Residual trend method based on regional modeling and machine learning for attribution of vegetation changes

Existing residual trend methods utilize a pixel-by-pixel modeling strategy, in which the ordinary least squares method is employed. These methods suffer certain limitations. On the one hand, the pixel-by-pixel modeling strategy causes each model to contain signal interference from human activities in local space. On the other hand, the ordinary least squares method is unfavorable for simulating commonly observed nonlinear characteristics. This study proposed an entirely new residual trend method based on regional modeling and machine learning. Besides, this study compared two types of environmental variables used to express spatial heterogeneity: ①direct-environmental variables (DEVs) such as terrain, hydrology, and land use; and ②proxy-environmental variables (PEVs) that combine the spatiotemporal series of vegetation and climate. First, a regional modeling strategy was adopted. After DEVs and PEVs were introduced individually, models for the vegetation-climate relationship were built using machine learning. Second, residuals were determined based on the definition of the residual trend method. Finally, the contributions of anthropogenic and climatic factors to vegetation change were assessed. The results indicate that compared to the previous pixel-by-pixel residual trend method that utilizes ordinary least squares, the new residual trend method can simulate the nonlinear features of the vegetation-climate relationship and exhibits enhanced resistance to human signal interference. For the new method, significantly higher performance can be achieved using PEVs compared to DEVs. PEVs can fully utilize the original modeling data, without increasing difficulties with data acquisition and avoiding additional data errors. The residual trend method based on regional modeling and machine learning proposed in this study allows for more effective attribution of vegetation changes.