Health assessment of poplar plantations in the Horqin Sandy Land, Inner Mongolia of northern China based on ChatGPT

Objective Conventional forest health assessment methods face limitations when handling complex, high-dimensional data, making it difficult to accurately reflect the true state of forest health. To address the issue, this paper proposes a novel assessment framework integrating the ChatGPT large language model with machine learning algorithms to improve forest health assessment and identify key indicators affecting the health of poplar (Populus spp.) plantations in the Horqin Sandy Land, Inner Mongolia of northern China.

Method A comprehensive health evaluation system comprising 15 indicators was developed using field survey data and forest landscape images. The ChatGPT model was applied for health level prediction via few-shot learning. YOLOv5 was used to extract visual features from images to enhance assessment accuracy. K-means clustering was employed to automatically classify forest health levels. Additionally, the predictions from ChatGPT were cross-validated and compared using the DeepSeek large language model, a fuzzy comprehensive evaluation method, and eight machine learning models, respectively. SHAP analysis was used to identify the main factors affecting forest health.

Result The proportions of forest health levels in poplar plantations were as follows: moderate health (53.3%) > unhealthy (21.7%) > sub-healthy (20.7%) > healthy (4.3%), indicating an overall moderately healthy forest condition. The prediction accuracy of DeepSeek model was 76.1%. No significant difference was observed between the results of ChatGPT and the traditional fuzzy evaluation method (p = 0.29). Among the eight machine learning models, the random forest classifier achieved the highest validation accuracy of 84.2%. Stand mean height and soil organic carbon density were identified as the main factors influencing forest health.

Conclusion The proposed ChatGPT-based evaluation approach, integrated with multi-source data, is scientifically reliable and effectively enhances both the accuracy and interpretability of forest health classification. This method provides a new pathway for monitoring and managing the health of plantations in arid and semi-arid regions.