=Paper=
{{Paper
|id=Vol-3126/paper47
|storemode=property
|title=Spectral-spatial analysis of data of images of plantings for identification of stresses of technological character
|pdfUrl=https://ceur-ws.org/Vol-3126/paper47.pdf
|volume=Vol-3126
|authors=Natalia Pasichnyk,Dmytro Komarchuk,Korenkova Hanna,Sergey Shvorov,Oleksiy Opryshko,Nikolay Kiktev
}}
==Spectral-spatial analysis of data of images of plantings for identification of stresses of technological character==
Spectral-Spatial Analysis of Data of Images of Plantings for Identification of Stresses of Technological Character Natalia Pasichnyk1, Dmytro Komarchuk2, Hanna Korenkova 3, Sergey Shvorov4, Oleksiy Opryshko5 and Nikolay Kiktev6 1,2,4,5,6 National University of Life and Environmental Sciences of Ukraine, Heroyiv Oborony st., 15, Kyiv, 03041, Ukraine 3 Odessa I.I.Mechnikov National University, Dvoryanskaya str., 2, Odessa, 65082, Ukraine Abstract Methods of spectral-spatial analysis are promising for the identification of technological stresses. The most common solution for interpreting the causes of stress is the use of machine learning technologies, namely neural networks. As at technological stresses in particular at chemical poisoning of crops, there can be various options of the coloring of the affected plants the possibility of providing a sufficient amount of initial data for training of neural networks is doubtful. An alternative is graph analysis of the distribution of stress areas on the field map. Given the urgency of the problem for promising technologies of precision agriculture, the work aimed to develop a spectral-spatial method of monitoring technological stresses, namely the algorithm and software for its. Experimental studies of the manifestation of technological stresses on winter crops on the example of wheat and rapeseed were conducted during 2018-2020 in production fields using universal cameras in the visible range and special multispectral Slantrange systems. For remote monitor, the state of winter crops, an algorithm for identifying technological stresses was developed, which is implemented in the developed software in Python for spectral-spatial analysis of stress index maps. It has been experimentally confirmed in the production fields that the use of the developed software allows identifying the contours of areas of plants with stresses of technological nature based on stress index distribution maps. Keywords 1 UAVs, winter crops, vegetation indices, stresses, herbicides 1. Introduction data processing systems for satellite data filtering and machine learning technology for object recognition. That is, in the spectral-spatial The prospects for agricultural production analysis, the field of the field as a whole was not management based on objective remote considered as the object of research. This monitoring data were obvious both at the state approach is used in particular in aviation for the level and for agricultural enterprises. implementation of orientation in the use of Accordingly, research was carried out to develop electronic warfare as shown in the work of S. various theories and methods for obtaining Shvorov and others (2018) in [2]. Regarding information about vegetation. Under uncertainty, agricultural production, Xianlong Zhang and M. Lotfi et al. (2009) in [1] proposed computer others (2019) in [3] proposed the division of ISIT 2021: II International Scientific and Practical Conference «Intellectual Systems and Information Technologies», September 13–19, 2021, Odesa, Ukraine EMAIL: N.Pasichnyk@nubip.edu.ua (A. 1); dmitruyk@gmail.com (A. 2); av.korenkova@gmail.com (A. 3); sosdok@nubip.edu.ua (A. 4); ozon.kiev@nubip.edu.ua (A. 5); nkiktev@gmail.com (A. 6); ORCID: 0000-0002-2120-1552 (A. 1); 0000-0003-3811-6183 (A. 2); 0000-0001-7207-3688 (A. 3); 0000-0003-3358-1297 (A. 4); 0000-0001-6433-3566 (A. 5) ©️ 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings (CEUR-WS.org) spectral-spatial monitoring methods into 2 need to be adapted to respond to such problem conditional categories. The first category uses the areas, as weakened plants are easily affected by spectral characteristics of terrestrial objects and pests and can become a breeding ground for them. then obtains vegetation information by comparing A possible technology for the analysis of the difference with the results of spectral spectral-spatial distribution is artificial neural monitoring. An example of such monitoring is the networks which, due to the rapid development of identification of trees in densely populated cities multi-core processors, have become available to based on satellite images shown in S.W. Myint et farmers. There is a positive experience of using al. (2013) in [4]. neural networks for various monitoring tasks The second category is based on a combination which, if necessary, can be adapted to monitor of external knowledge such as decision tree for technological stresses. Section 1 shows that in the image classification shown in Andrea S. Laliberte