Transfer Learning in Deep Neural Network for Land Cover Classification

Abstract

Land cover classification research contributes significantly to sustainable development by enhancing the baseline for urban planning, natural resource management, and environmental monitoring at the local, regional, national, and global levels. In practice, only the RGB and near-infrared (NIR) image bands are obtainable for identifying land cover since high-resolution satellite remote sensing data may not even be accessible due to cost, storage systems, or bandwidth restrictions. This work presents Transfer Learning in Deep Neural Networks (DNN) for land cover classification, where ResNet101 is considered as a DNN model. Numerical experiments are conducted using the EuroSAT dataset, a multispectral dataset based on Sentinel-2 satellite images of 34 European cities covering 13 spectral bands in visible, near-infrared, and short-wave infrared parts. Several experiments were conducted, and the proposed ResNet101-based technique effectively classifies the land cover classes with more than 99% accuracy.