Abstract:Wildfire monitoring and prediction are essential for understanding wildfire behaviour. With extensive Earth observation data, these tasks can be integrated and enhanced through multi-task deep learning models. We present a comprehensive multi-temporal remote sensing dataset for active fire detection, daily wildfire monitoring, and next-day wildfire prediction. Covering wildfire events in the contiguous U.S. from January 2017 to October 2021, the dataset includes 3552 surface reflectance images and auxiliary data such as weather, topography, land cover, and fuel information, totalling 71 GB. The lifecycle of each wildfire is documented, with labels for active fires (AF) and burned areas (BA), supported by manual quality assurance of AF and BA test labels. The dataset supports three tasks: a) active fire detection, b) daily burned area mapping, and c) wildfire progression prediction. Detection tasks use pixel-wise classification of multi-spectral, multi-temporal images, while prediction tasks integrate satellite and auxiliary data to model fire dynamics. This dataset and its benchmarks provide a foundation for advancing wildfire research using deep learning.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the accuracy and efficiency of wildfire monitoring and prediction by using satellite image time - series data through a multi - task deep - learning model. Specifically, the paper aims to: 1. **Improve the accuracy of wildfire monitoring and prediction**: Existing satellite products have reliability issues in active fire - point detection and burned - area mapping, such as false alarms, low precision, and low temporal resolution. The paper hopes to improve the accuracy of these tasks by introducing a deep - learning model and making full use of the temporal, spatial, and spectral information of satellite data. 2. **Integrate multiple related tasks**: Most current datasets and research focus only on a single task (such as active fire - point detection or burned - area mapping) and ignore the relationships between different tasks. The paper proposes a multi - task dataset, TS - SatFire, which covers three key tasks: - **Active fire - point detection**: Identify active fire - points in satellite images. - **Daily burned - area mapping**: Map the burned - area on a daily basis. - **Wildfire progress prediction**: Predict the future progress of wildfires based on the current fire - point location and environmental information. 3. **Provide benchmark datasets and models**: To evaluate the performance of different models on these three tasks, the paper provides detailed benchmark test results. By comparing different model architectures (such as GRU, LSTM, T4Fire, U - Net, Attention U - Net, UNETR, and SwinUNETR, etc.), researchers can better understand which model is most suitable for a specific task and provide references for future improvements. ### Specific problem description - **Active fire - point detection**: Although existing MODIS and VIIRS products can provide high temporal resolution, they are easily affected by clouds and high - temperature buildings, resulting in false alarms. The paper hopes to improve the detection accuracy through a deep - learning model, especially by using time - series data. - **Burned - area mapping**: Existing MCD45, MCD64, and VNP64A1 products can only provide monthly burned - area information and have significant omission errors. The paper proposes to use high - resolution VIIRS data and a deep - learning model to achieve accurate daily burned - area mapping. - **Wildfire progress prediction**: Traditional semi - empirical methods such as FARSITE and Prometheus rely on physical knowledge and experimental observations but have limited accuracy. The paper predicts the future progress of wildfires by introducing a deep - learning model and combining satellite images and auxiliary data (such as weather, terrain, and land cover). ### Dataset characteristics The TS - SatFire dataset contains 3,552 surface reflectance images that occurred in the continental United States from January 2017 to October 2021, as well as auxiliary data such as weather, terrain, land cover, and fuel information, with a total of 71GB. The life cycle of each wildfire event is recorded, and active fire - points (AF) and burned - areas (BA) are labeled. In addition, the dataset also supports three tasks: active fire - point detection, daily burned - area mapping, and wildfire progress prediction. Through this dataset, researchers can train and evaluate deep - learning models that can handle these three tasks simultaneously, thereby promoting further development in the field of wildfire monitoring and prediction.

TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction

Wildfire Detection Through Deep Learning Based on Himawari-8 Satellites Platform

Wildland Fire Detection and Monitoring Using a Drone-Collected RGB/IR Image Dataset

Active fire detection in Landsat-8 imagery: A large-scale dataset and a deep-learning study

Sen2Fire: A Challenging Benchmark Dataset for Wildfire Detection using Sentinel Data

Assessing Sentinel-2, Sentinel-1, and ALOS-2 PALSAR-2 Data for Large-Scale Wildfire-Burned Area Mapping: Insights from the 2017–2019 Canada Wildfires

An open flame and smoke detection dataset for deep learning in remote sensing based fire detection

Near Real-Time Monitoring of Fire Spots Using a Novel SBT-FireNet Based on Himawari-8 Satellite Images

FireRisk: A Remote Sensing Dataset for Fire Risk Assessment with Benchmarks Using Supervised and Self-supervised Learning

Active Fire Detection Using a Novel Convolutional Neural Network Based on Himawari-8 Satellite Images

Deep Learning Approaches for Wildland Fires Using Satellite Remote Sensing Data: Detection, Mapping, and Prediction

Analyzing Multispectral Satellite Imagery of South American Wildfires Using Deep Learning

SWIFT: Simulated Wildfire Images for Fast Training Dataset

Development and Application of a Sentinel-2 Satellite Imagery Dataset for Deep-Learning Driven Forest Wildfire Detection

Real-Time Wildfire Detection Algorithm Based on VIIRS Fire Product and Himawari-8 Data

Autonomous Satellite Wildfire Detection Using Hyperspectral Imagery and Neural Networks: A Case Study on Australian Wildfire

Uni-temporal Sentinel-2 imagery for wildfire detection using deep learning semantic segmentation models

A Deep Learning Framework: Predicting Fire Radiative Power From the Combination of Polar-Orbiting and Geostationary Satellite Data During Wildfire Spread

An Adaptive and Extensible System for Satellite-Based, Large Scale Burnt Area Monitoring in Near-Real Time

Deep Learning for Global Wildfire Forecasting

The Canadian Fire Spread Dataset