initial stages of the growing season, the et al. (2007) in [5], neural networks, and wavelet dimensions of plants may indicate their stress. transforms described in Mitch Bryson and others Neural networks for estimating plant height (2010) in [6]. This promising method has not been during rice lodging are shown in Ming-Der Yang widely used in satellite monitoring because the et al. (2020) [11]. According to the provided combination of time delay and low resolution results, it was possible to detect rice lodging with leads to unacceptably large errors, which was acceptable accuracy based on images from shown in the article by Passang Dorji et al. (2017) universal cameras in the visible range, but the in [7]. UAV monitoring is devoid of these calculations were performed using cloud services, shortcomings and accordingly, this method can be which is difficult to implement in our country. implemented on a new technological basis. Thus, Autonomous work of neural networks is shown in in the work of J. Senthilnath et all (2017) [8], it the work of Wojciech Gruszczyński and others was possible to successfully identify weeds in (2019) [12] to identify grass among general crops by fixing plants in automatically vegetation. When analyzing the image was determining technological tracks. segmented into parts and carried out training of Wavelet analysis methods do not require the the network on the distribution of the cloud of division of the image into blocks, because the points. This approach is promising for the required localization properties are already identification of low-growing grass because only embedded in the wavelet system. Accordingly, it one manifestation is considered, but under is possible to filter out a significant number of technological stress, there may be more. In errors inherent in pixel analysis methods. The principle, for neural networks, there can be method of wavelet analysis for the identification several options for identifying objects. They can of affected areas due to technological stresses, be used in particular to determine the state of rice namely the prolonged action of herbicides was yield at the stage of ripening, as shown in Qi Yang shown in the work of M. Dolia and others (2019) and others (2019) [13], or the state of mineral in [9]. The proposed solution proved its nutrition described in V. Lysenko and others effectiveness when, as a result of a dosing error in (2017) [14]. Spatial distribution was also a part of the field, a higher dose of herbicides was considered in Yan Pang et al. (2020) [15] to applied. Since the application was made by calculate the number of plants in a ridge. All these appropriate ground equipment, the authors in the works are combined by a limited number of analysis of the map image focused on the search classification options and a large sample of source for linear functions. The authors noted some data for neural network training. In this case, in difficulties in the established systems when contrast to the vegetation indices, which focus on choosing thresholds. The complexity of this pixel-by-pixel analysis, the training of neural controversial issue was confirmed by Yu-Hsuan networks was based on crop areas obtaining more Tu et al. (2020) [10] where it was the limit values accurate results. that were recommended to be studied at higher As at technological stresses in particular at resolutions. Due to this specificity of the method, chemical poisoning of crops, there can be various the analysis will be effective for the affected crops options of the coloring of the affected plants the on a large scale, which significantly limits its possibility of providing a sufficient amount of effectiveness. Large-scale impressions can be initial data for training of neural networks is easily identified by satellite technology or doubtful. ground-based monitoring, but small areas will be There are no ready-made software solutions difficult to detect. Crop management technologies for analyzing the distribution of stress areas on the field map to identify the nature of stress. Given controllers and flight control of UAVs, the the urgency of the problem for promising prospects of the Python language for these tasks technologies of precision agriculture, the work were emphasized primarily due to the use of aimed to develop a spectral-spatial method of technical means from different manufacturers on monitoring technological stresses, namely the different operating systems. algorithm and software for its implementation. 