Technology of intelligent agricultural crops recognition by neural network based on multispectral multitemporal Earth remote sensing data

Abstract

This article presents a technology for recognition of crops based on Earth remote sensing data. Approbation of the technology was carried out on agricultural lands of the Belgorod region, Russian Federation. According to the crop rotation statistics, the following crops were selected: wheat, barley, corn, sunflower, soybean, sugar beet, perennial grasses and fallow lands. As input data, multispectral images from the Sentinel-2 satellite of the MSI L2A processing level were used, namely the channels of the RGB, SWIR and NIR bands. From the data obtained, time series were compiled in the growing seasons for 2018-2020. An algorithm has been developed to eliminate data gaps on days of high cloudiness to improve the quality of recognition. A convolutional-recurrent neural network was used as a classifier model. The developed model on test data showed an overall measure of accuracy F-score of 88.7%. The proposed architecture of the neural network is also applicable in other regions with a similar sowing structure, phenological phases of crops and similar climatic conditions.