Reconstructing MODIS/VIIRS Fire Events in a Tropical Peatland Landscape: Sentinel-1 Audit and Robustness Assessment

This repository contains the data and supporting materials from the study “Reconstructing MODIS/VIIRS Fire Events in a Tropical Peatland Landscape: Sentinel-1 Audit and Robustness Assessment.” The research focused on reconstructing candidate fire events in the Tanjung Lago-Banyuasin II tropical peatland landscape in South Sumatra, Indonesia, using MODIS and VIIRS active-fire records. The original archive contained 437 active-fire detections. After confidence filtering, 415 detections were retained and grouped using a 1 km spatial threshold and a 24-hour temporal threshold. This process produced 218 candidate fire events, including 34 multi-product clusters and 184 single-product clusters, of which 139 were singletons. The purpose of this reconstruction was to reduce repeated counting of the same fire while keeping the original detection history and event provenance traceable. A Sentinel-1 VV-VH based audit was then developed using 18 development clusters. Once the indicators, thresholds, and decision rules had been established, they were frozen and applied without further adjustment to an independent stratified random sample of 60 evaluation clusters. This Evaluation60 sample consisted of 18 multi-product clusters, 18 single-product non-singleton clusters, and 24 singleton clusters. The stability of the results was further examined using seven robustness scenarios and leave-one-acquisition-out analyses. The final Sentinel-1 assessment identified 15 provisionally corroborated cases, 24 partially corroborated cases, and 21 unresolved cases. These were further mapped into 5 class-A, 7 class-B, 27 class-C, and 21 class-D cases according to their level of evidence and recommended use as provisional labels. Overall, 56 of the 60 main Sentinel-1 decisions remained unchanged after robustness testing. After accounting for the stratified sampling design, the estimated proportion of candidate events receiving at least partial Sentinel-1 support was 57.58% (95% CI: 44.41–70.76%). To provide an additional and independent source of evidence, the same 60 evaluation clusters were also examined using Landsat 8/9 imagery. The Landsat procedure was defined and frozen separately before its results were compared with the Sentinel-1 outcomes. Twenty-six clusters had sufficient optical coverage for spectral evaluation. Seven clusters met all conservative REF-FIRE criteria, while no cluster met the REF-NOFIRE criteria. The remaining cases were unresolved because of limited image availability or inconclusive spectral responses. Importantly, all seven REF-FIRE cases were already within the Sentinel-1-supported group: two were provisionally corroborated and five were partially corroborated. The repository includes the candidate-event inventory, retained hotspot records, evaluation-sample information, frozen Sentinel-1 outputs, robustness and leave-one-acquisition-out results, Landsat image-sufficiency and spectral-audit outputs, final optical-reference classifications, comparison tables, weighted estimates, quality-control documentation, Google Earth Engine scripts, and supporting verification materials. These materials are intended to support transparent and reproducible candidate-event reconstruction and provisional label assessment. They should not be interpreted as field-validated fire ground truth or as a burned-area product. The unresolved category does not mean that no fire occurred, and unresolved optical cases should not be treated as confirmed no-fire observations. Because no confirmed REF-NOFIRE cases were obtained and independent field reference data were not available, the results should not be used to estimate sensitivity, specificity, false-positive rate, overall classification accuracy, or burned-area accuracy.