2.3. Experimental research 2. The state of the issue 2.1. Identification of direction of Experimental studies of the manifestation of technological stresses on winter crops on the movement of technological example of wheat and rapeseed were conducted equipment during 2018-2020 in production fields. Photography was performed using: in 2018-2019, Stressful conditions of crops of technological to monitor the stationary experiment and character are caused by human actions which are fragments of production fields - hexacopter based realized by the use of the ground technological on multi-rotor platform CD600 with a set of equipment. The identification of equipment specialized sensor equipment in the digital action directions was considered in Junfeng Gao et al. camera GoPro HERO4, in 2019-2020 - (2018) [16] regarding the detection of weeds in multispectral system 3p, mounted on a DJI row crops, where all plants between rows were Matrice 600 hexacopter, which allowed to obtain considered weeds. In Carlos Henrique Wachholz orthophotos of industrial fields. It is the spectral- de Souza et al. (2017) [17], sugar cane rows were spatial analysis of the obtained orthophoto plan identified to estimate row gaps. In both cases, the that allowed us to establish the dependences on rows were considered to be the arrangement of the basis of which the identification of plants in a row, because this is how ground technological stresses is carried out. equipment moves. However, in agricultural Technological stresses on winter oilseed rape practices, the directions of ground equipment can be detected by means of leaf diagnostics movement should change from year to year, and, because in September-November there is an accordingly, the distribution of stress areas may abnormal color of the lower leaves, which is easy differ from the direction of crop rows. to establish both by ground visual assessment and Accordingly, the identification of stresses can be research using UAVs. For winter wheat, such based on the assessment of the contour of the manifestations suitable for reliable identification stress section, which for technological stresses from the UAV platform on an industrial scale must have the correct geometric shape inherent (height from 60 meters) could not be detected. In exclusively in artificial objects. In particular, in ground-based monitoring, it was noted that plants the case of chemical poisoning of plants, the have a characteristic deformation of the leaves, boundary between affected and healthy crops will which can be a characteristic criterion for be directly linear. identifying the nature of stress. Affected areas inside the field are more dangerous for industrial fields, which are difficult, often impossible to 2.2. Choice of the software visually detect by ground monitoring means. This situation is extremely dangerous, as areas with environment weakened plants appear in the field, which is more susceptible to pests and can become centers for Assessing the nature of stress for crops is an the spread of the latter. Accordingly, it is urgent task to be solved both by agronomists advisable to develop a technology that will directly in the fields and by relevant specialists identify stress areas regardless of their location. using cloud services. Accordingly, for the This was taken into account when choosing the versatility of the operating system used, it is experimental production field. advisable to use a cross-platform programming language such as Python, which is adapted to the 2.4. Analytical research fate of large data processing and machine learning. In the work of Emad Ebeid and others (2018) [18], devoted to the review of flight Laboratory studies accompanied all stages of plantation monitoring. A sampling of plants and soil was performed on the day of monitoring or 3.2. Data processing within two days thereafter. Soil samples were taken from a layer of 0-25 cm, prepared for 3.2.1. Evaluation of the contour of the analysis according to DSTU ISO 11464: 2007. map Agrochemical analysis was performed in scientific and research laboratories of the To manage the harvest, farms, regardless of Department of Agrochemistry and Plant Product weather conditions, need maps of the distribution Quality, Ukrainian Laboratory of Agricultural of vegetation indices in many production fields Products Quality, in compliance with accepted available on the farm. Based on these methods and techniques. circumstances, the Slantrange sensor complex was created to survey up to 10,000 ha/day, which can be provided on aircraft platforms. Since the 3. Algorithm for identification of average area of production fields in the plains of Ukraine is 70-100 hectares, it is desirable to technological stresses, its software survey several fields at once during one flight. In implementation, and results of the analysis of a particular field, it is necessary to experimental data processing determine its boundaries. The Python-supported OpenCV library contains ready-made procedures 3.1. Select the source data format for finding the contours of graphical objects that can be used in this case. An example of the result To process spectral monitoring data, the of card processing is shown in Figure 1. Slantrange sensor system has its own Slantview software, which allows you to save the received maps in several data formats, namely Shapefile, KMZ, GeoTiff. The shape format contains attributive information of geometric objects and is designed