Multi-Angle Imaging and Machine Learning Approaches for Accurate Rice Leaf Area Estimation

Rice leaf area is a critical physiological metric that indicates photosynthetic efficiency, energy conversion, and dry matter accumulation capacity. This study aimed to develop a simple and efficient rice leaf area imaging system and prediction method, so as to provide a theoretical foundation and technical support for rapid and accurate leaf area measurement.

The study utilized representative rice varieties—Xiushui 134 (indica), Huanghuazhan (japonica), and Yongyou 1540 (indica-japonica hybrid)—as experimental materials. Leaf area data were collected from the aboveground parts during critical growth periods, and both flat-overhead-view and side-view images were captured. Using the PlantScreen high-throughput modular plant phenotyping platform, morphological and color feature information was extracted. Based on these data, various feature selection methods (Pearson correlation coefficient, maximal information coefficient (MIC), and recursive feature elimination (RFE)) combined with machine learning models (support vector regression (SVR), random forest regression (RFR), and XGBoost) and deep learning models (ResNet50, AlexNet, VGG, and SeNet) were employed to develop a simplified and efficient rice leaf area prediction model.

(1) An imaging approach that integrated flat-overhead and multi-angle side views significantly outperformed single-view methods for leaf area prediction, with R² values of 0.76–0.82 and coefficients of variation (CV) of 5.5%–13.7%, compared with R² values of 0.51–0.78 and CVs of 9.7%–27.5% for single views. The optimal system used one flat-overhead-view and one side-view image, achieving R² = 0.79, root mean square error (RMSE) = 95.3, mean absolute error (MAE) = 77.02, and CV = 6.5%. (2) Using MIC algorithm for key feature selection combined with the random forest regression model achieved excellent results (R² = 0.84, RMSE = 81.8, and MAE = 63.3), noticeably outperforming other machine learning models. The deep learning model SeNet (R² = 0.80, RMSE = 98.1, and MAE = 74.7) outperformed traditional ResNet50 and AlexNet models but showed no significant advantage over the MIC-RFR model. (3) Feature analysis indicated that the projected area and plant height from side-view images, as well as leaf perimeter and green-yellow characteristics from flat-overhead-view images, significantly contributed to leaf area prediction. The contribution of the side-view projected area (+117.4) was substantially greater than that of other features (ranging from 1.48 to 18.87).

This study employed a simple and efficient leaf area prediction imaging system (one flat-overhead-view combined with one side-view image), integrated with the MIC-RFR model, to meet the high-precision and stable prediction requirements for individual rice leaf area. This method provided a powerful tool and technical support for precision agriculture and crop breeding.