Objective diagnosis of machine learning method applicability to land comprehensive carrying capacity evaluation: A case study based on integrated RF and DPSIR models

Abstract

The evaluation of land comprehensive carrying capacity (LCCC) is a popular topic in regional sustainable development and land science research. However, these evaluations are often not objective due to the complexity and nonlinear characteristics of the LCCC evaluation indicators. This study provides a new framework for using machine learning methods to objectively evaluate LCCC. The Driver-Pressure-State-Impact-Response (DPSIR) indicator conceptual framework was used for the spatial visualisation of indicators and was then combined with the random forest (RF) model to optimise the LCCC evaluation. The results showed the following: (1) The accuracy of the integrated RF and DPSIR model was better than that of traditional support vector machine (SVM) and principal component analysis (PCA) methods, indicating that RF model was more suitable for LCCC evaluations. (2) The contribution of the DPSIR indicators after RF optimisation was significantly different from that of the traditional analytic hierarchy process (AHP), the contribution of the influence subsystem (I) (48.7%) was enhanced, while the contribution of the drive subsystem (D) (6.8%) was weakened, demonstrating that after RF optimisation, the DPSIR was more conducive for handling complex nonlinear systems and objectively reflecting indicator contributions. (3) The objective attributes of the DPSIR indicators were optimised through the processing of spatial visualisation models, indicating that RF was suitable for processing evaluations with a multidimensional spatial heterogeneity index system. The LCCC evaluation model can be applied to other carrying capacity case studies and operational processes and provide a scientific method for handling the multidimensional nonlinear evaluation index system and objective evaluation of LCCC.