Cross-Scale Prediction of MODIS Land-Surface Phenology Reference Dates from Sentinel-1/2 Time Series Across Monsoon Rice Regions
Description
This data record provides the processed, de-identified input data, analysis code, configuration files, complete out-of-fold predictions, parameter-selection traces, summary tables and publication figures for the study ‘Cross-Scale Prediction of MODIS Land-Surface Phenology Reference Dates from Sentinel-1/2 Time Series Across Monsoon Rice Regions’. The study authors are Fan Liu, Yugeng Guo, Yi Sun, Haoze Zhang and Wenzhi Zeng. The package reproduces rule-based, Elastic Net, single-view RBF-SVR, feature-concatenation RBF-SVR and dual-view multiple-kernel SVR (MKL-SVR) comparisons for MCD12Q2-derived SOS, Peak and EOS reference dates. It contains 2,483 common-eligible field-year records from Battambang (Cambodia), Nam Ngum (Lao PDR) and Jinghong, Yunnan (China), covering 2019–2024. Each record contains Sentinel-1 and Sentinel-2 time-series features. MCD12Q2-derived dates are regional-scale proxy reference dates and do not constitute field-level agronomic ground truth. Precise coordinates, field boundaries, original field identifiers and raw archived Sentinel/MODIS files are excluded. Sample identifiers are pseudonymised; precomputed anonymised group and fold identifiers are included solely to reproduce the reported validation design. The public package cannot reconstruct original locations or spatial groups.
Files
Steps to reproduce
1. Download and extract the archive. 2. Create a Python 3.9.13 environment and install pinned dependencies: python -m pip install -r requirements.txt. 3. From the package root, run: python code/run_anonymized_v4_validation.py --package . --out results/recomputed --verify 4. Compare the regenerated out-of-fold predictions and outer-fold metrics with the supplied results. The portable runner uses de-identified features, fixed outer/inner split assignments and rule-candidate arrays to reproduce the seven reported model families. The README documents the expected full-run time, fixed random seed and limitations: the release cannot reproduce original image extraction, original field locations or coordinate-derived grouping; cluster-bootstrap confidence intervals can differ when anonymous group ordering is used.