Predicting rice crop height from field and Sentinel-2 data

Abstract

Improving agricultural methods, guaranteeing food security, and optimizing resource utilization depend on the accurate prediction of crop development and height. This study evaluates rice crop height in Tianjin and Linhai, China, by integrating Sentinel-2 vegetation indices with machine learning and deep learning algorithms. During multiple sampling seasons, data were gathered from 67 fields in Tianjin and 38 fields in Linhai. Each field contained 5 to 8 measurement locations. We gathered the height of the rice plants, the chlorophyll content (Soil Plant Analysis Development, SPAD), and the Leaf Area Index (LAI). We utilized Sentinel-2 imagery to derive vegetation indices such as the Canopy Nitrogen Content (CNC), the Green Normalized Difference Vegetation Index (GNDVI), and the Canopy Chlorophyll Content Index (CCCI). The indices and field data were utilized to train machine learning models such as Extreme Gradient Boosting (XGBoost), Ridge Regression, and Support Vector Regression (SVR), in addition to deep learning models like TabNet and Multilayer Perceptron (MLP). The dataset was partitioned into training (80%) and testing (20%) subsets, and validation (Linhai), ensuring stratification of sites to reflect all growth stages. XGBoost had superior prediction performance, achieving the lowest Root Mean Square Error (9.225 cm), Mean Absolute Error (8.044 cm), Mean Absolute Percentage Error (8.438%), and the highest coefficient of determination (R2 = 0.75). The robust correlation between Sentinel-2 indicators and crop height enhanced the model’s accuracy. The results indicate that employing sophisticated modelling techniques with multispectral satellite data is an effective method for reliably monitoring crops on a broad scale. Future research should focus on optimizing feature selection, incorporating environmental variables such as soil type and cultivar, and improving model generalizability across diverse agro-ecological locations.