A Pruned Multi-Scale Skip Connected and Asymmetric-Convolution-Based U-Net Model for Land Cover Classification

Recently, land cover classification became an evolutionary study in Remote Sensing (RS) satellite interpretations. While extracting features from satellite images, spatial information causes the overfitting issue. In addition, computational complexity is occurring in land cover segmentation due to the usage of complex architectures. Thus, in this research, Multi-scale Skip connected and Asymmetric-Convolution-based U-Net (MACUNet) with pruning method is introduced to reduce the computational complexity. To reduce the overfitting, Batch Normalization (BN) layer is employed, simultaneously 3D Convolutional Neural Network (3D CNN) for feature extraction. Initially, the input is taken as satellite images from the GID dataset, where spatial and temporal features are extracted using 3D CNN model with BN layer. After that, these features are given as input to the MACUNet method to improve segmentation performance. Then, classification is done with class_num value utilization based on identified pixel region maps and classified into six classes such as build up, forest, water, farmland, meadows and others. From the results, the proposed pruned-MACUNet model produces better results than existing GeohashNet in terms of Mean Intersection over Union (MIoU) as 80.45% and f1-score as 90.23% respectively.