Semantic Segmentation of Remote Sensing Images Using Optimized Pyramidal Dilation Attention Convolutional Neural Network with Transformation Consistency Regularization

Deep neural networks (DNNs) have received a lot of interest recently in the domains of remote sensing (RS) and computer vision. The scarcity of labeled samples in the higher-resolution remote sensing imagery complicates accurate semantic segmentation. The efficacy of traditional supervised learning techniques in real-world applications with a lack of labeled data is limited by their requirement for large, expensive and time-consuming labeled datasets.Inthis paper, semantic segmentation of remote sensing images using an optimized pyramidal dilation attention convolutional neural network with transformation consistency regularization (SSRSI-OPDACNN-TCR) is proposed. Initially, the input RS images are gathered from the Deep Globe land-cover categorization dataset. The gathered image is fed into the transformation consistency regularization (TCR) method to maintain pixel wise consistency in predictions by applying several random transformations to enhance method resilience and accuracy. Then, the image samples are fed into the PDACNN architecture, which improves parameter aggregation and learning by mixing the strengths of labeled and unlabeled samples to semantically segment RS images. Generally, the PDACNN does not express any adaption of optimization techniques to determine optimum parameters to assure exact segmentation. Hence, hermit crab optimization (HCO) is proposed to enhance the pyramidal dilation attention convolutional neural network (PDACNN) classifier that exactly segments the RS images. The proposed SSRSI-OPDACNN-TCR is implemented in Python. The performance of SSRSI-OPDACNN-TCR method attains 12.32%, 17.21%, and 23.41% high accuracy; 14.81%, 22.15% and 23.13% greater F1-score analyzed with existing methods such as transformation consistency regularization using semi-supervised deep learning for semantic segmentation of remote sensing image (TCR-S4GAN-RSIS), consistency-regularized region-growing network for semantic segmentation of urban scenes by point-level annotations (TCR-CRGNet-RSIS), consistency-guided lightweight network for semi-supervised binary change identification of buildings in RS imageries (TCR-PSPNet-RSIS) respectively.