Hybrid PSOBAGA-Tuned IARUMV2 for Segmentation of High-Resolution Remote Sensing Images

Abstract

The complexity of spatial structures as well as fine boundary details are present in the high-resolution remote sensing images (HR-RSI). Due to the abovementioned characteristics, the segmentation of HR-RSI is a typical task. Enhanced U-Net models and traditional techniques are still facing problems like suboptimal feature extraction and low convergence rates for segmenting remote sensing images. In order to overcome the aforementioned challenges, this work suggested an improved U-Net architecture named Inception Attention Residual MobileNetV2 U-Net (IARUMV2). IARUMV2 incorporates inception modules, attention mechanisms, residual links, and a MobileNetV2 encoder to acquire multi-scale contextual information very effectively. For fine-tuning the hyperparameters of IARUMV2, a new hybrid optimization algorithm is introduced, which combines the Particle Swarm Optimization (PSO), Bat Algorithm (BA), and Genetic Algorithm (GA) in order to balance the exploitation, exploration, and convergence speed. The model is tested on two datasets, i.e., Massachusetts Buildings and WHU datasets. Experimental results on the Massachusetts Buildings and WHU Building datasets demonstrate the effectiveness of the proposed PSOBAGA-tuned IARUMV2 framework. The model achieves segmentation accuracies of 96.67% and 98.26%, respectively, with corresponding IoU values of 83.04% and 93.19%. Compared with existing approaches such as U-Net, MATU-Net, TCNet, and SegNet, the proposed method consistently achieves higher accuracy, precision, recall, and F1-score while exhibiting faster convergence and improved hyperparameter optimization efficiency showing its effectiveness for HR-RSI segmentation.