CPD-YOLOv8: A Causality-Driven Physical Modality Fusion Framework for Early Micro-Fire Detection to Protect Complex...

The early detection of wildland forest fires is a crucial task for protecting vulnerable plant ecosystems, yet it remains physically challenging due to the extreme limitations imposed by micro-scale targets and severe environmental interference within dense vegetation. Traditional monocular vision systems typically rely on statistical fitting and are highly prone to false alarms when facing complex optical interference from vegetation canopies, such as penetrating sunlight through foliage and specular reflections. Although multispectral fusion aims to mitigate this issue, existing dual-stream architectures frequently fall into the "heterogeneous negative transfer" trap, where modal conflicts and misaligned physical priors degrade the detection performance. To address these issues, this study introduces a causality-driven and physical-feature-decoupled YOLO (CPD-YOLOv8) framework. Adhering to the principle of physical fidelity, we implement an early R-G-NIR channel replacement strategy to ensure the fundamental physical alignment of spatial and thermodynamic properties. Furthermore, an Adaptive Cross-Modal Attention (ACMA) module is incorporated to facilitate non-destructive feature decoupling, and a Regional Causal-Preserving Downsampling (RCPD) module is designed to intercept environmental confounders (e.g., high-frequency leaf reflections) at the local micro-scale. This framework explores the transition of detection methods from pure statistical correlation toward physical causality. Experimental results on the Corsican wildland multispectral forest fire dataset demonstrate that CPD-YOLOv8 yields an mAP@0.5 of 70.3% and a precision of 87.1%. By mitigating the constraint between precision and recall, this method provides a low-false-alarm solution, offering a technical reference for proactive vegetation management and the reliable deployment of fire early-warning algorithms in real-world complex wildland terrains.