection) of key points. The most widespread of points on subsequent frames of the same area due to various them were investigated in the work. frame changes. Images of the same area taken by different One of them is the Harris method [6]. The principle of the spacecraft may differ owing to the deviations in sensors, shooting method is that for the image under consideration 𝐼 a window 𝑊 conditions (position of the vehicle, season of the year, is allocated with the center at the point(𝑥, 𝑦) then it is diddled atmosphere). In the conducted researches the work of detectors towards (𝑢, 𝑣). The size of the selected window depends on the was contrasted taking into consideration the following types of size of the image. Then the sum of squared differences between distortions: blurring and darkening (they can occur due to the initial and diddled window (𝐸(𝑢, 𝑣)), is calculated using the atmospheric phenomena). A number of images of the specified formula: area were taken, then artificially with the help of graphic editors 2 𝐸(𝑢, 𝑣) = ∑ 𝑤(𝑥, 𝑦)(𝐼(𝑥 + 𝑢, 𝑦 + 𝑣) − 𝐼(𝑥, 𝑦)) ≈ , distortions such as darkening and blurring were made. After that, (𝑥,𝑦)∈𝑊 measurements of operating time took place and a number of 2 𝑥 points in the initial images was identified. Besides, the key points ≈ ∑ 𝑤(𝑥, 𝑦)(𝐼𝑥 (𝑥, 𝑦)𝑢 + 𝐼𝑦 (𝑥, 𝑦)𝑣) ≈ (𝑥, 𝑦)𝑀 (𝑦) on the distorted images were also searched and the found points (𝑥,𝑦)∈𝑊 were compared with the reference ones. Thus, a percentage where 𝑤(𝑥, 𝑦) is impulsive response (Gaussian function is the discrepancy between the key points of the reference image and most popular); 𝑀 is autocorrelation matrix: the distorted one was defined. 𝐼𝑥2 𝐼𝑥 𝐼𝑦 Less resistant to blurring turned out to be the FAST method. 𝑀 = ∑ 𝑤(𝑥, 𝑦) [ ], 𝐼𝑥 𝐼𝑦 𝐼𝑦2 With this type of distortion, the method makes an error of about (𝑥,𝑦)∈𝑊 With large changes in the direction of the function 𝐸(𝑥, 𝑦) in 38% of the found points. For the same set of images, the Harries the direction 𝑥 and 𝑦 large modulo the eigenvalues of the matrix method made approximately 1.5% of errors. So did the Shi- 𝑀. are obtained. Due to the complexity the matrix eigenvalues Tomasi method. Thus, the invariance of the Harries and Shi- calculations, a response measure 𝑅, is often used, determined Tomasi methods in regard to distortions of blurring type can be confirmed. In the second type of distortions (darkening) the from the formula: 𝑅 = 𝑑𝑒𝑡𝑀 − 𝑘(𝑡𝑟𝑀)2 > 𝑘, where 𝑘 is the FAST method also gave worse results. In this case the error was empirical constant from the interval [0,04; 0,06]. about 43%. And both methods: of Harries and of Shi-Tomasi In this case the value 𝑅 will be positive for the angular key showed invariance with respect to this type of distortion. points. The local maxima of the response function in the However, it is necessary to take into account the fact that neighbourhood of a given radius are calculated in the midst of these methods are adapted to the objects of artificial origin, while the identified points, and these obtained points are selected as space images are more natural, that is they have got a more angular key points. monotonous texture. Therefore, the criterion of the number of The advantage of the considered method is its stability to found key points is more important. rotations and exceptionally to affine transformations of the In studied images of the forest both Harries and Shi-Tomasi image. But as for disadvantages, a significant sensitivity to noise methods found approximately 73.16% fewer key points than the in the image should be mentioned. There is one more detector of FAST method. At the same time the search time for points was key point detection similar to the Harris method. It is the Shi- approximately 78.57% of the search time consumed by the FAST Tomasi angle detector [15], which differs in the calculation of method. the response measure. This method computes eigenvalues On the basis of the conducted researches it is possible to draw directly, so finding angles will be more stable. The bottom line a conclusion about further application of the FAST method in the