Amazon next-day wildfire spread dataset at 375 m (2020–2024)

Amazon next-day wildfire spread dataset at 375 m (2020–2024) This dataset supports next-day wildfire spread prediction in the Brazilian Amazon biome at the native 375 m resolution of the VIIRS I-band active fire product. Each sample is a 64 × 64 pixel tile (24 km × 24 km) with the fire mask at day t, 23 environmental predictor channels for day t, and the fire mask at day t+1 as the target. It accompanies the paper “Next-day wildfire spread prediction in the Brazilian Amazon at 375 m resolution” (XXII Brazilian Symposium on Remote Sensing, SBSR 2027) [DOI of the paper, when available]. Contents 48,288 tiles from 2020 to 2024 (2020: 9,481; 2021: 5,756; 2022: 10,076; 2023: 8,395; 2024: 14,580), stored in compressed NumPy shards ({year}_{k}.npz) with one Parquet metadata file per shard. In the paper, tiles with at least 50% valid pixels at both t and t+1 are used (48,146 tiles), split by year: 2020–2022 training, 2023 validation and 2024 testing. Normalization statistics computed on the training split (stats_ano.json), the tile index (indice_tiles.parquet) and the list of days removed because of S-NPP data gaps (datas_invalidas.csv). [Code: the Jupyter notebooks used to build the dataset and run the benchmark / link to the code repository.] Arrays in each shard X (float16, N × 17 × 64 × 64): continuous predictors, in this order: topography: elevation (m), slope (degrees), aspect_sin, aspect_cos; vegetation: NDVI, EVI, NBR; surface temperature: LST_day (°C); weather: u10, v10, wind_speed (m s⁻¹), temperature_max, temperature_min (°C), precipitation (mm day⁻¹), sph (kg kg⁻¹); drought: precip_anom_30d (standardized anomaly); population: world_pop (persons per cell). Missing values are NaN. lc (uint8): MapBiomas land-cover group for year t−1 (0 forest, 1 savanna, 2 grassland, 3 farming, 4 other anthropic, 5 water, 255 no class). prev_fire, fire (int8): fire masks at t and t+1 (1 fire, 0 no fire, −1 invalid). vnp_t, vnp_t1 (uint8): raw VNP14A1 FireMask classes at t and t+1. id_rotulo: sample identifier. Grid. EPSG:4326 with a pixel size of 375/111,320 degrees, anchored at (−180°, 90°). Each tile’s upper-left cell is given by col0 and row0 in the metadata, so tiles can be georeferenced and new layers can be extracted on the same grid. Construction Fire detections. VIIRS S-NPP 375 m active fires (FIRMS, standard processing), vegetation fires only (type 0), nominal or high confidence, grouped by UTC date. They were rasterized on the grid with the scan × track footprint of each detection. Sampling. Tiles were sampled around daily DBSCAN clusters of detections inside the Amazon biome (haversine distance, ε = 3 km, minimum 4 points), with up to three tiles per cluster. Inclusion criterion. A tile was kept when at least five new fire pixels appeared at t+1 within 3 km of the fire at t. Invalid pixels. Pixels outside Brazil, and pixels flagged as cloud, unprocessed or unknown in VNP14A1, are invalid (−1) unless a fire was detected. Predictors. They were extracted in Google Earth Engine on the same grid: MODIS composites ending before t and the most recent LST up to t; ERA5-Land daily aggregates of day t; the 30-day precipitation anomaly ending at t−1, relative to the previous ten years; WorldPop 2020; MapBiomas Collection 10 for year t−1. Finer layers were aggregated to the grid, and coarser layers were bilinearly resampled. Source data and attribution. This dataset is derived from the following sources, which should also be acknowledged when it is used: NASA FIRMS / VIIRS S-NPP 375 m active fire product (Schroeder et al., 2014). VNP14A1 V002 (doi:10.5067/VIIRS/VNP14A1.002). MODIS MOD13A1, MOD09A1 and MOD11A1, Collection 6.1 (NASA LP DAAC). SRTM (Farr et al., 2007). ERA5-Land (Muñoz-Sabater et al., 2021), generated using Copernicus Climate Change Service information. Neither the European Commission nor ECMWF is responsible for any use of this information. WorldPop 2020 (doi:10.5258/SOTON/WP00647). MapBiomas Collection 10 (Souza et al., 2020). IBGE biome boundaries (2019). We acknowledge the use of data from NASA’s Fire Information for Resource Management System (FIRMS), part of NASA’s Earth Science Data and Information System (ESDIS). License. CC BY 4.0. Please cite this dataset and the accompanying paper.