Spectral bands vs. vegetation indices: An AutoML approach for processing tomato yield predictions based on Sentinel-2 imagery

Abstract

Predicting tomato yield is crucial for agricultural production planning, management, market supply, and risk management. While yield prediction in precision agriculture presents a complex challenge, advancements in remote sensor technologies have greatly improved the accuracy and feasibility of these predictions. In this paper, three modeling approaches were implemented to estimate processing tomato yield based on satellite data, covering 152 fields across two growing periods. Sentinel-2 spectral bands and Vegetation Indices (VIs) captured at 5-day intervals during the growth periods, used as input parameters in an AutoML pipeline. The first approach aimed to identify the optimal timeframe and the most effective spectral bands and VIs, such as NDVI and RVI. Modelling results indicated that Red/Red Edge/NIR bands performed well in predicting yield, with the period between 75 to 90 days post-transplanting identified as the optimal timeframe for yieldpredictions (R² of 0.56). The second approach incorporated inter-date VIs, utilizing bands from different dates, leading to a significant improvement in performance with an R² of 0.61 and root mean square error (RMSE) of 12ton/ha. The third approach involved band combinations to enhance performance, where specific bands, including the Bands 4, 6, and 12, collectively achieved the highest R² of 0.65. Using feature extraction algorithms such as PCA, UMAP, and autoencoder partially contributed to improved performance while using the same VIs on consecutive different dates. By utilizing a higher number of bands and dates without the constraint of a VI formula demonstrated the potential for enhanced model accuracy.