Remote Sensing Image Segmentation Methods Based on Deep Learning Models

Abstract

Remote sensing image segmentation plays a pivotal role in Earth observation tasks by transforming raw satellite and aerial imagery into meaningful semantic regions. This process underpins numerous applications, such as urban planning, precision agriculture, disaster response, and ecological monitoring. With the advent of deep learning, segmentation accuracy has improved significantly due to the capacity of neural networks to learn complex spatial and semantic representations. This paper presents a comprehensive comparative study of three representative deep learning models — U-Net, SegNet, and DeepLabv3+ — applied to the ISPRS Potsdam dataset. We analyze performance across various dimensions, including segmentation accuracy, efficiency, robustness to noise, parameter complexity, and category-specific behaviors. Furthermore, we propose a hybrid model architecture that fuses U-Net’s spatial detail preservation with DeepLab’s contextual aggregation capabilities. To address label scarcity and enhance generalization, we incorporate self-supervised pretraining and transfer learning strategies. We also provide preliminary benchmarking with Transformer-based models. The findings contribute to the body of knowledge guiding the design and deployment of segmentation models in real-world remote sensing scenarios.