lies in defining a threshold value and if the calculated value is further experiments. Since the number of key points found in presented (fig. 3). The Harries method as well as the Shi-Tomasi images is of higher priority than the invariance in respect to method had 0% of matching. image distortions. The second stage of the study was devoted to the application of detectors capable of determining the unsuitability of the image in the task of forest pathology monitoring based on the algorithm proposed in the previous section. If the image is suitable, the FAST method finds about 89- 93% of the reference points in the image. Fig. 1 shows pictures taken 23.05.2019 and 02.06.2019. Fig. 1 illustrates the results of the FAST method for these images with the coincidence of key points reaching 95%. Fig. 4. Key points of the FAST method for images in fig. 3 All these conducted researches have resulted in the conclusion that Harries and Shi-Tomasi methods are more preferable than the FAST method. This is due to the fact that these two methods have a much larger difference between the thresholds of usable and unusable images comparing to the FAST method. 5. Research results The final studies were the experiments aimed at finding Fig 1. Satellite images of the lake «Krugloe» area, taken satellite images free of clouds and other noise. To compare the 23.05.2019and02.06.2019 results of the developed algorithm, the QGIS SAC module was used. The result of the search for images with 0% nebulosity in the SAC module accounted for 6 images, with a maximum nebulosity of 100% there were 61 images. The developed algorithm used in the same area and time limitations having the threshold value of suitable images equal to 89% gave the result of 9 images, which is 30% better than SAC. To confirm the operation of the algorithm a manual selecting of all images took place, which resulted in 15 suitable images. 6. Conclusion Fig 2. Key points of the FAST method for images in fig.1 The article analyzes the process of forest pathology However, in cases of the image impropriety, this method monitoring. The necessity of search automation applied for finds about 50-70% of the reference points. fig. 3 shows images processing suitable forest images obtained from satellites was taken 23.05.2019 and 25.06.2019 and fig. 4 shows the results of realized. Empirical studies connected with the application of key the FAST method, where the coincidence of key points is 70%. point detectors in the image were conducted in order to assess the applicability of the image for further processing and monitoring of forest pathology. The studies revealed that a large number of key points is a hindrance in determining the suitability of the image, that is why the FAST detector was not introduced in further studies. In the final experiments, it was confirmed that the proposed method of identifying suitable images for forest pathology monitoring produces results by 30% better than the well-proven QGIS SAC module. As a result, more images can be obtained for better tracking of pathology dynamics. However, it should be borne in mind that the result of the work depends not only on the Fig. 3. Satellite images of the lake «Krugloe» area, taken chosen method of key point detection, but also upon the threshold 23.05.2019 и 25.05.2019 value of the key point correlation accuracy In addition to this investigation there were studies in regard to other methods, when the number of points is not as large as 7. References with the FAST method. These are Harries and Shi-Tomasi [1] Forest code of the Russian Federation as amended on methods. December 27, 2018 (part 4, article 60.5) The results of similar test of Harries and Shi-Tomasi methods [2] The order of April 5, 2017 N 156 «On approval of the state differ by about 2-3%. In the case of good images, reference point forest pathology monitoring procedure». correlation is approximately 80-95%. [3] Showengerdt R. Remote sensing. Models and methods of Exemplarily the results of the methods on the same images image processing. M., 2010. 