Satellite-derived vegetation indices and machine learning for food crop yield prediction: A systematic review of methods, performance and research gaps

Crop yield prediction is essential for global food security, but the current methods for yield estimation are labour-intensive, time-consuming and have limited spatial and temporal coverage. This systematic review focuses on consolidating the evidence from the 32 peer-reviewed studies published from 2018–26 that are concerned with using both satellite-derived vegetation indices (VIs) and machine learning (ML) to predict food crop yields. A preferred reporting items for systematic reviews and meta-analyses (PRISMA)-based search resulted in the identification of studies on rice, wheat, maize, soybean, sugarcane, sorghum and barley across different geographic regions. The evidence clearly shows that across all food crops, Random Forest (RF) is the most consistent ML technique, with an R² of 0.70–0.96 based on crop and region and deep learning approaches like long short-term memory (LSTM) and convolutional neural network (CNN) show higher temporal pattern recognition, but much greater numbers of labelled data are needed. Multi-index approaches consistently outperform single-index approaches, with those extracted from Sentinel-2 performing well in situations where normalised difference vegetation index (NDVI) loses sensitivity due to canopy saturation at high biomass levels in dense-crop canopies. Sentinel-2 is the most frequently used satellite platform (over 65 % of the analysed studies) and Google Earth Engine (GEE) has become the most used data processing environment. There are still gaps in the standardisation of benchmarking metrics, the representation of smallholder and tropical farming systems, the integration of data across phenological stages and the cross-regional transferability of models. This review consolidates evidence and presents clear priorities to move precision agriculture yield forecasting toward global food security applications.