SCD-MoE: Adaptive Expert Routing for Efficient Remote Sensing Semantic Change Detection

Abstract

Semantic change detection (SCD) aims to simultaneously localize changed regions and identify land-cover transitions from bi-temporal images. However, existing methods usually process bi-temporal features with fixed interaction structures or shared fusion functions, providing limited flexibility to explicitly adapt interaction preferences to diverse change patterns. In addition, SCD requires both change localization and semantic recognition, which emphasize different information cues: the former relies more on discrepancy-sensitive features, while the latter requires semantic-preserving and region-consistent representations, yet such task-specific preferences are often insufficiently modeled. To address these challenges, we propose SCD-mixture-of-expert (MoE), an efficient dynamic framework that leverages conditional computation to model sample-adaptive interactions and task-specific feature preferences. First, we introduce an interaction-adaptive MoE (IA-MoE), which routes bi-temporal features to difference-aware and semantic-preserving experts for adaptive interaction modeling. Through selective expert weighting, IA-MoE enhances change-related cues while maintaining semantic context. Second, we develop a task-aware routing MoE (TAR-MoE), which employs task-specific gates over a shared expert pool to generate localization- and recognition-oriented representations. This design enables controlled information sharing while allowing each task to emphasize suitable feature transformations. Experiments on three benchmarks show that SCD-MoE consistently achieves superior performance, including a 5.04% separated Kappa coefficient (SeK) improvement on Landsat-SCD, while using only 48% of the parameters and substantially fewer FLOPs compared with the previous state-of-the-art method. Code and pretrained models will be made available at https://github.com/HuQ1an/SCD-MoE