Next-day forest fire detection forecasting for Indian forest cells: data set, stored predictions and code

Next-day forest fire detection forecasting for forest cells in India. One row of the data is one forest cell of 0.1 degrees on one UTC day. The label is whether the cell has a VIIRS satellite fire detection (NOAA-20 and Suomi NPP) inside forest boundaries on the next day. It is a satellite detection, not a confirmed forest fire. The data cover 7,148 forest cells in 17 states, used for training, validation and testing, and 526 cells in 3 blind states (Uttar Pradesh, West Bengal and Kerala) that were never used for training, scaling or tuning. They span 2,557 days, from 1 January 2019 to 31 December 2025. Each cell-day has 41 inputs: 33 observed inputs (weather, vegetation, land cover, terrain, people, forest context, fire history, calendar) and 8 inputs from the Global Forecast System weather forecast of the same day. The record also holds the stored predictions of every model run (logistic regression, random forest, XGBoost, MLP, LSTM, CNN, ConvLSTM, a ConvLSTM with attention and a graph network, five seeds each, plus three references that need no training: persistence, climatology and the Fire Weather Index), the saved models of the first seed, the result tables, the audit of the data and the code. The settings of all runs were fixed in a checksum before the blind states were scored. Start with README.md for the layout and a short loading example. DATA_DICTIONARY.csv describes every file and column, and SOURCES_AND_LICENCES.md lists the source data and the notices to keep. Licence: CC BY 4.0 for the data, predictions and tables. The road-distance column is derived from OpenStreetMap and is available under ODbL 1.0. The code is MIT. Source data keep the terms of their providers.