primarily to create tasks for ground equipment. KMZ files are 3D data in Google Earth and represent a map of the distribution of vegetation indices on satellite images. According to the results of experimental studies on the recognition of the values of vegetation indices, it was found that the data was distorted during the overlay of the images - the recorded colors were Figure 1: Green Chlorophyll index distribution missing in the palette for the specified vegetation map (a) for wheat crops affected by the after- indices. Probable explanation in image correction effects of herbicides, (b) - a photograph in for overlay on the satellite image to facilitate visual perception by the user. By comparing the pseudo-colors of the area highlighted by the data for the distribution points of the distribution square on the map, and the results of the field map from the working window of the SlantView contour search by the proposed software (c) program, it was found that for the GeoTiff format color distortion and, consequently, the values of It should be noted that the forest-steppe zone vegetation indices do not occur. Unlike the KMZ of Ukraine is characterized by strong forest belts, format, the file does not have positioning labels, the leaf cover of which, as well as the shadow but when you save the map, the program retains from them are also fixed by the system. From the the scaling, and, accordingly, when using available experience, maintenance of forest belts landmarks, the calculation of positioning is quite and their renewal is not carried out regularly and possible. In view of the above, the GeoTiff format there are many cases when tree crowns was adopted for analysis. completely cover the road surface due to which significant errors are possible in determining the contours of crops. Since the field boundaries are stable, to analyze the presence of technological stresses on the maps of the distribution of vegetation indices stored in the Geotiff format, it is advisable to enter them manually, using certain amount of data that does not belong to the field reference points. under study, it is necessary to change the orientation from geographical to local reference to 3.2.2. Estimation of the orientation of the dimensions of the field. Due to the change in the orientation of the image, the number of pixels the sections of the field caused by of the image obtained from the GeoTIFF file stresses of technological character decreased from 1100 × 1660 to 245 × 1521, ie the amount of data decreased almost 5 times. With the identification of crop rows, the Figure 3 shows the interface of the developed direction of crop rows was stable, but this is not a program in python to identify stressful areas of prerequisite for technological stresses. Thus, technological nature. Figure 1 (a) shows the presence of a green band on the left and top, which for this index Green Chlorophyll index corresponds to healthier crops than those with yellow. This condition may be due to the best condition of mineral nutrition at the field boundaries because it is there that the equipment slows down, turns, adjusts the operation of nozzles, augers, and more. The width of such a layer, as a rule, does not exceed the radius of reversal of ground equipment, which can be taken into account when analyzing the distribution of stress areas The distribution in the field of stress areas caused by phytotoxic action (aftereffect) of Figure 2: Distribution maps of stress areas for the herbicides, as well as violation of the seeding rate, GreenNDVI index for winter wheat crops from is related to the direction of technological tracks, April 27, 2019. Image of the Slantview software the organization of which meets certain rules. This map window (left) and converted from a GeoTIFF is due to the fact that the introduction of chemical reagents or seeds during sowing is not carried out file to jpeg format in an arbitrary manner, namely in compliance with the laid technological tracks. The directions of technological tracks in one field can change from year to year to maintain soil fertility, but their number usually does not exceed 2, in some cases 3 directions. Soil loosening can be carried out in any order, but technological stresses cannot be caused by this operation. Determining the direction of technological tracks has certain prerequisites, so mechanics when planning work are interested in the maximum length of the runs. Accordingly, in the absence of data from technological maps for the implementation of mechanical tillage, the orientation of the Figure 3: Picture of the program command-line experimental field should be carried out along the interface when entering analysis parameters maximum length of the field. Figure 2 presents a map of the distribution 3.2.3. Convert data from color format of stress areas for winter wheat crops where to numeric view chemical poisoning of winter wheat crops as a result of the after-effects of herbicides from Figure 4 shows the map window