Changelog

Release notes for the framework and the documentation site.

The framework uses DateVer (vYYYY.MM) for releases. Patches within a release use vYYYY.MM.N.

v2026.07 - correctness and provenance

Corrections found while reconciling this documentation site against the code and against the thesis value ledger. The correction algorithm itself did not change, so results produced with v2026.05 remain valid.

Fixed. The batch cells in notebooks 02 and 03 were left at development ranges, covering 9 and 3 of the 36 dekadal windows while reporting that they covered all 36; both now run the full year. existing_model_action was read from config.yml but never consulted, so overwrite silently reused a cached model. apply_deeplearning_model froze blend_alpha at import time, so a value set by a later initialize_config() had no effect; it and the GPD threshold percentile are now resolved when the function runs. Docstrings describing a saturation_count default of 3 were corrected to 2.

Added. Corrected precipitation files now carry Conventions: CF-1.8, a references DOI, and run-context attributes recording the release, commit and the three sensitivity parameters. The precipitation variable no longer carries a standard_name: the previous value is not in the CF standard-name table, and CF treats the attribute as optional, so omitting it is correct when no controlled term applies. units now read mm/day, following the IMERG-L source. Data values are unchanged, and files in the archived Indonesia bundle predate this.

Changed. Zenodo citations now use all-versions DOIs so they follow each release. Documentation was corrected throughout: function names that no longer existed, output filename patterns, the CNN architecture listing, the metric inventory, runtime figures, the validated-station count, and correlation values that had drifted from the value ledger.

Companion publication. The journal manuscript on the methodology and Indonesia validation is now published in Remote Sensing 2026, 18, 2298 (doi:10.3390/rs18142298), superseding the under-review status recorded at v2026.05.

v2026.05 - first public release

The first public release of the Hybrid Bias Correction framework. This is the codebase developed for the MSc thesis on daily satellite precipitation bias correction over Indonesia, refactored into a reproducible Python package for the broader community.

What’s in this release

The framework (src/) - Linear Scaling, Empirical Quantile Mapping with a Generalized Pareto tail, and a CNN refinement gated by station-density confidence. Implemented as a Python package with one driver config per region.
Notebooks (notebooks/) - end-to-end pipeline from AOI definition and data acquisition through bias correction, metrics, QA, station validation, and visualisation.
Bali example bundle (data/example_bali/, 11 MB) - ships with the repo and runs in about 72 minutes across notebooks 02 to 06 on a free Colab CPU.
Indonesia data bundle - 2001-2025 inputs, masks, and outputs (~40 GB) published at Zenodo 10.5281/zenodo.20287846.
Documentation (this site) - Methodology, Implementation, Tutorials, Case Studies (Bali + Indonesia, with interactive figures), FAQ, and auto-generated API Reference.

Backlog

Replace the CNN architecture with a U-Net to avoid the Flatten -> Dense bottleneck.
tf.keras.backend.clear_session() between dekads in nb02 for memory predictability on Indonesia-scale runs.

--- title: "Changelog" description: "Release notes for the framework and the documentation site." --- The framework uses [DateVer](https://github.com/datever/datever) (`vYYYY.MM`) for releases. Patches within a release use `vYYYY.MM.N`. ## v2026.07 - correctness and provenance Corrections found while reconciling this documentation site against the code and against the thesis value ledger. The correction algorithm itself did not change, so results produced with v2026.05 remain valid. **Fixed.** The batch cells in notebooks 02 and 03 were left at development ranges, covering 9 and 3 of the 36 dekadal windows while reporting that they covered all 36; both now run the full year. `existing_model_action` was read from `config.yml` but never consulted, so `overwrite` silently reused a cached model. `apply_deeplearning_model` froze `blend_alpha` at import time, so a value set by a later `initialize_config()` had no effect; it and the GPD threshold percentile are now resolved when the function runs. Docstrings describing a `saturation_count` default of 3 were corrected to 2. **Added.** Corrected precipitation files now carry `Conventions: CF-1.8`, a `references` DOI, and run-context attributes recording the release, commit and the three sensitivity parameters. The precipitation variable no longer carries a `standard_name`: the previous value is not in the CF standard-name table, and CF treats the attribute as optional, so omitting it is correct when no controlled term applies. `units` now read `mm/day`, following the IMERG-L source. Data values are unchanged, and files in the archived Indonesia bundle predate this. **Changed.** Zenodo citations now use all-versions DOIs so they follow each release. Documentation was corrected throughout: function names that no longer existed, output filename patterns, the CNN architecture listing, the metric inventory, runtime figures, the validated-station count, and correlation values that had drifted from the value ledger. **Companion publication.** The journal manuscript on the methodology and Indonesia validation is now published in *Remote Sensing* 2026, 18, 2298 ([doi:10.3390/rs18142298](https://doi.org/10.3390/rs18142298)), superseding the under-review status recorded at v2026.05. ## v2026.05 - first public release The first public release of the Hybrid Bias Correction framework. This is the codebase developed for the MSc thesis on daily satellite precipitation bias correction over Indonesia, refactored into a reproducible Python package for the broader community. ### What's in this release - **The framework** (`src/`) - Linear Scaling, Empirical Quantile Mapping with a Generalized Pareto tail, and a CNN refinement gated by station-density confidence. Implemented as a Python package with one driver config per region. - **Notebooks** (`notebooks/`) - end-to-end pipeline from AOI definition and data acquisition through bias correction, metrics, QA, station validation, and visualisation. - **Bali example bundle** (`data/example_bali/`, 11 MB) - ships with the repo and runs in about 72 minutes across notebooks 02 to 06 on a free Colab CPU. - **Indonesia data bundle** - 2001-2025 inputs, masks, and outputs (~40 GB) published at [Zenodo 10.5281/zenodo.20287846](https://doi.org/10.5281/zenodo.20287846). - **Documentation** (this site) - Methodology, Implementation, Tutorials, Case Studies (Bali + Indonesia, with interactive figures), FAQ, and auto-generated API Reference. ## Backlog - Replace the CNN architecture with a U-Net to avoid the Flatten -> Dense bottleneck. - `tf.keras.backend.clear_session()` between dekads in nb02 for memory predictability on Indonesia-scale runs.