Knowledge distillation based lightweight building damage assessment using satellite imagery of natural disasters

Accurate and timely assessment of post-disaster building damage is of great significance for national development and social security concerns. However, due to the high timeliness requirements of disaster emergency response and the conflict that sufficient computing resources are not easily available in harsh environments, and therefore the lightweight AI-driven post-disaster building damage assessment model is highly needed. In this paper, we introduced a knowledge distillation-based lightweight approach for assessing building damage from xBD high-resolution satellite images with the purpose of reducing the dependence on computing resources in disaster emergency response scenarios. Specifically, an ensemble Teacher-Student knowledge distillation method was designed and compared with the xBD baseline model. The result has shown that, the knowledge distillation reduces the parameter number of the original model by 30%, and the inference speed is increased by 30%-40%. In the building localization task, the accuracy of teacher and student model are 0.879 and 0.832 (IOU) respectively. In the damage classification task, the accuracy of teacher and student are 0.798 and 0.775 respectively. In addition, we proposed a dual-teacher-student knowledge distillation strategy, which cannot use the pre-training skills of curriculum learning in student model training, but achieve the same effect through more direct knowledge transfer. In the experiment, our dual-teacher-student method improves the knowledge distillation baseline by 3.7% with 30 epoch training. With only 70% parameters, our student model performs close to the teacher model at a degradation within 5%.This study verifies the effectiveness and prospect of knowledge distillation method in building damage assessment for disaster emergency.