DKI-Net: Dynamic Kernel-Interaction Network for Semantic Change Detection in Remote Sensing Images
Abstract
Semantic change detection (SCD) aims to simultaneously locate changed regions and recognize their “from–to” semantic transition types from bitemporal remote sensing images. Recently, Transformer-based architectures have significantly advanced SCD performance by modeling long-range bitemporal dependencies. However, the feedforward network (FFN) in Transformers is input-agnostic, lacks spatial awareness, and lacks explicit modeling of relationships among different response patterns, which may limit the model's ability to capture fine-grained semantic changes in complex remote sensing scenes. To address these issues, we propose DKI-Net, an end-to-end multitask network with a three-branch architecture, whose core component is a kernel-interaction dynamic FFN (KID-FFN). Specifically, KID-FFN consists of three parts. First, a dynamic kernel branch leverages global context to generate input-adaptive $\text{5} \times \text{5}$ spatial filter kernels for each sample and each channel. Second, a static convolution branch combines a bottleneck structure with depthwise convolution to restore local spatial structure. Third, a kernel interaction branch extracts $K$ semantic kernel tokens and performs multihead attention interaction to explicitly model dependencies among different response patterns. In addition, a dual-scale bidirectional change feature enhancement module is introduced to further strengthen change feature representations. Extensive experiments on SECOND and Landsat-SCD demonstrate that DKI-Net achieves superior performance over existing competing methods.