The bottom line: how fast a drone fleet spots a wildfire depends less on clever routing math and more on whether the risk map guiding it is any good. That is the central finding of WFDroneBench, a new open-source benchmark that tests sensor placement and drone routing strategies against thousands of simulated wildfires, and it has real consequences for fire agencies deciding where to spend limited monitoring budgets.
Wildfire monitoring today mixes fixed ground sensors, satellite passes, and increasingly, drones that can fly to a suspected ignition point and confirm it visually. The problem is a layered one: where do you place stationary sensors and charging stations before a fire starts, and once a fire risk signal appears, how do you route a limited number of drones to check it out fastest? WFDroneBench, built by researchers releasing their code and data publicly, is the first standardized way to test answers to both questions side by side, rather than in isolation.
The benchmark works by combining two ingredients that used to be evaluated separately: machine-learned wildfire risk maps, which estimate the probability of ignition across a landscape, and optimization strategies that decide where infrastructure goes and how vehicles move. It draws its fire-spread scenarios from Sim2Real-Fire, a dataset built from established simulators such as FARSITE and WRF-SFIRE, and layers in real historical ignition records across 49 locations, producing 7,746 distinct test scenarios in total.
On the routing side, the core problem drones face resembles a classic operations research puzzle called the orienteering problem: given a set of candidate sites with associated risk (reward) and a time or distance budget per drone, choose a subset of sites and a visiting order that maximizes the expected value of what gets covered before time runs out. For a fleet of $K$ drones with route duration limit $T$, this is naturally a Team Orienteering Problem, roughly: $$\max \sum_{k=1}^{K}\sum_{i \in R_k} r_i \quad \text{s.t.} \quad \text{time}(R_k) \le T \;\; \forall k, \quad R_k \cap R_{k'} \text{ managed to avoid excess overlap}$$ where $r_i$ is the risk score at site $i$ and $R_k$ is the route assigned to drone $k$. In plain terms: send each drone on the path that visits the highest-risk locations it can reach before its battery or a decision deadline forces it back.
The benchmark compares this Team Orienteering Problem approach and a simpler risk-weighted Max-Coverage heuristic against baseline strategies (including two ground-detector configurations) across scenarios of varying difficulty. According to the results reported in the paper and its OpenReview listing, the risk-aware strategies — Max-Coverage and TOP — significantly outperform simpler baselines when the risk map is sufficiently accurate, with TOP achieving the fastest detection on the hardest fires in one version of the results. Drone-based detection also consistently beat ground-sensor-only setups, and pre-positioning risk-aware static infrastructure helped even when the risk map wasn't fully accurate.
That last point matters more than it might first appear. It means an agency doesn't need a perfect predictive model to benefit from optimization — a decent risk estimate paired with a smart placement and routing strategy still beats naive, uniform coverage. But the benchmark also surfaces a harder truth: when risk maps are poor, the sophistication of the routing algorithm stops mattering much, because the drones are being sent to the wrong places in the first place.
The paper flags two open problems that should interest both the ML and OR communities. First, small fires remain stubbornly hard to detect quickly and reliably, regardless of strategy — they're the scenarios where every approach in the benchmark struggles most. Second, and more fundamentally, there's a measurable gap between the ground-truth ignition patterns in the simulation data and what today's risk maps actually predict, which the authors frame as an opportunity for better ML models rather than better routing algorithms.
For optimization practitioners, the takeaway is a reminder that routing and siting algorithms are only as good as the data pipeline feeding them. A Team Orienteering Problem solver tuned to perfection is wasted effort if the risk scores it's optimizing against are noisy or systematically wrong. WFDroneBench, by open-sourcing its code, scenario data, and evaluation metrics (time to detection, fire size at detection, and drone distance traveled), gives both ML researchers and OR practitioners a shared benchmark to iterate against — and a concrete way to measure whether a new risk model or a new routing heuristic is actually worth deploying before it gets tested on a real fire season.
Sources: Puech et al., 'WFDroneBench: A Benchmark for Sensor Placement and Drone Routing for Wildfire Detection', arXiv:2609.11829 · WFDroneBench OpenReview listing, https://openreview.net/forum?id=A8UKo53z3R · RomainPuech/wildfire_drone_routing GitHub repository, https://github.com/RomainPuech/wildfire_drone_routing · 'Sim2Real-Fire: A Multi-modal Simulation Dataset for Forecast and Backtracking of Real-world Forest Fire', NeurIPS 2024 Datasets and Benchmarks Track
