Disaster reports stop at masks no longer: DisasterAgent ties every claim to image evidence and lifts report scores 13.5%
Synopsis
Remote sensing disaster assessment has long stopped at segmentation masks, bounding boxes and category labels, while responders need checkable written reports. DisasterAgent feeds bitemporal imagery to a vision expert toolkit, organizes the tool outputs into a spatiotemporal perception graph, and draws inline-cited evidence from that graph to drive a vision-language model that was never fine-tuned on disasters. On DisasterM3 its average disaster report generation score is 13.5% above the Qwen3.5-27B model serving as its central brain; on RSCC, ST5-SCS is 15.34% above that same base model.
Interpretation
The output of post-disaster remote sensing assessment moves from pixel-level masks and category labels to written reports whose statements can each be traced back to image evidence. Vision-language models previously generated end to end without exposing the visual evidence behind a statement, so checking whether a sentence was truly supported by the imagery was hard; existing remote sensing agents were mostly confined to single-image building counting or damage classification, and their tool outputs were isolated intermediate results. On RSCC, DisasterAgent reaches 67.18% ST5-SCS, above the RSCC-fine-tuned RSCCM-7B at 58.52%; against EarthAgent and OpenEarthAgent, which use the same expert toolkit, it also has the highest ST5-SCS.
The system takes a bitemporal image pair and a free-form query, has a pretrained vision-language model pick a skill and plan tool calls, then fuses the evidence into a report. A spatiotemporal perception graph acts as structured memory, holding pre-disaster and post-disaster region nodes together with cross-phase evolution edges instead of leaving tool results as disconnected fragments. Experiments run on the purely optical DisasterM3 Bench set and the RSCC test set, with Qwen3.5-27B as the central model, no disaster-specific fine-tuning, and greedy decoding.
Each key statement in the report carries an inline citation that maps back to specific nodes, evolution edges and confidence scores in the spatiotemporal perception graph. Semantic change evidence is first emitted as atomic claims from a perception brief, then scored by a separate visual verification call, and claims below the threshold are filtered out before final generation. On DisasterM3, 94.86% of 2,004 samples carry semantic change evidence citations; on RSCC the share is 85.32% of 988 samples, with each citation linked to 5.23 and 8.02 graph entities on average.
Perspective
The demonstrated scope is limited to bitemporal optical imagery: the purely optical DisasterM3 Bench set and the RSCC test set, covering change captioning and disaster report generation. The system targets emergency analysts who need to check the source of each conclusion, and reports carry a reference list for human tracing. The authors propose extending it to multimodal sources such as synthetic aperture radar and to multi-temporal disaster imagery.
Report quality is scored by an LLM judge on a five-point rubric rather than verified item by item by humans. The ablation shows restoration advice dropping from 3.15 to 2.96 when the spatiotemporal graph is added alone, recovering to 3.22 only with full evidence augmentation. In the qualitative case the base model misread an explosion as a tsunami, so upstream perception errors remain a risk source.