A comparative analysis of machine learning and vegetation index-based modeling approaches for durum wheat yield assessment using Sentinel-2 imagery

Abstract

Durum wheat is a globally important cereal. Therefore, precise yield assessment is crucial for informed decision-making in precision agriculture. Remote sensing techniques, specifically high-resolution multispectral data from the Sentinel-2 mission, provide valuable insights. Additionally, machine learning algorithms offer a powerful alternative to traditional modeling by efficiently processing multi-dimensional datasets and extracting complex relationships from remote sensing data. This study presents two modeling approaches for the estimation of durum wheat yield based on Sentinel-2 data, covering 157 fields in Thessaly plain (Greece) across five growing periods. The first approach comprises multiple linear regression modeling based on vegetation indices (VI-MLR) that capture plant, water and soil signals (NDVI, EVI, NMDI, NDWI). The results demonstrate moderate accuracy (R 2 : 0.476-0.503, RMSE: 873-897 kg ha -1 ). In the second approach, the reflectance data of all Sentinel-2 bands for several dates during the growth periods are used as input parameters in three machine learning model algorithms, i.e., Random Forest (RF), K-Nearest Neighbors (KNN) and Boosting Regressions (BR), resulting in significantly improved accuracy. Utilizing all 12 bands from Sentinel-2 demonstrates high accuracy (RF and KNN: R 2 > 0.94, RMSE < 318 kg ha -1 ) throughout the growing season, that remains high when images from the start of the growing period until February, i.e., four months before harvest, are used (RF and KNN: R 2 > 0.92, RMSE < 357 kg ha −1 ). Performance remains high when using only the 6 most important bands (RF and KNN: R 2 > 0.92, RMSE < 349 kg ha −1 with all images, R 2 > 0.92, RMSE < 364 kg ha −1 with images until February). This study underscores the potential of machine learning models for precise durum wheat yield assessment using satellite imagery, even in early stages. However, further research is necessary to evaluate their generalizability across regions with varying climatic and growing conditions.