TRNet AI maps rice fields with 85% IoU, beating U-Net by 19 points

Mapping paddy rice in mountainous regions is notoriously difficult because terrain alters optical appearance and creates confusion with visually similar vegetation. To tackle this, Kaiwen Xiao and colleagues from the arXiv paper present TRNet, a multimodal segmentation framework that fuses 0.5-meter GaoJing-1 RGB imagery with 5-meter TanDEM-X digital elevation models and derived slope data. The architecture uses separate visual and terrain encoders to preserve modality-specific features, then applies a novel Topographic Energy-Spectral Rectification (TESR) module at an early encoder stage. TESR performs terrain-conditioned low-frequency modulation and asymmetric high-frequency regulation, effectively suppressing steep-slope clutter while enhancing compatible low-slope rice cues. A Topography-guided Paddy Structure Decoder (TPSD) then combines semantic, boundary, and interior cues, using coarse terrain as context to refine segmentation.

TRNet was evaluated on two test sets: Area A (internal) and Area B (held-out, with steeper terrain and lower rice prevalence). It achieved rice intersection-over-union (IoU) scores of 85.10% and 80.68%, respectively—significantly outperforming the original Dual-Encoder U-Net by 9.15 and 18.83 percentage points. Ablation studies and slope-stratified analyses confirmed that the gains come from frequency rectification and structure learning, which notably reduce false positives on steep terrain. The paper demonstrates that coarse topography can serve as a powerful contextual prior for precision agriculture mapping, potentially improving crop monitoring and food security assessment in challenging geographies.