Satellite-Based Landslide Event Detection Leveraging EfficientDet Model

Timely detection and monitoring of landslides are essential for effective disaster preparedness and mitigation. With the notable advancement of remote sensing technology and the computational environment, enriched solutions are anticipated for landslide hazards. Hence this study evaluates a deep learning-driven approach for detecting landslides from satellite images. Particularly, EfficientDet, a popular single-staged object detection algorithm. A series of EfficientDet models namely D0, D1, D2, D3, and D4 are applied and implemented. For experimentation, publicly available Bijie landslide detection data is used. Furthermore, the model's performance is assessed using well-known indicators i.e., average precision and recall. The experimental results indicated that the accuracy of EfficientDet_D3 is best (AP@0.50=0.80) followed by EfficientDet_D4, EfficientDe_D1, EfficientDet_D2, EfficientDet_DO. The qualitative and the quantitative outcomes exhibit the power of deep learning algorithms in the studies of hazard analysis. Our findings contribute to proactive disaster management schemes, aiding in the growth of enhanced mitigation measures to minimize the effect of landslides on communities and the environment.