Classification of remote sensing images using deep learning for sugar cane population studies

Abstract

The Sentinel-2 satellite platform launched by the European Space Agency has seen wide application in precision agriculture. Its advantage, in addition to all the technological possibilities, in the case of Cuba, is its free access, which substantially lowers the costs of technological solutions developed on this basis. On the other hand, in the current circumstances, it is necessary to improve the accuracy in the methods of automated estimation of the population of the sugarcane crop, as an essential element for the appreciation of the productions to be obtained. Taking into account the conditions of this crop in Cuba and, in turn, the need for a tool to help decision-making in the companies of the AZCUBA group, we propose the improvement of the computational methodology for estimating the population of the crop. A deep learning-based model based on semantic genesis concepts for Sentinel-2 image classification is presented. The proposed model takes as a characteristic for the classification the normalized difference vegetation index (NVDI). The previous existing methodology is improved. This model, integrated into the methodology, was applied in sugarcane areas of the country. The results were compared with those obtained by crop specialists and with those obtained by the methodology without the proposed model, demonstrating superior accuracy and the feasibility of its application.