Dataset of Deep Learning from Landsat-9 Satellite Images for Segmentation Burned Areas in South Sumatera

This dataset comprises a curated collection of Landsat-9 Level-1 Top of Atmosphere (TOA) satellite imagery, rigorously processed to facilitate the training and benchmarking of deep learning algorithms, specifically semantic segmentation architectures such as U-Net, for burned area delineation. The geographic scope of this study encompasses South Sumatra Province, Indonesia, specifically within the WRS-2 path/rows 124/062 and 125/062. This region represents a critical area of interest due to its high susceptibility to recurrent forest and land fires. The dataset provides a robust baseline for remote sensing research aimed at automating the detection of fire-induced land cover changes using multispectral data.

Each data instance within this repository constitutes a co-registered pair consisting of a multispectral input image and a corresponding binary ground truth mask, standardized to a spatial resolution of 30 meters. The input data is encoded in GeoTIFF raster format and features a five-channel spectral stack designed to maximize the separability between burned and unburned features. These channels include the Normalized Difference Vegetation Index (NDVI), the Normalized Burn Ratio (NBR), Band 4 (Red), Band 5 (Near-Infrared/NIR), and Band 7 (Shortwave Infrared/SWIR-2). Prior to patch generation, the imagery underwent a stringent pre-processing pipeline, which included atmospheric correction via the Dark Object Subtraction (DOS) method and the removal of cloud and shadow artifacts utilizing the Quality Assessment (QA) band to ensure radiometric integrity.

The ground truth annotations were generated via an automated thresholding mechanism applied to the differenced Normalized Burn Ratio (dNBR) index. The dNBR metric was computed as the temporal difference between pre-fire acquisitions (January 1 – July 30, 2023) and post-fire acquisitions (November 1, 2023 – June 30, 2024). A threshold of dNBR > 0.1 was applied to classify pixels as "burned" (class 1) versus "unburned" (class 0), thereby isolating significant spectral alterations indicative of vegetation loss.

The final dataset consists of 1,008 image patches, each with dimensions of 512 x 512 pixels, extracted using a sliding window technique with a 50% spatial overlap. To ensure data quality, an automated filtering protocol was employed to exclude patches containing excessive void pixels resulting from cloud masking or scene boundaries. It is pertinent to note that the dataset provided herein is raw and unsplit to allow for flexible experimental design. However, the foundational research associated with this dataset employed a stratified split ratio of 70:15:15 (yielding 706 training samples, 151 validation samples, and 151 testing samples). Researchers utilizing this dataset are encouraged to replicate this partitioning strategy for comparative analysis or to implement alternative cross-validation schemes as required.