Estimating Total Rice Biomass and Crop Yield at Field Scale Using PlanetScope Imagery Through Hybrid Machine Learning Models

Abstract Near real-time crop monitoring has been a challenging due to the lack of high-resolution remote sensing images suitable for agricultural applications. The PlanetScope constellation, comprising approximately 130 Dove satellites, collects images of the entire Earth daily, with a resolution of 3.7 m. The high-resolution images from the PlanetScope satellite, along with vegetation indices, geo-environmental data, and soil and crop parameters, were utilized and analysed using machine learning models to enhance the accuracy of predicting total biomass and rice crop yield at the field scale. The study area, covering nearly 214 sample rice plots, was located in the Tarekswar block of Hooghly, West Bengal, India. Alongside ten vegetation indices and three Principal Component Analysis (PCA) soil nutrient levels, approximately thirty-six factors were analyzed to predict rice total biomass and crop yield using ten machine learning (ML) models, namely Random forest (RF), Extreme Gradient Boosting (XGB), Support Vector Machine (SVM), Bagging Tree (Treebag), Generalized Additive Models (gamSpline), Elastic Net (enet), Ordinary regression with LASSO penalty (rqlasso), Tree Models from Genetic Algorithm (evtree), Bayesian Regularized Neutral Networks ( brnn), cubist models, and there hybrid of ensembles. Boruta and multi-collinearity analysis were also conducted for the selected factors to explore their influence levels. The study area exhibited robust rice yields ranging from 5 to 10 t/ha, accompanied by healthy biomass growth. Four ML models ─cubist, random forest, enet, and the hybrid model—showed promising predictions with R 2 > 0.88. Most models classified less than 20 ha of the study area as falling into the “very-low suitable class”, showing the region’s suitability for rice cultivation due to its highly fertile alluvial soil. Boruta sensitive analysis revealed that nearly 24 individual factors significantly influenced the final crop yield including, organic carbon (OC), phosphorus (P), electrical conductivity (EC), mechanization level, and the majority of the vegetation indices. A critical analysis carried out through the Map query tool showed that five vegetation indices estimated via PlanetScope displayed strong correlations (exceeding 89%) in identifying areas with high to very high rice yields. The study can serve as a guideline for near-real-time crop monitoring in the near future, using high-resolution PlanetScope images.