COMPARATIVE ANALYSIS OF CONVOLUTIONAL NEURAL NETWORKS (CNN) FOR LAND USE CLASSIFICATION BASED ON AGRICULTURAL SATELLITE IMAGES

The use of satellite imagery has grown considerably in recent years for land cover detection, particularly for agricultural land classification. This article aims to address these challenges by applying several machine learning algorithms to multispectral data from the Sentinel-2 satellite to obtain accurate land classification models. This study presents a comparative analysis of five convolutional neural networks (CNN) architectures for the automatic classification of land use from satellite images. The models evaluated include Res Net 50, Mobile Net, VGG16, Google Net and Efficient Net. These were applied to a dataset comprising four land use classes: Annual Crop, Forest, Permanent Crop and Residential. The results show that Mobile Net and Google Net perform best with a validation accuracy of 99%, whereas ResNet50 is limited to 66%. This underlines the importance of using advanced machine learning techniques, including Mobile Net or Google Net, to accurately classify changes in agricultural land use.