Rock mass weakness planes are mechanically unfavored discontinuities that control slope stability. However, their spatial pattern is poorly understood, limiting advances in regional landslide assessment. In this study, we propose a multi-explainable machine learning framework to predict the distribution of weakness planes and quantify their contribution to landslides. Focusing on southeastern Tibet, the study investigates 194 field outcrops of rock mass weakness planes, integrating geological section comparisons and driving factor analyses to reveal spatial heterogeneity and the evolvement of weakness plane. Rock mass weakness planes are denser in lithologically weak and structurally damaged rock masses, and less developed in gentler, sparsely faulted terrains. Comparative analysis of geological sections indicates that pre-existing discontinuities are essential preconditions for developing weakness planes, whereas precipitation acts as an activator. The controlling factors vary with lithology, indicating that material and structure govern how efficiently exogenic processes transform discontinuities into weakness planes. By considering weakness plane during susceptibility assessment, performances were improved (Recall increased by 4.6-6.8%, AUC increased by 3.1-5.6%). This work constructs the regional-scale model for continuous prediction of rock mass weakness planes. It links lithology-dependent formation mechanisms of rock mass weakness plane to slope instability, providing a process-based framework for interpreting landslide development and improving landslide susceptibility assessment.