Improved Wildfire Spread Prediction with Time-Series Data and the WSTS+ Benchmark

TL;DR

This paper conducts a rigorous, controlled comparison of data-driven wildfire spread models under single-day (T=1) and time-series (T=5) inputs using the public WildfireSpreadTS (WSTS) benchmark, and introduces an extended WSTS+ benchmark with four additional years of data. It shows that time-series inputs generally yield higher accuracy, with the time-series UTAE models and a Res18-Unet family achieving new state-of-the-art performance on WSTS, including a highest overall AP of 0.478 for Veg features. To enable broader research, the authors expand the benchmark (WSTS+) and propose a four-fold cross-validation scheme, but find that cross-year covariate shift limits gains from more data, indicating domain shift as a fundamental challenge for scaling data-driven wildfire modeling. The work highlights the importance of temporal information, careful model tuning, and domain-shift awareness for deploying predictive wildfire models in real-world settings.

Abstract

Recent research has demonstrated the potential of deep neural networks (DNNs) to accurately predict wildfire spread on a given day based upon high-dimensional explanatory data from a single preceding day, or from a time series of T preceding days. For the first time, we investigate a large number of existing data-driven wildfire modeling strategies under controlled conditions, revealing the best modeling strategies and resulting in models that achieve state-of-the-art (SOTA) accuracy for both single-day and multi-day input scenarios, as evaluated on a large public benchmark for next-day wildfire spread, termed the WildfireSpreadTS (WSTS) benchmark. Consistent with prior work, we found that models using time-series input obtained the best overall accuracy, suggesting this is an important future area of research. Furthermore, we create a new benchmark, WSTS+, by incorporating four additional years of historical wildfire data into the WSTS benchmark. Our benchmark doubles the number of unique years of historical data, expands its geographic scope, and, to our knowledge, represents the largest public benchmark for time-series-based wildfire spread prediction.

Figures (24)

• Figure 1: The wildfire prediction models take as input a geospatial map of several variables: vegetation, topography, and weather features, alongside the current day fire mask. We consider two scenarios: one in which the model receives input features from one preceding day, denoted $t-1$, and one in which it receives input from five previous days
• Figure 2: Sample predictions made by the Res18-Unet gerard2023wildfirespreadts, our Res18-Unet, and Res50-Unet. The two leftmost columns show the current fire spread $y(t-1)$ and the next-day label $y(t)$. True positive pixels are colored in green, while false positives are colored in red
• Figure 3: Geographic distribution of the fire events in each year of WSTS (blue) and WSTS+ (red)
• Figure 4: New cross-validation folds used for WSTS+. Each pair of consecutive years is used as validation/testing once. Color code: blue: training, orange: validation, green: test
• Figure 5: Performance breakdown by test year. Blue bars represent models trained on the original WSTS data, termed Res18-Unet(WSTS), while red bars represent those trained on WSTS+, termed Res18-Unet(WSTS+). The bolded x-axis ticks highlight original test years from WSTS. For Res18-Unet(WSTS+), we stratify its performance by year. For Res18-Unet(WSTS), we stratify by year to obtain performance for 2018 to 2021. To obtain performance on the remaining years, we select the cross-validation fold with the best-performing model (as judged by its test fold error) and report its performance on the newly added WSTS+ years.