Crop Harvest Prediction Using Remote Sensing and Machine Learning: An Integrated Framework

Accurate estimation of the optimal harvest date (OHD) is critical for maximising crop yields and grain quality. This paper presents a machine learning pipeline that uses real soil and climate agronomic data alongside simulated Sentinel-2 Normalised Difference Vegetation Index (NDVI) time-series to classify crop types and compute an Extended Harvest Readiness Index (HRI) for four major crops: Maize, Wheat, Rice, and Cotton. NDVI time-series are modelled using the double-logistic phenology model (Zhang et al., 2003), with parameters derived from published Sentinel-2 field studies (Zhong et al., 2014; Xu et al., 2019). Five NDVI-derived features are extracted per field sample: peak NDVI, day of peak NDVI, NDVI at harvest date, season-integrated NDVI, and the rate of NDVI decline during senescence. Four classifiers are evaluated — Random Forest (RF), Gradient Boosting (GB), Support Vector Machine (SVM with RBF kernel), and Multilayer Perceptron (MLP) — on both tabular-only and NDVI-augmented feature sets. The best-performing model, Random Forest, achieves 100% test accuracy on tabular features and maintains this accuracy with NDVI augmentation. The Extended HRI combines model confidence, temperature suitability, and NDVI-based senescence indicators into a single actionable readiness score. Key findings confirm that NDVI characteristics — particularly ndvi_peak and ndvi_peak_doy — rank among the most important features, supporting the value of remote sensing augmentation for precision agriculture. Keywords: crop classification, remote sensing, ndvi phenology, harvest readiness, machine learning, random forest, sentinel-2