Time Series Extrinsic Regression for Smart Agriculture: Field‐Level Crop Yield Prediction From Multisource Agricultural Time Series

ABSTRACT The growing availability of temporal data in agriculture has created new opportunities for data‐driven decision support systems aimed at improving the efficiency, sustainability, and adaptability of agricultural systems. Remote sensing platforms, weather‐related datasets, and field‐management records provide information on crop development and environmental conditions throughout the growing season. In this context, predictive tasks such as crop yield prediction, irrigation requirement assessment, and nutrient demand estimation can be naturally formulated as Time Series Extrinsic Regression (TSER), a supervised learning task in which a continuous target is predicted from one or more time series. This paper presents TSER within the Smart Agriculture domain and investigates its applicability to field‐level crop yield prediction using multisource agricultural time series. The study adopts a field‐based temporal approach where Sentinel‐2 observations are processed to derive vegetation‐index time series, spatially aggregated at parcel level, and integrated with ERA5‐Land weather‐related variables and static vineyard attributes. Representative neural and non‐neural TSER methods are compared under temporal and spatial generalization scenarios. Results show that TSER provides a suitable supervised learning framework for crop yield prediction from heterogeneous agricultural time series. Conditional LSTM achieves the best overall performance under temporal generalization across vintages, while non‐neural methods remain competitive under spatial generalization to unseen vineyard blocks. These findings suggest that TSER can support field‐level predictive modeling in Smart Agriculture, with model choice depending on the target generalization setting, data availability, and operational constraints.