Assessing the temporal transferability of machine learning models for predicting processing pea yield and quality using Sentinel-2 and ERA5-land data

Accurate pre-harvest prediction of yield and quality (tenderometric reading, TR) is crucial for the processing pea industry due to a narrow optimal harvest window. Machine learning (ML) models offer potential, but their real-world utility depends on their performance stability across different years. This ability, known as temporal transferability, is often overestimated by standard validation methods, creating a gap between theoretical performance and operational reliability. This study aims to rigorously quantify this temporal transferability gap for both pea yield and TR prediction. Four ML algorithms (RF, XGBoost, GPR, SVMr) were evaluated using Sentinel-2 and ERA5-Land data from 2018 to 2024 in northern Italy. A comparison was made between a standard group-based cross-validation (LOGOCV) and a temporally rigorous Leave-One-Year-Out Cross-Validation (LOYOCV). For yield prediction, ML models outperformed a baseline (NullModel) under LOGOCV (RF nRMSE = 14.6%), but performance degraded significantly under LOYOCV, revealing a clear transferability gap. TR prediction was more challenging while RF showed promising results in LOGOCV (nRMSE = 18.7%), all ML models were outperformed by the NullModel in the more realistic LOYOCV scenario. The findings highlight a critical temporal transferability gap, especially for the TR parameter, limiting the current operational readiness of standard ML models. It is recommended that future work focus on more robust approaches, such as models designed for temporal data (e.g., RNNs, Transformers) and higher-resolution data, to bridge the gap towards reliable real-world application.