Analysis of deep learning architecture for patch-based land cover classification

In recent years, the usage of computer vision methods for mapping land cover and land use has increased. Patch-based change detection may produce much superior results than pixel-based change detection and generate accurate change maps. Convolutional Neural Network (CNN) is an excellent option for remote sensing applications using hyperspectral data. The training of models using deep learning techniques takes a lot of time. This research aims to determine the most efficient deep learning model for patch-based land-cover classification by examining and comparing three CNN architectural models for land cover classification: LeNet-5, VGG-16, and ResNet-50. EuroSAT data derived from Sentinel-2A remote sensing imagery are utilized in this work. Comparing the three CNN architectures indicates that ResNet-50 has the highest validation accuracy, with a testing accuracy of 0.877, and a training time that is neither too quick nor too slow. The LeNet-5 model has the quickest training time but the lowest accuracy. VGG-16 has the longest training period yet has the highest test score of 0.878.