This work aims to support emergency command decision-making by enabling rapid detection of landslides and volume changes through overcoming the traditional dependency on dual-temporal topographic data. We proposed a method that integrates instance segmentation in deep learning-based machine vision, Kriging interpolation of regular grid remote sensing images, and pixel volume integration. It achieves rapid 3D pre-landslide terrain reconstruction using single-temporal aerial high-resolution visible remote sensing data. A progressive training strategy was introduced during the segmentation model training, which has effectively suppressed the "overfitting" phenomenon during landslide body inference. The proposed rapid 3D terrain reconstruction algorithm was validated through two landslide case studies. The results show that: (1) In small-sample datasets, the progressive training strategy of "deep freezing-thawing-global fine-tuning" significantly improved the model's mAP 50-95 box and mask metrics compared to the baseline model. (2) The use of single-temporal centimeter-resolution aerial data enabled sub-pixel localization of landslide boundaries (error < 0.5 GSD). The image segmentation speeds for the two landslides reached 241.4 ms and 256.7 ms, with relative errors of 0.29% and 0.31% for projected area inversion, and 1.81% and 3.36% for volume inversion, respectively. This study deeply integrates machine vision deep learning with geology and surveying disciplines. It has established a mapping relationship from remote sensing image features to landslide physical parameters, which provides a new technical pathway for emergency geological disaster response.