*Result*: Integrating UAV hyperspectral imaging with machine learning techniques to predict tomato ecophysiological parameters and yield.

*Purpose: Unmanned aerial vehicle (UAV)-based hyperspectral (HS) imaging enables precise monitoring of crop growth parameters. Machine learning (ML) has recently gained significant attention in precision agriculture as a powerful statistical learning technique for processing complex, multi-dimensional remote sensing data. This study aims to evaluate the integration of UAV-based HS imaging and ground measurements, applying various ML techniques to predict ecophysiological, agronomic and quality parameters of processing tomatoes in the Mediterranean region. Methods: Ground-truth data were collected at different growth stages over three crop years, including leaf chlorophyl (CHL) content, leaf water potential (LWP), fruit yield (YLD), and total soluble solids (TSS). In parallel, a UAV-mounted HS camera acquired images in the VIS–NIR region (400–1000 nm), from which spectral reflectance was retrieved and 19 vegetation indices (VIs) were calculated following image calibration and processing. Recursive feature elimination (RFE) was performed using support vector machine (svm) and random forest (rf) models to select the most relevant features before proceeding with ML analysis. Several ML algorithms, including linear modelling (LM), RF, SVM, k-nearest neighbors (KNN), and partial least squares (PLS) regression, were implemented to predict the crop parameters based on the HS bands and VIs. Results: Results showed that both RFE approaches effectively selected relevant features, with svm-based RFE performing better for HS bands, while rf-based RFE was more suited for VIs. The best selected models were mainly preceded by rf-based RFE and built on limited number of VIs, with model performance varied for each parameter: rf-LM yielded the best predictions for CHL and LWP using VIs, achieving R2 values of 0.52 (RMSE = 3.03 Dualex unit) and 0.56 (RMSE = 0.16 MPa), while rf-PLS outperformed yield prediction using HS bands (R2 = 0.73; RMSE = 0.47 kg/plant), and rf-RF performed better for TSS relying on VIs (R2 = 0.43; RMSE = 0.42°Brix). Various combinations of optimal HS bands and VIs were selected for each single parameter as best performing predictive features. Moreover, the study identified key insights to optimize the timing and sampling strategies to improve prediction efficiency and sustainability. Conclusion: The results demonstrate that UAV-based hyperspectral imaging, combined with machine learning and RFE-selected features, is effective for assessing and predicting ecophysiological traits, yield, and quality in processing tomato. This integrated approach offers a valuable tool for advancing precision agriculture through targeted crop monitoring and management. [ABSTRACT FROM AUTHOR]*