Improving rice yield prediction with multi-modal UAV data: hyperspectral, thermal, and LiDAR integration
Rice yield prediction is critical for ensuring food security, particularly in major rice-producing countries like China. While Unmanned Aerial Vehicles (UAVs) equipped with hyperspectral imaging are widely used for yield prediction due to its ability to capture detailed spectral information, they ma...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-07-01
|
| Series: | Geo-spatial Information Science |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/10095020.2025.2535524 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Rice yield prediction is critical for ensuring food security, particularly in major rice-producing countries like China. While Unmanned Aerial Vehicles (UAVs) equipped with hyperspectral imaging are widely used for yield prediction due to its ability to capture detailed spectral information, they may not fully account for key factors such as canopy structure and moisture content. To improve accuracy, integrating Thermal Infrared (TIR) data, which reflects canopy moisture, and Light Detection and Ranging (LiDAR) data, which provides crop height and canopy density, is essential. However, the role of biophysical features provided by these sensors in yield prediction across different phenological stages remains unclear. This study addresses this gap by evaluating the combined use of hyperspectral, TIR, and LiDAR data collected during key rice growth stages: tillering, booting, heading, and filling. Two key questions were explored: (1) Does integrating multi-modal data at multiple phenological stages consistently improve yield prediction accuracy? (2) What is the optimal phenological stage for accurate rice yield prediction at the sub-field scale? Multi-modal information, including 2D/3D spectral indices, textural features, temperature data, and canopy structural attributes, was derived and integrated for rice yield prediction using ensemble Machine Learning (ML) models. Single-temporal and multi-temporal modeling strategies were compared. Results showed that hyperspectral data alone achieved satisfactory accuracy during the booting stage (R2 = 0.806), mainly driven by 2D texture and 3D spectral features. Combining TIR-derived temperature features and LiDAR-derived structural features did not improve early-stage predictions but significantly enhanced accuracy during mid-to-late stages, particularly at heading. The highest prediction accuracy (R2 = 0.837) was achieved using a multi-stage model combining data from the tillering, booting, and heading stages. This study provides valuable insights into optimizing sensor fusion strategies and identifying the most informative growth stages for rice yield prediction. |
|---|---|
| ISSN: | 1009-5020 1993-5153 |