WaterMamba: A Hybrid Mamba-Enhanced Detection Transformer for Small River Obstacle Detection in Complex Hydrological Environments

Small river obstacles are difficult to distinguish from reflections, ripples, turbidity, and low-contrast backgrounds, while spatial-only detectors may inadequately preserve small-target detail and long-range water-surface context. We propose WaterMamba,a hybrid Mamba-enhanced detection transformer built on RT-DETR. The model combines a MobileMamba backbone with Multi-Receptive Field Feature Interaction (MRFFI), a Hybrid Spatial-Spectral Mamba (HybridMamba) encoder with Frequency-Domain Attention (FDA), a Triple-Channel Window Block (TCWB) for multi-scale fusion, and an Adaptive Scale-Aware Detection Loss (ASDL). Under a unified RGB-only protocol, WaterMamba improves over RT-DETR-L by 2.8 percentage points in mAP50 and 4.5 points in APS on WaterScenes; the corresponding APS gains on LaRS and IWHR are 4.8 points. Full-evaluation-set COCO-style metrics and controlled ablations establish the quantitative comparison. At the common confidence threshold of 0.3, the qualitative panels complement these statistics by showing the operating-point behavior on difficult small and low-contrast instances; similarity among boxes jointly detected by strong models is expected and does not negate differences in difficult-target recall. WaterMamba contains 28.6 M parameters, requires 96.2 G FLOPs, and reaches 78.4 FPS on an NVIDIA A100; embedded-platform latency remains outside the present evaluation. The source code is available at the repository specified in the Data Availability Statement.