AGRO-ECOLOGICAL CLASSIFICATION OF AGRICULTURAL LAND USING MACHINE LEARNING, GIS AND REMOTE SENSING DATA

Abstract

The paper is devoted to agro-ecological classification of agricultural land using modern methods of geo-information data analysis and machine learning.
 Background. There are few works in the literature, which cover the issues of accuracy of machine learning (MLL) models for agro-ecological grouping of agricultural land. A large number of raster information layers are used to improve the accuracy of land classification from satellite images. This considerably increases the time of training and testing MMOs, producing thematic maps of agricultural land classification. This approach requires considerably high computing resources and a considerable amount of computer RAM. Raster GIS data models occupy a much larger volume than vector models. In this regard, research on automated agro-ecological (grouping, classification) of agricultural land using vector GIS models is of practical importance.
 Purpose. The aim of the study is to apply GIS methods, remote sensing (ERS) data and machine learning methods for agricultural land grouping.
 Materials and methods. The materials used were synthesized multispectral high spatial resolution Sentinel-2A images, maps of vegetation indices NDVI (Normalized Difference Vegetation Index), OSAVI (Optimized Soil Adjusted Vegetation Index), EVI2 (Enhanced Vegetation Index2), NDWI (Normalized Difference Water Index), SAVI, PVI, GDVI, MCARI, NDRE, TSAVI; topographic map, ALOS DSM (30 m/pixel) and ALOS PALSAR (12.5 m/pixel) satellite images, soil map and field survey results. Field measurements were carried out using the Triump-2 satellite geodetic receiver and included determination of coordinates of characteristic points of land plot boundaries, relief elements, and soil survey.
 Digital spatial model of land use was created using GIS ArcGIS and QGIS, Python engineering libraries were used in the machine learning process.
 Results. Agro-ecological grouping of lands was realized by the example of the farm "Zerno Sibiri" of Novosibirsk region using the following machine learning methods: Random Forest (RF) method, Decision Tree (DT), k-nearest neighbours method (KNN). The best accuracy is the RF machine learning model. The accuracy of the model averaged 97.9% (with training 99.9%, testing 98.8%, and cross validation 95.0%). The Root Mean Square Error (RMSE) is 0.006: for training sample 0.001; test sample 0.076; validation sample 0.123 respectively). The mean kappa coefficient was 0.97 (1.00 for the training sample; 0.982 for the test sample, and 0.927 for the validation sample). 
 Conclusion. The offered method of agro-ecological grouping of agricultural lands by means of GIS, RS data and machine learning methods enabled to distinguish informative quantitative indicators of the relief. The main essence of the proposed method is to create a machine learning model (MLM) based on a spatial dataset. The spatial dataset is formed using geoinformation analysis methods and includes: geomorphometric maps, maps of agrometeorological parameters, soil map, on-farm land management map and operational-territorial units of land classification. The application of vector data model allowed for agro-ecological grouping of agricultural lands in automated mode, to accelerate labor-intensive process of raster data recognition, to increase objectivity of the work. The suggested method of agro-ecological agro-ecological grouping of lands allows taking into account the totality of relief and soil-ecological conditions indicators with the help of geoinformation analysis methods, remote sensing and machine learning.