palette, which the predecessor crop was recorded. is used to encode data and save them in tiff format The specificity of SlantView software data according to the method presented in S. Shvorov processing is the observance of the north-south et all (2020) [19]. Since the NDVI indices for geographical orientation. As a result, to reduce the plants change in the range 0… 1 for visualization in the 8-bit color model, the index values were Figure 5: Image of the command line interface of multiplied by 255 (Fig. 4) the program during the program with the output of intermediate results of the analysis (in areas highlighted in green probable stress of a technological nature) 3.2.6. Filtering of the received data Figure 4: The image of the Green NDVI map is Previous studies have found that stress should listed from the Slantview software palette be determined by the magnitude of the standard (encoding "shades of gray" where the index value deviation. For filtration, a limit value was set at is multiplied by 255. which the stress status was set for the plots. 3.2.4. Image segmentation 3.2.7. Graph analysis for in-depth search To assess the presence of stress of a technological nature, the field image was divided The stressful state of plantations is caused by into separate sections. The size of the plot was chemical poisoning of plants or their thickening determined based on the resolution of the due to non-compliance with production distribution map and the standard nomenclature of technology when moving ground equipment. ground equipment available on the farm. The size Accordingly, stress areas will form bands. The was 13 × 13 pixels (6.5 × 6.5 m). DFS (Depth-first search) method was used to identify such stress areas. That is, single 3.2.5. Calculation of distribution manifestations of plant stress due to differences from the total mass of the water supply regime, parameters etc. are not taken into account. The results obtained are presented in Figure 6. The GaussAmp equation was used to The developed software passed a production approximate the experimental data on the color test, during which its accuracy and selectivity intensity when color-coding the values of the were confirmed. intensity of the GreenNDVI index. Determined the value of the average value. Figure 5 shows the image of the program interface when indicating the intermediate results for statistical processing of the distribution of index values on the map segment. Figure 6: Picture of the command line interface of the program during the output of the analysis results 4. Conclusions High-resolution maps of high-resolution stress indices can be considered as a separate object of study on the interpretation of the causes of the [5] Andrea S. Laliberte (2007) Combining stress of complex biological objects, such as Decision Trees with Hierarchical Object- winter crops. For remote monitoring of the state oriented Image Analysis for Mapping Arid of winter crops, an algorithm for the identification Rangelands / Andrea S. Laliberte, Ed of technological stresses has been developed on L.Fredrickson, Albert Rango // the basis of the spectral-spatial analysis of the Photogrammetric Engineering & Remote nature of the location of stress areas. The Sensing, Vol.73(2), pp. 197-207(11), algorithm is implemented in the developed https://doi.org/10.14358/PERS.73.2.197. software for spectral-spatial analysis of stress [6] Mitch Bryson (2010) Airborne vision‐based index maps to identify stress areas due to mapping and classification of large farmland technological factors. environments / Mitch Bryson, Alistair Reid, It has been experimentally confirmed in the Fabio Ramos, Salah Sukkarieh // Special production fields that the use of the developed Issue: Visual Mapping and Navigation software allows identifying the contours of areas Outdoors, Vol.27 (5), рр. 632-655, of plants with stresses of a technological nature on https://doi.org/10.1002/rob.20343. the basis of stress index distribution maps. [7] Passang Dorji (2017) Impact of the spatial resolution of satellite remote sensing sensors 5. References in the quantification of total suspended sediment concentration: A case study in turbid waters of Northern Western [1] M. Lotfi (2009) Combining wavelet Australia / Passang Dorji, Peter Fearns // transforms and neural networks for image PLoS ONE, Vol.12(4), e0175042. classification / M.Lotfi, A.Solimani, https://doi.org/10.1371/journal.pone.017504 A.Dargazany, H.Afzal, M.Bandarabadi //. 