TRNet: Topography-Guided Segmentation of Paddy Rice from VHR Imagery
A new deep learning model named TRNet has been created by researchers to effectively segment paddy rice fields from very-high-resolution (VHR) satellite images in hilly and mountainous areas. This model combines 0.5-meter GaoJing-1 RGB images with a 5-meter TanDEM-X digital elevation model (DEM) and slope data. TRNet utilizes distinct visual and terrain encoders to maintain unique modality characteristics. An early-stage Topographic Energy-Spectral Rectification module modulates low frequencies based on terrain and regulates high frequencies to reduce clutter from steep slopes while enhancing low-slope rice signals. The Topography-guided Paddy Structure Decoder integrates semantic, boundary, and interior cues. Testing was performed on an internal set (Area A) and a separate Area B, which had steeper terrain and less rice. TRNet demonstrated exceptional segmentation capabilities, especially in difficult terrains, addressing the challenges of differentiating rice from similar-looking vegetation in mountainous environments. The architecture details are published in a paper on arXiv (arXiv:2608.04154).
Key facts
- TRNet is a deep learning model for paddy rice segmentation.
- It uses 0.5-m GaoJing-1 RGB imagery and 5-m TanDEM-X DEM.
- Separate visual and terrain encoders are employed.
- Topographic Energy-Spectral Rectification module is used early in the encoder.
- Topography-guided Paddy Structure Decoder combines semantic, boundary, and interior cues.
- Experiments used Area A (internal test) and Area B (held-out with steeper terrain).
- TRNet achieved superior segmentation performance in challenging topographies.
- Paper available on arXiv with ID 2608.04154.
Entities
Institutions
- arXiv