Semantic-Aware Dense Representation Learning for Remote Sensing Image Change Detection

Supervised deep learning models depend on massive labeled data. Unfortunately, it is time-consuming and labor-intensive to collect annotate bitemporal samples containing desired changes. Transfer from pretrained effective alleviate label insufficiency in remote sensing (RS) change detection (CD). We explore the use of semantic information during pretraining. Different traditional supervised pretraining that learns mapping image label, we incorporate supervision into self-supervised (SSL) framework. Typically, multiple objects interest (e.g., buildings) are distributed various locations an uncurated RS image. Instead manipulating image-level representations via global pooling, introduce point-level per-pixel embeddings learn spatially sensitive features, thus benefiting downstream dense CD. To achieve this, obtain points class-balanced sampling overlapped area between views using mask. embedding space where background foreground pushed apart, aligned across pulled together. Our intuition resulting semantically discriminative invariant irrelevant changes (illumination unconcerned land covers) may help recognition. large-scale image-mask pairs freely available community for Extensive experiments three CD datasets verify effectiveness our method. Ours significantly outperforms ImageNet pretraining, in-domain supervision, several SSL methods. Empirical results indicate improves generalization data efficiency model. Notably, competitive 20% training than baseline (random initialization) 100% code at https://github.com/justchenhao/SaDL_CD .