Using XGBoost-SHAP for understanding the ecosystem services tradeoff effects and driving mechanisms in ecologically fragile areas

Peiyu Du ^1,2 Heju Huai ² Xiaoyang Wu ³ Hongjia Wang ² Wen Liu ² Xiumei Tang ^2*

• ¹ National Engineering Research Center for Information Technology in Agriculture, Shandong Agricultural University, Taian, China
• ² Beijing Research Center for Information Technology in Agriculture, Beijing, Beijing Municipality, China
• ³ Shandong Agricultural University, Taian, China

Understanding the spatial and temporal variability of Ecosystem services (ES), along with the tradeoffs and synergies among different services, is crucial for effective ecosystem management and sustainable regional development. This study focuses on Wensu, Xinjiang, China, as a case study.Ess and their trade-offs were assessed from 1990 to 2020 there. Additionally, explainable machine learning models were used to quantify the nonlinear effects and threshold effects of ES trade-offs to explore their driving factors. The findings indicate that: (1) From 1990 to 2020, there was an inverted "N"-shaped downward trend in water yield (WY) and soil conservation (SC) in Wensu County. The mean annual WY decreased from 22.99 mm to 21.32 mm, and SC per unit area dropped from 1440.28 t/km² to 1351.3 t/km². Conversely, the quantity of windbreak and sand fixation (WS) showed a "N"-shaped trend, increasing from 2.32×10⁷t to 3.11×10⁷t. The habitat quality (HQ) index initially improved, then declined, with values of 0.596, 0.603, 0.519, and 0.507, respectively. (2) The relationships between WY-WS, WY-HQ, WS-HQ, SC-WS, and SC-HQ are primarily characterized by trade-offs, while the relationship between WY-SC is mainly driven by synergistic interactions.Trade-offs between SC-HQ, WY-HQ, and WS-HQ are notably stronger, whereas the trade-off between WY and SC is weaker. (3) The main contribution of this study is to identify the nonlinear and threshold effects of the drivers behind the trade-offs in ESs. The XGBoost-SHAP model identified land use type (Land), precipitation (Pre), and temperature (Tem) as the dominant drivers of trade-offs. And the drivers showed significant nonlinear responses and threshold effects. For example, the WY-SC trade-off is exacerbated when precipitation exceeds 17 mm, while the temperature threshold determines the variation between the WY-HQ trade-off and the synergistic state. The model's ability to capture these nonlinearities, along with its interpretability, provides a clearer understanding of the complex mechanisms driving ESs. The findings of this study can be applied to other ecologically vulnerable regions, offering valuable guidance for ecosystem management and conservation strategies in areas facing similar ecological challenges.