TDNet: Bidirectional LSTM Gated Triple-Decoder Network for Remote Sensing Change Detection
Cui Zhang, Zhouyang Sha, Hailong Wang
Abstract
Remote sensing change detection (CD) identifies land-use/land-cover changes by analyzing multi-temporal imagery. Mainstream methods employ siamese multi-level encoders and a single multi-level decoder: encoders extract multi-scale features, while the decoder fuses them to generate change maps. However, the effective exploitation of multi-scale features in existing methods remains limited, often falling short in fully unleashing the discriminative potential at each scale and in adequately achieving synergistic cross-scale fusion. To address this limitation, we propose the bidirectional LSTM gated triple-decoder network (TDNet), which introduces two innovations: 1. Triple-decoder architecture: We abandon the single-decoder paradigm and construct three cascaded decoder branches of varying depths (2, 3, and 4 layers). Each branch specializes in processing features at its corresponding level—shallow, middle, or deep—fully unleashing the discriminative potential at every scale. 2. Bidirectional LSTM-based multi-scale feature interaction module (BLMIM): Leveraging LSTM’s gating mechanism, BLMIM extends the bidirectional LSTM to handle multi-scale feature interactions: it adaptively retains upstream features via the forget gate, filters current-layer information via the input gate, and controls contributions to subsequent layers via the output gate. This yields interpretable multi-scale interactions and, to our knowledge, constitutes the first adaptation of LSTM to supervised change detection. Comprehensive experiments on LEVIR-CD, CDD, and WHU-CD demonstrate that TDNet surpasses state-of-the-art methods, achieving F1 improvements of 0.60 %, 0.49 %, and 0.50 %, respectively.