Wheat yield prediction using integrated optical and radar remote sensing with machine learning across key phenological stages.

Abstract

Precise and timely prediction of wheat yield is pivotal for ensuring global food security and optimizing agricultural management practices, particularly through advanced remote sensing and machine learning techniques. In this study, wheat yield was accurately estimated by leveraging remote sensing-derived soil and vegetation indices. Yield data from 189 study points were collected, and Sentinel-2 (10-meter resolution) imagery from the Tillering and Anthesis growth stages was used. The models incorporated 25 variables, including 16 optical indices (e.g., NDVI, SAVI, MSAVI2) and three topographic factors. The key novelty of this research is the rigorous comparison of the predictive value and synergistic contribution of Sentinel-1 Synthetic Aperture Radar (SAR) data when integrated with Sentinel-2 optical indices within machine learning frameworks. Three machine learning approaches (Multiple Linear Regression (MLR), Support Vector Machine (SVM), and Random Forest (RF)) were employed and evaluated using 70% training and 30% testing subsets. Results revealed that the RF model, leveraging data from the Anthesis phenological stage, exhibited superior performance in wheat yield estimation, achieving an R² of 0.92 and an RMSE of 0.14 ton ha -1 for the training set, and an R² of 0.90 and an RMSE of 0.29 ton ha -1 for the testing set. To enhance model accuracy, Sentinel-1 radar data were integrated into the RF framework. This addition reduced the training set RMSE to 0.13 ton ha -1 but increased the testing set RMSE to 0.33 ton ha -1 , with R² values remaining stable at 0.92 and 0.90 for the training and testing sets, respectively. Variable importance analysis indicated that optical soil and vegetation indices were the dominant predictors. Although the inclusion of Sentinel-1 SAR data offered additional insights, it did not outperform the predictive capacity of optical indices. These findings validate the efficacy of the combined Sentinel-2 remote sensing approach for generating reliable wheat yield forecasts approximately 50 days prior to harvest.

Plant phenotyping relevance

Sentinel-1/2リモートセンシングと機械学習による小麦収量という植物形質の推定が中心で、モデル比較・評価と予測精度検証を実施している。

Code and data availability

植物フェノタイピング解析を再現する公開資産であることを、入力本文と直接リンクから確認できなかったため保留しました。

No evidence-backed public reproduction asset is currently recorded.