The dynamics of rural human-environment systems in developing countries have increasingly been attracting attention. Agent-based modeling (ABM) is a valuable simulation tool for detecting complex feedback loops in rural human-environment systems with a 'bottom-up' approach. However, such models require the prerequisite analysis of household typology to simulate households' decision-making process, where a gap exists between having accurate classification criteria and a simplified modeling framework. This study aimed to develop a household typology for two selected counties in southwest China based on multivariate analysis techniques and the classification tree method. Four categories of socioeconomic variables, including labor conditions, resource endowments, economic status, and social connections, were screened as possible factors impacting agriculture practice decisions. The results showed that household diversification in the study area was mainly determined by diversified livelihood strategies of off-farm work, livestock breeding, subsidy dependence, and traditional planting. Five distinct household types were identified: non-farm households, part-time households, livestock breed households, subsidized households, and traditional planting households. The household types were associated with specific cropland use behaviors, and their decision-making behaviors were verified with bounded rationality theory (where the maximization of profits is the primary goal). The quantitative classification criteria obtained in this study were clear and could be easily identified and used by ABMs. Our study provides a basis for further simulation of the complicated rural human-environment systems in southwest China.