2. System Theory, Ssst 2009. Southeastern [8] J. Senthilnath (2017) Application of UAV Symposium on. IEEE. pp. 44–48. imaging platform for vegetation analysis https://doi.org/10.1109/SSST.2009.4806819 based on spectral-spatial methods / . J.Senthilnath, M.Kandukuri, A.Dokania, [2] S. Shvorov (2018) UAV Navigation and K.N.Ramesh // Computers and Electronics in Management System Based on the Spectral Agriculture. Vol.140, pp.8–24; Portrait of Terrain / S.Shvorov, https://doi.org/10.1016/j.compag.2017.05.02 D.Komarchuk, N.Pasichnyk, О.Opryshko, 7. Y.Gunchenko, S.Kuznichenko // 2018 IEEE [9] M. Dolia (2019) Information Technology for 5th Int. Conf. on Methods and Systems of Remote Evaluation of after Effects of Navigation and Motion Control (MSNMC), Residues of Herbicides on Winter Crop Rape – Proceedings. рр. 68-71, / M.Dolia, V.Lysenko, N.P asichnyk, http://dx.doi.org/10.1109/MSNMC.2018.85 O.Opryshko, D.Komarchuk, V.Miroshnyk, 76304. T.Lendiel, A.Martsyfei // 2019 3rd [3] Xianlong Zhang (2019) New research International Conference on Advanced methods for vegetation information Information and Communications extraction based on visible light remote Technologies, AICT 2019 - Proceedings рр. sensing images from an unmanned aerial 469-473. vehicle (UAV) / Xianlong Zhang, Fei Zhang, https://doi.org/10.1109/AIACT.2019.88478 Yaxiao Qi, Laifei Deng, Xiaolong Wang, 50. Shengtian Yang // International Journal of [10] Yu-Hsuan Tu (2020) Optimizing drone flight Applied Earth Observation and planning for measuring horticultural tree Geoinformation, Vol.78, рр.215-226, crop structure / Yu-Hsuan Tu, Stuart Phinn, https://doi.org/10.1016/j.jag.2019.01.001. Kasper Johansen, Andrew Robson, Dan Wu [4] S. W. Myint (2013) Per-pixel vs. Object- // ISPRS Journal of Photogrammetry and based classification of urban land cover Remote Sensing, Vol. 160, рр. 83–96, extractio n using high spatial resolution https://doi.org/10.1016/j.isprsjprs.2019.12.0 imagery / S.W.Myint, P.Gober, A.Brazel, 06. S.Grossman-Clarke, Q. Weng // Remote [11] Ming-Der Yang (2020) Adaptive Sens. Environ. Vol.115 (5), рр.1145–1161, autonomous UAV scouting for rice lodging https://doi.org/10.1016/J.RSE.2010.12.017. assessment using edge computing with deep learning EDANet / Ming-Der Yang, Jayson [17] Carlos Henrique Wachholz de Souza (2017) G. Boubin, Hui Ping Tsai, Hsin-Hung Tseng, Mapping skips in sugarcane fields using Yu-Chun Hsu, Christopher C. Stewart // object-based analysis of unmanned aerial Computers and Electronics in Agriculture, vehicle (UAV) images / Carlos Henrique Vol.179, 105817, Wachholz de Souza, Rubens Augusto https://doi.org/10.1016/j.compag.2020.1058 Camargo Lamparelli, Jansle Vieira Rocha, 17. Paulo Sergio Graziano Magalhães // [12] Wojciech Gruszczyński (2019) Application Computers and Electronics in Agriculture, of convolutional neural networks for low Vol.143, рр. 49–56, vegetation filtering from data acquired by http://dx.doi.org/10.1016/j.compag.2017.10. UAVs / Wojciech Gruszczyński, Edyta 006. Puniach, Paweł Ćwiąkała, Wojciech Matwij [18] Emad Ebeid (2018) A Survey of Open- // ISPRS Journal of Photogrammetry and Source UAV Flight Controllers and Flight Remote Sensing, Vol.158, рр.1-10, Simulators / Emad Ebeid, Martin Skriver, https://doi.org/10.1016/j.isprsjprs.2019.09.0 Kristian Husum Terkildsen, Kjeld Jensen, 14. Ulrik Pagh Schultz // Micro- processors and [13] Qi Yang (2019) Deep convolutional neural Microsystems, networks for rice grain yield estimation at the https://doi.org/10.1016/j.micpro.2018.05.00 ripening stage using UAV-based remotely 2. sensed images / Qi Yang, Liangsheng Shi, [19] S. Shvorov (2020) The method of Jinye Han, Yuanyuan Zha, Penghui Zhu // determining the amount of yield based on the Field Crops Research, Vol.235, рр. 142-153, results of remote sensing obtained using https://doi.org/10.1016/j.fcr.2019.02.022. UAV on the example of wheat / S. Shvorov, [14] V. Lysenko (2017) Determination of the not V. Lysenko, N. Pasichnyk, Y. Rosamakha, uniformity of illumination in process А. Rudenskyi, V. Lukin, A. Martsyfei // 2020 monitoring of wheat crops by UAVs / IEEE 15th International Conference on V.Lysenko, D.Komarchuk, O.Opryshko, Advanced Trends in Radioelectronics, N.Pasichnyk, N.Zaets, // 2017 4th Telecommunications and Computer International Scientific-Practical Conference Engineering (TCSET), рр. 245-248. Problems of Infocommunications Science http://dx.doi.org/10.1109/PICST47496.2019 and Technology, PIC S and T 2017 – .9061238Wang, Xin, Tapani Ahonen, and Proceedings. Volume 2018-January, 28 June Jari Nurmi. "Applying CDMA technique to 2017, Pages 265-267. network-on-chip." IEEE transactions on very https://doi.org/10.1109/INFOCOMMST.201 large scale integration (VLSI) systems 15.10 7.8246394. (2007): 1091-1100. [15] Yan Pang (2020) Improved crop row detection with deep neural network for early- season maize stand count in UAV imagery / Yan Pang, Yeyin Shi, Shancheng Gao, Feng Jiang, Arun-Narenthiran Veeranampalayam- Sivakumar, Laura Thompson, Joe Luck, Chao Liu // Computers and Electronics in Agriculture, Vol.178, 105766, https://doi.org/10.1016/j.compag.2020.1057 66. [16] Junfeng Gao (2018) Fusion of pixel and object-based features for weed mapping using unmanned aerial vehicle imagery / Junfeng Gao, Wenzhi Liao, David Nuyttens, Peter Lootens, Jürgen Vangeyte,Aleksandra Pižurica, Yong He, Jan G. Pieters // International Journal of Applied Earth Observation and Geoinformation Vol. 67, рр.43-53, https://doi.org/10.1016/j.jag.2017.12.012.