Hybrid Bias Correction of IMERG Late Run V07 over Indonesia (2001-2025): Input Data, Land-Sea Masks, and Corrected Products

This deposit contains the full input, mask, and output data bundle for the hybrid statistical-deep learning bias correction of daily IMERG Late Run V07 precipitation over Indonesia (95-141 deg E, -11 to 6 deg N) for the period 2001-2025. The bundle accompanies the companion paper, Istanto, B., Boer, R., Santikayasa, I. P. (2026), A Modular and Transferable Framework for Enhancing Satellite-Derived Daily Precipitation: Adjusting Values, Aligning Distributions, and Preserving Extremes, Remote Sensing 18(14), 2298, https://doi.org/10.3390/rs18142298, and the open-source code at https://github.com/bennyistanto/hybrid-bias-correction. The framework combines three correction stages applied per 10-day window (dekad). Linear Scaling matches the long-term mean. Empirical Quantile Mapping with a Generalized Pareto Distribution tail (80th-percentile threshold) reshapes the full daily distribution. A lightweight two-layer convolutional network refines the spatial residual on extreme days. A station-density confidence mask, built from the BMKG gauge network, modulates the CNN influence so that the correction degrades gracefully to LSEQM in gauge-sparse areas. input.zip contains 1. IMERG Late Run V07 (idn_imergl.nc4, 2001-01-01 to 2025-12-31)2. IMERG Final Run V07 baseline (idn_imergf.nc4, 2001-01-01 to 2025-09-30)3. CPC-UNI regridded to 0.1 deg (idn_cpcuni.nc4) and at its native 0.5 deg resolution (idn_cpcuni_native05.nc4)4. BMKG station location and daily observation CSVs. mask.zip contains the Indonesia land-sea mask at 0.1 deg together with the area-of-interest subsets used for clipping (idn_subset.nc). output.zip contains 1. Per-method corrected products (corrected_ls, corrected_lseqm, corrected_lseqmdl)2. Per-method performance metrics (metrics_ls, metrics_lseqm, metrics_lseqmdl)3. Continuous Quality Index outputs (quality_ls, quality_lseqm, quality_lseqmdl)4. Station-density confidence mask5. Independent BMKG station validation CSVs6. Figures trained_models.zip contains the trained CNN models, one per dekad. window_diagnostics.zip contains the derived arrays behind the temporal-alignment analysis, together with the per-station half-hourly cache they are built from (2001 to 2021, one file per year, matching the span of the BMKG station record). These are computed from the raw half-hourly IMERG-L archive rather than from the daily inputs in this deposit, so the other bundles cannot reproduce them. The gridded arrays are split at the satellite transition: 2001 to 2013 for the TRMM-input era and 2015 onward for the GPM era, with 2014 excluded as a transition year. Two GPM windows are included, one ending in 2021 so that it is comparable with the station-based analysis, and one running to the end of the satellite record. The full-record array was assembled by pooling the shorter window with a later increment that was not retained; because these are additive sufficient statistics, that increment is recovered by subtraction from the two files included here, and the README gives the four lines needed. The README also maps each array to the scripts that write and read it. Those scripts are in the thesis repository at https://github.com/bennyistanto/msc-thesis under scripts/; this bundle contains data only. All gridded files are float32 NetCDF with zlib compression on a common 0.1 degree lat-lon grid, with precipitation in mm/day following the IMERG-L source. The metrics, quality-assessment, and station-density files declare Conventions: CF-1.8; the corrected-precipitation files in this bundle do not, because that attribute was added to the corrected-product writer in framework release v2026.07, after these data were produced. The CPC-UNI 0.5 degree file is kept on the original gauge-analysis grid and is used by the parameter-fitting path that avoids the 5x5 nearest-neighbour block boundary. The run used blend_alpha = 0.70, gpd_threshold_percentile = 80, and station-density saturation_count = 2. These three parameters were not formally optimised and were chosen by inspection. Users adapting the framework to another region should treat them as sensitivity parameters rather than fixed defaults. Outputs written by v2026.07 and later additionally record these parameters, the framework version, and the source commit as global attributes; files in this deposit predate that change. To reproduce the outputs from the inputs, clone the code repository, point config.yml at the directories in this bundle, and run notebooks 02 through 06 in sequence. A smaller Bali example (about 11 MB) ships with the code repository so users can validate the toolchain on a free Google Colab session before downloading the full Indonesia bundle. Full documentation is available at https://bennyistanto.github.io/hybrid-bias-correction. If you use these data, please cite both this Zenodo deposit and the companion paper (https://doi.org/10.3390/rs18142298). NASA GPM IMERG V07 data are courtesy of the NASA Goddard Earth Sciences Data and Information Services Center. CPC Unified gauge-based daily precipitation data are provided by the NOAA Physical Sciences Laboratory. BMKG station observations are provided by the Indonesian Agency for Meteorology, Climatology and Geophysics.