Abstract:

As one of China’s major pastoral regions, Xinjiang faces significant threats from grassland degradation, particularly poisonous weed proliferation, which adversely affects animal husbandry and ecological stability. Traditional remote sensing techniques encounter limitations in accurately identifying poisonous weed species. Conversely, machine learning models integrating multisource environmental factors offer substantial improvements in prediction accuracy concerning the poisonous weed distribution. This study employed a Random Forest model to analyze ecological drivers and poisonous weed distribution data, aiming to identify the primary environmental variables influencing the spatial patterns of five representative poisonous weed species in Xinjiang. Furthermore, the potential distribution areas of these species were predicted under current (1970-2020) and near-future (2021-2040) climatic conditions based on different shared socioeconomic pathway scenarios. The model demonstrated high predictive performance, with evaluation metrics, including Accuracy, Precision, Recall, and F1-score, all exceeding 0.90, confirming its robustness and generalization capability. The key ecological variables influencing poisonous weed distribution include isothermality, temperature seasonality, and elevation. According to the model projections, the current and future potential distribution areas are predominantly located in northern Xinjiang, particularly in Altay, Urumqi, Changji, Tacheng, and Ili. Under the SSP126 scenario, the poisonous weed distribution exhibited relatively stable spatial shifts, with slight northward and southward movements and shorter migration distances. Contrarily, under the SSP245 scenario, increased ecological stress portended greater distributional fluctuations, longer migration distances, and a notable decline in ecosystem stability. This research highlights the importance of integrating ecological and climatic variables with machine learning approaches for effective species distribution modeling. The findings provide valuable insights into the ecological behavior of poisonous weeds and offer a scientific basis for regional grassland management strategies in Xinjiang. By forecasting the spatial responses of poisonous species under varying climate change scenarios, this study supports the development of adaptive management plans to mitigate the ecological and economic effects of poisonous weed encroachment in arid and semi-arid pastoral systems.

Key words: Random Forest model, ecological factors, poisonous weeds, potential distribution area