Abstract

Semantic change detection aims to localize changed regions and identify the corresponding land-cover transitions from bi-temporal remote sensing images, which is crucial for applications such as urban expansion analysis, disaster assessment, and environmental monitoring. Although vision foundation models such as the Segment Anything Model 2 provide strong visual priors and powerful feature representations, directly transferring them to semantic change detection remains challenging. In particular, the high-frequency details required for precise boundary delineation are often weakened during feature extraction, while the joint optimization of binary change localization and semantic recognition can introduce task interference. To address these challenges, we present FreqSCD, a SAM2-based framework built on a frozen backbone with three task-specific components: a High–Low-Frequency Adapter for frequency-aware feature adaptation, Task-Decoupled Decoding and Semantic Consistency for reducing task interference, and Local Spatial–Semantic Alignment for improving multi-scale feature aggregation. Experiments on the SECOND and Landsat-SCD benchmarks show that FreqSCD achieves strong semantic change detection performance, obtaining an F1 score of 56.72% and a SeK of 24.17% on SECOND, as well as an F1 score of 85.46% and a SeK of 53.76% on Landsat-SCD.

IPC Classification

Keywords