EVENT DOSSIER
TRNet proposes a multi-modal segmentation method for high-resolution rice mapping using terrain priors. This method combines 0.5-meter Gaojing-1 RGB images with 5-meter TanDEM X digital elevation models to suppress steep slope clutter through low-frequency modulation of terrain conditions and asymmetric high-frequency adjustment. It also utilizes a terrain-guided decoder to integrate semantic and structural information. On the Area A and Area B test sets, this method increased the rice IoU to 85.10% and 80.68%, respectively, by 9.15 and 18.83 percentage points compared to the original dual-encoder U-Net. Experiments covering internal test sets, geographically independent regions, and cross-year seasonal analyses verified the effectiveness of terrain as a stable context prior, significantly reducing false positives on steep slopes and missed detections on gentle slopes.
TRNet 提出一种利用地形先验进行高分辨率水稻映射的多模态分割方法，在 Area A 和 Area B 测试集上分别将水稻 IoU 提升至 85.10% 和 80.68%，较原始双编码器 U-Net 提升 9.15 和 18.83 个百分点。该方法结合 0.5 米高景一号 RGB 影像与 5 米 TanDEM X 数字高程模型，通过地形条件低频调制和不对称高频调节抑制陡坡杂波，并利用地形引导的解码器整合语义与结构信息以优化预测。实验涵盖内部测试集、地理独立区域及跨年度季节性分析，验证了地形作为稳定上下文先验的有效性，显著降低了陡坡误报和低坡漏检。