Data and code for: Data leakage and spatial cross-validation in machine-learning crop yield prediction: a Sentinel-2 assessment of silage maize at district scale (TR22, Türkiye)

This dataset contains the analysis code and the derived analysis-ready dataset supporting the associated article. It enables full reproduction of the reported results: the variance decomposition, the cross-validation matrix (random, leave-one-district-out, spatial-block, leave-one-year-out), the signal-versus-structure attribution, the significance tests, and the figures. The study covers the TR22 South Marmara region of Türkiye (Balıkesir and Çanakkale provinces), 30 districts over nine seasons (2017-2025; 270 district-year observations). District silage-maize yield statistics are from the Turkish Statistical Institute (TÜİK); Sentinel-2 Surface Reflectance and ERA5-Land meteorology are openly available via Google Earth Engine. Contents: Python scripts for Earth Engine extraction, data harmonisation, modelling (hierarchical models, validation regimes, attribution, significance testing, feature importance), and figure generation; the analysis-ready panel (CSV); and the TÜİK yield series. See README.md for reproduction steps.