DepthCD: Depth prompting in 2D remote sensing imagery change detection

Abstract

Change detection with multi-temporal remote sensing images has wide applications in urban expansion monitoring, disaster response, and historical geographic information updating. In recent years, advancements in artificial intelligence have spurred the development of automatic remote sensing change detection methods. However, the existing change detection methods focus on variations in the spectral characteristics of objects, while ignoring the differences and variations in the Earth surface elevation of the different targets. This results in false alarms and missed detections in complex scenarios involving shadow occlusion, spectral confusion, and differences in imaging angles. In this paper, we present a depth prompting two-dimensional (2D) remote sensing change detection framework (DepthCD) that models depth/height changes automatically from 2D remote sensing images and integrates them into the change detection framework to overcome the effects of spectral confusion and shadow occlusion. During the feature extraction phase of DepthCD, we introduce a lightweight adapter to enable cost-effective fine-tuning of the large-parameter vision transformer encoder pre-trained by natural images. Inspired by domain knowledge of the dimensional correlation in land surface changes, we propose a depth change prompter to explicitly model depth/height changes at the feature, depth, and slope levels. In the change prediction phase, we introduce a binary change decoder and a semantic classification decoder that couple the depth change prompts with high-dimensional land-cover features, enabling accurate extraction of changed areas and accurate change types. Extensive experiments on six public change detection datasets validate the advantages of the DepthCD framework in binary and semantic change detection tasks. Detailed ablation studies further highlight the significance of the depth change prompts in remote sensing change detection.