Multi-sensor fusion using random forests for predicting plant height and yield in radish cultivation
In this study, we developed data-driven models for daily plant height prediction and harvest-day yield prediction of radish by integrating crop phenotypic and environmental sensor data. We applied Random Forests (RF) to predict plant height on the next day as well as the minimum, mean, and maximum yield values on the harvest day. We also assessed the variable importance and response curves to better understand the importance of each variable and the yield response to a given condition, respectively. The input data include time-series crop data (e.g., plant height and estimated SPAD value) obtained from iPad LiDAR-based point clouds and environmental data (e.g., air temperature, relative humidity, dew point temperature, solar radiation intensity, and irrigation amount) measured by multiple sensors. The RF model achieved high prediction accuracy, with the Pearson’s correlation coefficient (COR) exceeding 0.94 and the root mean squared error (RMSE) as low as 2.44 cm for plant height prediction. In yield prediction, models incorporating root cracking information outperformed those without, yielding COR values exceeding 0.85, compared to 0.47 without such data. Variable importance indicated that plant height on the previous day was a significant variable for plant height prediction, while yield prediction was influenced by cultivar, plant height, irrigation amount, and root cracking status. These findings demonstrate the effectiveness of multi-modal sensing and machine learning-based information fusion for predicting belowground traits from aboveground and environmental information, contributing to non-destructive, precise crop monitoring and yield forecasting in root crops.
点群解析によって得られる前日の草高などの作物の情報や気温や相対湿度といった物理環境情報を統合して入力変数とし、草高および収量を予測するデータ駆動型の作物モデルを開発しました。
草高予測では、ピアソン相関係数が0.94、RMSEが2.44 cmと高い精度を達成しました。また、裂根情報を作物の入力情報とすることで、ピアソン相関係数が0.85といった収量予測が可能でした。
これらの知見は、マルチモーダルセンシングと機械学習に基づく情報融合が地上部および環境情報から地下部の特性を予測する上で有効であることを示しており、根菜類における非破壊的で高精度な作物モニタリングと収量予測に貢献します。
詳細は、論文をご覧ください。
Yuto Kamiwaki, Shinji Fukuda. Multi-sensor fusion using random forests for predicting plant height and yield in radish cultivation. Smart Agricultural Technology. 101436. 2025. https://doi.org/10.1016/j.atech.2025.101436
