Investigation of Remote Sensing Images Using Deep Neural Network Analysis

Abstract

Semantic segmentation of Remote Sensing (RS) imagery challenging task due to many factors such as scene complexity, irregular object shapes and size, sizes, visual similarity between different categories, and shaded objects. Semantic segmentation using Deep learning (DL) in RS domain yielded promising results. However, many challenges still need to be addressed. These include addressing highly imbalanced class distributions, efficiently handling large-scale datasets while minimizing computational demands, dealing with occluded road regions, and distinguishing similar texture regions. This study addresses these issues by segmenting natural and non-natural objects in RS imagery resources using four DL models. The first method addresses shaded road extraction without using down sampled architectures. It learns road features by zooming road and background by shrinking the road. Dilated inception and pyramid pooling enhanced shadowed road by 6 percent. These results were further improved in post-processing using geometrical shape analysis. The first model achieves an IoU of 76.65 percent on Massachusetts road dataset as compared to the state of art methods. The convergence of the non-down sampled network is faster than that of the down sampled network. The second model is a variant of SegNet called as DR SegNet. The proposed network catches the shaded road with improved road junction detection. Here, we propose the basic layer as a Dense Residual Block and replace each convolutional layer of SegNet with this basic layer, while the feature variation with each successive down sampling is kept to 64. DR SegNet achieves better completeness, with 20 percent less processing time as compared to the state-of-the-art methodology. Our findings demonstrate that DR SegNet detects occluded roads and improves connectivity at road junctions. Moreover, it exhibits faster convergence, reduced memory requirements, and faster training than the variable number of filters with each successive down sampling.