Object-oriented classification of remote sensing earth images using machine

Abstract

The results of research on the development of automated classification of remote sensing images of the Earth for on-farm land use based on the use of an object-oriented approach, machine learning and geoinformation modeling are presented. The classification methodology included three stages: analysis of digital images with the selection of spatial objects through preliminary segmentation, classification of spatial objects using the ,Random Forest (RF) and Support Vector Machine (SVM) machine learning algorithms, and assessment of the overall accuracy of the result. For processing, satellite images Sentinel-2 from May to April for the land use area of the experimental station «Elitnaya» and Individual Enterprise of State Farm (Collective Farm) Kovalev S.M. of the Novosibirsk region with a spatial resolution of 10 m per pixel were used. The processing of the resulting multispectral images was carried out using the software product SAGA GIS version 8.5.1 and QGIS with opensource code, the creation of classification models was carried out in the package of the statistical programming language R. It was established that the overall accuracy of classification of land use objects displayed onsatellite images, for the territory of the experimental station «Elitnaya» the SVM algorithm was 87.1% (kappa coefficient 0.74), and using the RF algorithm – 90.3% (kappa coefficient 0.87). For the land use area of the Individual Enterprise of State Farm (Collective Farm) Kovalev S.M. using the SVM algorithm – 78.4% (kappa coefficient 0.78), and using the RF algorithm – 82.3% (kappa coefficient 0.82). The object-oriented approach, in integration with machine learning, facilitates efficient segmentation and classification of remote sensing images for the delineation of spatial objects, provides the ability to automate the mapping process of land use areas, and to incorporate this information into geoinformation modeling for evaluation and classification of agricultural lands.