Rapid identification of hazardous events is essential for next-generation Earth Observation (EO) missions supporting disaster response. However, current monitoring pipelines remain largely ground-centric, introducing latency due to downlink limitations, multi-source data fusion constraints, and the computational cost of exhaustive scene analysis. This work proposes a hierarchical multi-agent architecture for onboard EO processing under strict resource and bandwidth constraints. The system enables the exploitation of complementary multimodal observations by coordinating specialized AI agents within an event-driven decision pipeline. AI agents can be deployed across multiple nodes in a distributed setting, such as satellite platforms. An Early Warning agent generates fast hypotheses from onboard observations and selectively activates domain-specific analysis agents, while a Decision agent consolidates the evidence to issue a final alert. The architecture combines vision-language models, traditional remote sensing analysis tools, and role-specialized agents to enable structured reasoning over multimodal observations while minimizing unnecessary computation. A proof-of-concept implementation was executed on the engineering model of an edge-computing platform currently deployed in orbit, using representative satellite data. Experiments on wildfire and flood monitoring scenarios show that the proposed routing-based pipeline significantly reduces computational overhead while maintaining coherent decision outputs, demonstrating the feasibility of distributed agent-based reasoning for future autonomous EO constellations.
Figures & tables
Figure 1: Event-driven hierarchical architecture activated upon candidate event detection.
Figure 2: Two wildfire specialist nodes with different sensing modalities (TIR and HSI) observing the same area of interest after candidate wildfire detection by the Early Warning node.
Figure 3: Diamond-Topology Multi-Agent Demonstrator and information flow between agents.
Figure 4: (B12, B11, B8) composite from a Sentinel-2 scene used to enhance the visualization of active wildfire signatures. The segmentations produced by the three proposed methods are shown for the same scene.
Figure 5: Segmentation of floods and permanent water using the ML model, achieving an IoU 0.50flood of 0.554 on the CEMS dataset from SenForFlood. In the segmentation mask, permanent water and flood water are shown in blue and green, respectively.
Event
Speed-up
Reduction (%)
No event
4.78±2.54
73.2±14.1
Event (wildfire or flood)
1.3±0.45
13.5±30.8
Table 1: Average speed-up (mean ± standard deviation) measured on the testbench for 27 samples, grouped by event presence
Figure 6: Scatter plot of speed-up versus scene area for event and non-event samples. Lines correspond to linear regression fits for each group.
Figure 7: Example where specialist evidence contradicts the Early Warning prediction. The scene is shown using Sentinel-2 RGB and SWIR false-color composites (B12, B11, B8). The wildfire specialist finds no supporting evidence, leading the decision agent to reject the alert.
Figure 8: Comparison of wildfire reports generated by the baseline configuration (two specialists and a decision agent) and the proposed hierarchical pipeline. The scene is shown using Sentinel-2 RGB and SWIR false-color composites (B12, B11, B8). Both configurations detect the wildfire, but the proposed pipeline provides a more focused explanation aligned with the routed specialist evidence.
Figure 9: Comparison of flood reports generated by the baseline configuration (two specialists and a decision agent) and the proposed hierarchical pipeline. The scene is shown using Sentinel-2 RGB and SAR false-color composites (VV, VH, VV/VH). Both detect the flood, but the proposed pipeline provides a richer semantic explanation of the event and its potential impact.
Figure 10: Comparison between the baseline configuration (two specialists and a decision agent) and the proposed hierarchical pipeline in a scene containing previously burned areas. The Sentinel-2 scene is shown using RGB and SWIR false-color composites (B12, B11, B8). The baseline interprets the burn scar as an active wildfire, whereas the proposed pipeline correctly identifies it as a past event with no active fire.
State Key Laboratory of Space Information System and Integrated Application, Beijing 100095, China · Beijing Institute of Satellite Information Engineering, Beijing 100095, China · Shenzhen International Graduate School, Tsinghua University, Shenzhen, China +2