Multi-Layered Fusion Approach (MLFA) for Robust Land Cover Detection and Classification From Remote Sensing Data

Abstract

Nowadays, many countries are focusing on increasing greenery to prevent pollution. Land cover mapping is an essential task in computer vision for monitoring urban development and environmental changes and it applies to managing natural resources. Many current models require guidance to locate true landscapes, leading to the attenuation of true regions. Many existing algorithms have been developed to accurately identify land cover regions to understand urbanization. These algorithms failed to identify the correct regions due to poor image quality. This research primarily focused on presenting the new Multi-Layered Fusion Approach (MLFA) model for classifying land cover regions from satellite images. The proposed MLFA, a combination of the Vision Transformer (ViT) and Capsule Networks (CapsNets), has a significant impact on classification. The pre-trained ResNet-50 model is used to train on land cover image datasets collected from various online sources. The ResNet-50 was trained on multiple regions, including forest, urban, and agricultural land, and shows differences among these regions over the years. The proposed approach also focused on extracting the significant features and fusing them with the classification for effective output. Finally, the proposed MLFA achieved high performance in land cover mapping of forest, urban, and agricultural areas.