560 p. similar to the FAST method are presented (fig. 1). The Harries [4] Coluzzi, Rosa & Imbrenda, Vito & Maria, Lanfredi & method matched 93% and the Shi-Tomasi method matched 95%. Tiziana, Simoniello. (2018). A first assessment of the As applied to unsuitable images, the key point correlation is Sentinel-2 Level 1-C cloud mask product to support approximately 0-33%. For purpose of illustration the results of informed surface analyses. Remote Sensing of the methods in the same images used by FAST method are Environment. 217. 426-443. 10.1016/j.rse.2018.08.009. [5] Copernicus Open Access Hub. – URL: Environment (2015), https://scihub.copernicus.eu. http://dx.doi.org/10.1016/j.rse.2014.12.014. [6] Harris, C and Stephens, M (1988). "A Combined Corner and [21] Zi, Y., Xie, F., Jiang, Z.: A cloud detection method for Edge Detector". Alvey Vision Conference Landsat 8 images based on PCANet. Remote Sens. 10(6), [7] Hughes, M. J., Hayes, D. J. "Automated detection of cloud 877 (2018) and cloud shadow in single-date Landsat imagery using neural networks and spatial post-processing", Remote Sens., vol. 6, no. 6, pp. 4907-4926, 2014. [8] Jeppesen, J.H., Jacobsen, R.H., Inceoglu, F., Toftegaard, T.S. A cloud detection algorithm for satellite imagery based on deep learning. Remote Sensing of Environment, 229, 2019, pp. 247-259. https://doi.org/10.1016/j.rse.2019.03.039. [9] Kwan, C., Hagen, L., Chou, B. et al. Simple and effective cloud- and shadow-detection algorithms for Landsat and Worldview images. SIViP, 2019, pp. 1-9. https://doi.org/10.1007/s11760-019-01532-2 [10] Man D.C., Luu V.H., Hoang V.T., Bui Q.H., Nguyen T.N.T. (2015) Cloud Detection Algorithm for LandSat 8 Image Using Multispectral Rules and Spatial Variability. In: Nguyen VH., Le AC., Huynh VN. (eds) Knowledge and Systems Engineering. Advances in Intelligent Systems and Computing, vol 326. Springer, Cham [11] Pettorelli, N., Vik, J. O., Mysterud, A., Gaillard, J.-M., Tucker, C. J., Stenseth, N. C. Using the satellite-derived NDVI to assess ecological responses to environmental change // Trends in Ecology and Evolution. 2005. Vol. 20. P. 503–510. DOI: 10.1016/j.tree.2005.05.011. [12] Qi, J., Chehbouni, A., Huete, A.R., Kerr, Y.H., Sorooshian, S. A modified soil adjusted vegetation index. Remote Sensing of Environment. V. 48, № 2, 1994, pp. 119–126. [13] Rosten, Edward; Tom Drummond (2005). Fusing points and lines for high performance tracking. IEEE International Conference on Computer Vision. 2. pp. 1508–1511. CiteSeerX 10.1.1.60.4715. doi:10.1109/ICCV.2005.104. ISBN 978-0-7695-2334-7. [14] Scaramuzza, P.L., Bouchard, M.A., Dwyer, J.L.: Development of the Landsat data continuity mission cloud- cover assessment algorithms. IEEE Trans. Geosci. Remote Sens. 50(4), 1140–1154 (2012) [15] Shi, J. and Tomasi, C. (June 1994). "Good Features to Track". 9th IEEE Conference on Computer Vision and Pattern Recognition. Springer. pp. 593–600. CiteSeerX 10.1.1.36.2669. doi:10.1109/CVPR.1994.323794 [16] Shiffman, S. Cloud detection from satellite imagery: a comparison of expert-generated and automatically generated decision trees. In Proceedings of the Eighth Workshop on Mining Scientific and Engineering Datasets, held in conjunction with the 2005 SIAM International Conference on Data Mining, April 21-23, Newport Beach, CA, 2005. [17] Suzanne Angeli, Arnaud Quesney and Lydwine Gross (November 21st 2012). Image Simplification Using Kohonen Maps: Application to Satellite Data for Cloud Detection and Land Cover Mapping, Applications of Self- Organizing Maps, Magnus Johnsson, IntechOpen, DOI: 10.5772/51352. [18] Tucker C. J. Red and photographic infrared linear combinations for monitoring vegetation // Rem. Sens. of Env. 1979. Vol. 8. P. 127–150. DOI: 10.1016/0034- 4257(79)90013-0. [19] Xie, F., Shi,M., Shi, Z., Yin, J., Zhao, D.: Multi-level Cloud Detection in Remote Sensing Images Based on Deep Learning. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 10(8), 3631–3640 (2017) [20] Zhu, Z., et al., Improvement and expansion of the Fmask algorithm: cloud, cloud shadow, and snow detection for Landsats 4–7, 8, and Sentinel 2 images, Remote Sensing of