TrSC2Y: A transfer-learning-based model from UAV hyper-spectra imagery for field-scale canola yield prediction by integrating DSSAT with PROSAIL

Abstract

Rapid and accurate acquisition of field crop yield is of great significance for agriculture management optimization, food security and crop productivity. By the non-destructive and high-throughput data acquisition, the unmanned aerial vehicle (UAV) remote sensing has become a key tool for crop growth monitoring. However, the scarcity of in-situ samples poses technical barriers and efficiency challenges to yield model training. This study has developed a new yield estimation framework that integrates process models, optical remote sensing, and transfer learning to improve the stability and accuracy of crop yield estimation under small sample conditions. The DSSAT was calibrated with hyperspectral UAV derived crop growth variables, to describe the spatial-temporal variation of small-scale field winter canola leaf nitrogen content during growing season. Firstly, a process-interpretative crop yield estimation framework, TrSC2Y, was pre-trained using the PROSAIL radiative transfer model and the DSSAT crop growth model. Secondly, TrSC2Y was fine-tuned using field observations and UAV hyper-spectra images from three-years canola experiment. Finally, the actual performance and application potential of fine-tuned TrSC2Y in canola yield estimation were evaluated with machine learning as a benchmark test. The results show that: (1) Pre-trained by the crop spectra dataset (from PROSAIL) and yield dataset (from DSSAT), TrSC2Y can accurately extract crop phenotype parameters from theoretical canopy spectra. The joint use of phenotype parameters from multiple growth stages can achieve the best yield estimation (R²= 0.98;RMSE= 33.07 kg/ha;MAE= 1.26 %);(2) Fine-tuned TrSC2Y can be transferred to the field winter canola yield estimation task and shows stable performance (R²= 0.86;RMSE=224.42 kg/ha;MAE=6.5 %). Compared with the machine learning benchmark test, the demand of modeling samples for TrSC2Y is reduced by 50 %; (3) TrSC2Y supports the visualization of field-scale winter canola yield and captures the spatial variability of winter canola yield caused by irrigation-fertilizer treatments.The above results provide a lightweight, cost-effective, and innovative method for field crop yield estimation, promoting the development of precision agriculture management and intelligent applications.

Plant phenotyping relevance

UAVハイパースペクトル画像と作物モデル・転移学習を統合し、収量などの植物表現型を抽出・推定する枠組みを開発・評価しており、取得・推定手法が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.