High-Spatial-Resolution Remote Sensing Scene Method Based on Spatial Adaptive Convolution

The High-Spatial-Resolution (HSR) remote sensing is categorized through its ability to generate fine-grained spatial data. The image resolution is typically calculated through Ground Sample Distance (GSD) that presents the smallest ground object size that can be solved in an images. This paper presents a remote sensing image semantic segmentation for resolving issues through Hierarchical Atrous Pyramid (HASP) and Spatial Adaptive Network (SPANet) based framework. The HASP is utilized to resolve the issues of difficult background and high-scale variances through expanding the receptive domain of network. The SPANet has a capability to adjust convolutional operations adaptively that reduces the computational load. It can lead to quicker inference time and make it is feasible for real-time applications. The loss function is calculated which assists network to enhance the segmentation result. The parameters like accuracy, Intersection over Union (IoU), mean IoU (mIoU) and mean F1 (mF1) are utilized to estimating model performance on iSAID dataset. The SPANet obtains high accuracy of 98.74%, IoU of 84.68%, mIoU of 72.26% and mF1 of 81.59% which is comparatively higher than existing methods like Foreground-Aware Relation Network (FarSeg++) and Cascade Class-Aware Enhanced Network (CCENet).