Rapid Wildfire Hotspot Detection Using Self-Supervised Learning on Temporal Remote Sensing Data

TL;DR

This work tackles rapid wildfire hotspot detection by leveraging self-supervised learning on multi-temporal MSG15 remote-sensing sequences to learn robust hotspot representations without extensive labeling. It introduces a Europe-focused time-series hotspot dataset by integrating EFFIS burn-area data, MSG15 imagery, and land-cover information, and adapts a Presto-inspired SSL framework to process temporal pixel sequences with a three-component loss $L_{tot} = L_{eo} + L_{lc} + L_{cls}$. The model achieves a maximum test $F_1$ score of $F_1 = 63.58$ with a Cosine scheduler, demonstrating the effectiveness of temporal SSL for near real-time hotspot detection and providing a public benchmark for future research. The work highlights the potential of SSL to improve transferability and scalability of wildfire monitoring across large regions with limited labeled data.

Abstract

Rapid detection and well-timed intervention are essential to mitigate the impacts of wildfires. Leveraging remote sensed data from satellite networks and advanced AI models to automatically detect hotspots (i.e., thermal anomalies caused by active fires) is an effective way to build wildfire monitoring systems. In this work, we propose a novel dataset containing time series of remotely sensed data related to European fire events and a Self-Supervised Learning (SSL)-based model able to analyse multi-temporal data and identify hotspots in potentially near real time. We train and evaluate the performance of our model using our dataset and Thraws, a dataset of thermal anomalies including several fire events, obtaining an F1 score of 63.58.

Figures (3)

• Figure 1: Geographical distribution of train, validation and test split of our dataset.
• Figure 2: How we build positive (greed) and negative (red) pixel timeseries, starting from MSG15 native files.
• Figure 3: Architecture of the proposed approach.