A dual pixel-level and subpatch-level network with cross-temporal super resolution for change detection across spatial resolutions

Remote sensing image change detection is critical for monitoring earth surface dynamics. Although deep learning has significantly improved change detection performance, traditional and existing deep super-resolution techniques for cross-resolution change detection often assume bi-temporal images share the same resolution, at least in the training phase. They also suffer from limitations including dependency on expensive high-resolution paired training data, suboptimal performance transfer from super-resolution to change detection accuracy, and heavy reliance on extensive pixel-level annotations. To address these limitations, we propose a novel dual pixel-level and subpatch-level network with cross-temporal super resolution (DPSNet) for change detection across spatial resolutions. Our method DPSNet, comprises two core components: 1) a Reference Image-Guided Generative Adversarial Network (RefIGM GAN) for cross-temporal super resolution; and 2) a Semi-supervised Dual-Path Network (SDNet) for pixel-level and subpatch-level change detection. A resource-efficient alternating optimization strategy is employed between RefIGM GAN and SDNet, creating a virtuous cycle in which super-resolution improves detection accuracy, and detection results optimize super-resolution reconstruction. Experiments were conducted on three datasets, i.e., CDD, SYSU, and HTCD, each characterized by distinct resolution variations. The CDD and SYSU datasets include bi-temporal images with 4 × and 8 × resolution differences, respectively, while the HTCD dataset contains both satellite and UAV imagery with inherent resolution disparities. The results demonstrate that by integrating reference-guided super resolution and semi-supervised learning, effective cross-resolution change detection can be achieved with only limited high-resolution data and pixel-level labels, showing great practical significance in scenarios where solely low-resolution historical images are available. Our source code will be released at https://github.com/Flandre7155/DPSNet .