AI-driven microcharcoal recognition: Advancing intelligent automation for paleofire reconstruction

Charcoal in sediments is a reliable proxy for paleofire, playing a key role in reconstructing past fire activity and its links to climate and vegetation. However, traditional manual or semi-automated methods for identifying microcharcoal (MC) are constrained by limitations in efficiency, consistency and robustness. In this study, we developed an AI-driven software tool (FireFly) for automatic MC recognition, leveraging the advanced YOLOv8 (You Only Look Once, version 8) deep learning algorithm. This represents the first application capable of simultaneously recognizing both MC and the marker Lycopodium clavatum spores (LYP), while also enabling automated result calculation and export. The model was trained on an expert-validated dataset comprising 21,094 MC particles, 6,214 LYP particles, and 19,964 impurity particles (including mineral fragments, amorphous organic debris, and other microfossils). Performance evaluations on a dedicated test dataset and two long-term sedimentary sequences demonstrated high concordance with manual identification results, achieving an accuracy of ∼94.3% for MC and ∼ 98.6% for LYP. Notably, the software completes recognition and quantification of MC for an entire pollen slide in just 20 s, marking a substantial efficiency improvement compared to the several-hour-long manual process. These results confirm that the proposed method is accurate and outperforms traditional approaches in efficiency. Moreover, it enables the reconstruction of wildfire history across large spatial scales, thereby laying a robust foundation for research into the wildfire-ecology-climate nexus.