<p>Slope instability and the reactivation of pre-existing landslides in Xiji County, Ningxia Hui Autonomous Region, China, are driven by regional surface changes such as urban expansion, farmland construction and climatic triggers. Accurately assessing spatio-temporal changes in landslide susceptibility and risk under these evolving conditions is crucial for effective risk management. Current approaches often face challenges in effectively integrating complex machine learning models like random forest (RF) and back propagation neural networks (BPNN) to enhance predictive accuracy. Furthermore, utilizing historical data to reliably predict future landslide risk under dynamically changing factors remains a significant difficulty. To address these challenges, an integrated methodology combining RF and BPNN within an entropy-weighted framework is proposed. This model also incorporates key dynamic factors including land use and land cover (LULC) change, population density, temperature variation, and rainfall. This approach enables robust assessment of landslide susceptibility across both historical (2000—2020) and future scenarios. Results demonstrate that the entropy-weighted integrated model outperforms single models. High susceptibility zones are concentrated within the loess hills of the Hulu River Basin. Analysis reveals an overall decrease in high-susceptibility areas from 2000 to 2020. Projections indicate this decreasing trend continues under future scenarios, despite increased regional rainfall (primarily summer), due to concurrent decreases in population density and arable land. This integrated analysis provides critical insights for developing effective landslide risk mitigation and management strategies in Xiji County and similar regions.</p>