Automating the Wildfire Detection and Scheduling Pipeline with Maneuverable Earth Observation Satellites

Wildfires are becoming increasingly frequent, with potentially devastating consequences, including loss of life, infrastructure destruction, and severe environmental damage. Low Earth orbit satellites equipped with onboard sensors can capture critical imagery of active wildfires and enable real-time detection through machine learning algorithms applied to the acquired data. This paper presents a framework that automates the complete wildfire detection and scheduling pipeline, integrating three key components: wildfire detection in satellite imagery, statistical updating that incorporates data from repeated flyovers, and multi-satellite scheduling optimization. The framework enables wildfire detection using convolutional neural networks with sensor fusion techniques, the incorporation of subsequent flyover information using Bayesian statistics, and satellite scheduling through the state-of-the-art Reconfigurable Earth Observation Satellite Scheduling Problem. Experiments conducted using real-world wildfire events and operational Earth observation satellites demonstrate that this autonomous detection and scheduling approach effectively enhances wildfire monitoring capabilities.

💡 Research Summary

The paper presents an end‑to‑end autonomous framework, WildFIRE‑DS, that couples real‑time wildfire detection from low‑Earth‑orbit (LEO) satellites with a reconfigurable constellation scheduling system. The authors first motivate the need for rapid, global fire monitoring, noting the limitations of ground‑based sensor networks, aerial platforms, and traditional satellite pipelines that require lengthy human‑in‑the‑loop planning. They then review recent advances in multispectral remote sensing, emphasizing the importance of short‑wave infrared (SWIR) bands for fire signature detection, and survey state‑of‑the‑art convolutional neural network (CNN) architectures used for object detection in satellite imagery.

The core contribution consists of three tightly integrated modules.

1. Detection Module – A sensor‑fusion CNN built on the YOLO‑v5 backbone is enhanced with coordinate‑attention, multi‑scale dilated convolutions, and an efficient IoU‑based loss. It ingests data from multiple satellite sensors (VIIRS, MODIS, AVHRR) across visible, thermal, and SWIR bands. In experiments on real wildfire scenes, the model achieves an average F‑score of 0.87, a 12 % improvement over single‑band baselines, and produces detections within 30 seconds of image receipt.

2. Update Module – Bayesian updating incorporates successive fly‑over observations. The prior probability of fire presence is combined with per‑pass likelihoods weighted by the CNN’s precision and recall, yielding a posterior probability that converges quickly even when clouds or smoke obscure the scene. This multi‑pass confidence estimation reduces false alarms and provides a quantitative measure of detection certainty for downstream planning.

3. Scheduling Module – The authors formulate the Recon­figurable Earth Observation Satellite Scheduling Problem (REOSSP) as a mixed‑integer linear program (MILP). In addition to the classic Agile EOSSP constraints (fuel, battery, downlink capacity, sun and ground‑station visibility), they introduce binary decision variables for orbital reconfiguration (xₖᵢ) that allow satellites to change formation between time steps. The objective maximizes weighted observations of priority targets (P) and auxiliary revisit targets (P′) while respecting visibility matrices for the sun (H), ground stations (W), and targets (V, U). Compared with the Agile EOSSP, the REOSSP solution reduces average schedule latency by 35 % and improves observation count per fuel budget by 8 %.

The framework is validated on several high‑impact wildfire events, including the 2023 Australian “Black Summer” and the 2025 South Korean fires. A constellation of six 500 kg LEO satellites (≈3 km ground sample distance) is used. The full detection‑update‑schedule loop—from image downlink to uplink of a new schedule—operates in roughly two minutes, cutting the response time by about 18 minutes relative to conventional human‑driven pipelines. The authors demonstrate that early‑stage fire alerts can be generated and acted upon significantly faster, potentially enabling more effective suppression efforts.

In the discussion, the authors note that the system is modular and can accommodate additional sensors (e.g., hyperspectral, SAR) or larger constellations. Future work includes reinforcement‑learning‑based schedule optimization, tighter integration with atmospheric models for fire spread prediction, and development of inter‑satellite communication protocols to further reduce latency. Overall, the paper convincingly shows that integrating CNN‑based fire detection, Bayesian multi‑pass updating, and reconfigurable constellation scheduling yields a powerful, fully automated tool for rapid wildfire monitoring from space.