Change in Fractional Vegetation Cover and Its Prediction during the Growing Season Based on Machine Learning in Southwest China

Abstract

Fractional vegetation cover (FVC) is a crucial indicator for measuring the growth of surface vegetation. The changes and predictions of FVC significantly impact biodiversity conservation, ecosystem health and stability, and climate change response and prediction. Southwest China (SWC) is characterized by complex topography, diverse climate types, and rich vegetation types. This study first analyzed the spatiotemporal variation of FVC at various timescales in SWC from 2000 to 2020 using FVC values derived from pixel dichotomy model. Next, we constructed four machine learning models—light gradient boosting machine (LightGBM), support vector regression (SVR), k-nearest neighbor (KNN), and ridge regression (RR)—along with a weighted average heterogeneous ensemble model (WAHEM) to predict growing-season FVC in SWC from 2000 to 2023. Finally, the performance of the different ML models was comprehensively evaluated using tenfold cross-validation and multiple performance metrics. The results indicated that the overall FVC in SWC predominantly increased from 2000 to 2020. Over the 21 years, the FVC spatial distribution in SWC generally showed a high east and low west pattern, with extremely low FVC in the western plateau of Tibet and higher FVC in parts of eastern Sichuan, Chongqing, Guizhou, and Yunnan. The determination coefficient R2 scores from tenfold cross-validation for the four ML models indicated that LightGBM had the strongest predictive ability whereas RR had the weakest. WAHEM and LightGBM models performed the best overall in the training, validation, and test sets, with RR performing the worst. The predicted spatial change trends were consistent with the MODIS-MOD13A3-FVC and FY3D-MERSI-FVC, although the predicted FVC values were slightly higher but closer to the MODIS-MOD13A3-FVC. The feature importance scores from the LightGBM model indicated that digital elevation model (DEM) had the most significant influence on FVC among the six input features. In contrast, soil surface water retention capacity (SSWRC) was the most influential climate factor. The results of this study provided valuable insights and references for monitoring and predicting the vegetation cover in regions with complex topography, diverse climate types, and rich vegetation. Additionally, they offered guidance for selecting remote sensing products for vegetation cover and optimizing different ML models.

Keywords

• feature importance analysis
• FVC
• heterogeneous ensemble model
• machine learning
• southwest China
• spatiotemporal variation characteristics