Title: ForestHG-Trace: Enabling Traceable Long-Horizon Ecological Reasoning in Extensive Forest Landscapes

Abstract:

Remote sensing question answering (RS-QA) frequently demands capabilities that extend beyond simple semantic prediction. This is particularly true in vast forest environments, where ecological analysis necessitates multi-step filtering, numerical aggregation, neighborhood reasoning, and the verification of evidence. To address these challenges, we present ForestHG-Trace, a novel framework designed for traceable, long-horizon ecological reasoning within forest ecosystems.

Our approach models multimodal NEON forest scenes as ecological hypergraphs. Unlike traditional pairwise scene graphs, this structure incorporates tree instances, spatial units, semantic groups, and neighborhood relations to facilitate higher-order reasoning. Guided by an LLM-based agent, the system utilizes deterministic tools to perform reading, filtering, expansion, aggregation, comparison, and auditing. This process generates compact evidence records and replayable execution traces, moving beyond the generation of free-form answers alone.

Additionally, we introduce ForestTraceQA, an executable benchmark established to evaluate ecological QA performance across various task categories and reasoning depths. Our experimental results demonstrate that ForestHG-Trace significantly enhances both answer accuracy and execution faithfulness compared to single-step baselines and scene-graph agents. Furthermore, the study identifies execution depth as the primary bottleneck for long-horizon ecological QA.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC