Work

Cross-Field Transfer of UAV-Based Winter Wheat Yield Prediction: Model Transferability and Sample Efficiency

Year: 2026

Type: article

Abstract: Unmanned aerial vehicle (UAV) remote sensing is widely used for plot-scale crop yield prediction, but model performance may decline when models are transferred across fields with different crop, environmental, management, and sensing conditions. This study evaluated the cross-field transferability of UAV-based winter wheat yield models using multispectral (MS), RGB, and thermal infrared (TIR) information. Categorical Boosting (CatBoost) and multilayer perceptron (MLP) models were evaluated under five strategies: target-only training (T1), direct transfer (T2), nitrogen-based adaptation (T3), cultivar-based adaptation (T4), and mixed adaptation (T5). Direct transfer increased the normalized root mean square error (NRMSE) from 12.7% to 14.0% for CatBoost and from 12.9% to 15.9% for MLP. Mixed adaptation achieved the best overall performance, with coefficient of determination (R2) values of 0.84 and 0.87 and NRMSE values of 11.7% and 10.9% for CatBoost and MLP, respectively. After controlling for calibration sample size, the difference between T3 and T4 became smaller, indicating that adaptation performance was influenced by both calibration-data composition and sample availability. CatBoost showed diminishing returns near the 40% calibration level in this experiment, whereas MLP continued to benefit from additional target-domain samples. Prediction errors were higher for samples beyond the source-domain yield range, while mixed adaptation reduced prediction errors in both yield-range groups. SHapley Additive exPlanations (SHAP) and sensor-level ablation showed that MS features provided important predictive information, while RGB and TIR provided complementary information. These results indicate that target-domain adaptation can improve cross-field UAV yield prediction and highlight the importance of calibration design, yield-range mismatch, and multi-sensor information when transferring models to new fields.

Source: Remote Sensing

DOI: 10.3390/rs18203484

Authors: Jingjing Wang, Yanhao Liu, Yuncheng Xu, Wentao Wang, Dongyu Cai, Haijun Yan

Institution: China Agricultural University

Language: English

Topic: Remote Sensing in Agriculture

Subfield: Ecology

Field: Environmental Science

Domain: Physical Sciences

SDG: Zero hunger

Open Access status: gold

Estimated APC paid: $3,258