Unveiling the Spatio-Temporal Patterns and mechanisms of poverty incidence in Southwest China through multi-sourced data and interpretable machine learning

China's poverty alleviation practice stands out globally, yet illuminatingly decoding its complex process remains challenging due to data and method fragmentation. This study proposed an integrated analytical framework combining multi-source data fusion, ensemble learning, and interpretable machine learning. By integrating multi-dimensional socioeconomic survey (SES) data with spatiotemporally continuous nighttime light (NTL) observations, the framework enables robust poverty prediction and mechanistic insights in data-rich and data-scarce contexts. The method was applied to examine poverty reduction dynamics in Yunnan-Guizhou-Guangxi area (YGGA) of Southwest China. Across 328 county-level units, the average poverty incidence markedly decreased from 33.38% to 12.42% (2000–2019), characterized by three phases: widespread high poverty (2000–2005), uneven regional improvement (2006–2012), and transformative poverty reduction (2013–2019). Spatio-temporal analysis uncovered the transformation from highly clustered poverty to a more dispersed distribution. Through interpretable machine learning, the study analyzed 24 driving factors in three categories. While economic & demographic indicators became increasingly dominant, the persistent influence of geographical & environmental, and social & infrastructure indicators underscored the necessity for an integrated approach to poverty governance. This study provided insights for China's post-poverty alleviation era while contributing towards inclusive growth within global sustainable development framework.