Landslide Extraction Using Fused Local and Nonlocal Attentional Features on Edge Device Toward Embedded UAV Emergency Response

Unmanned aerial vehicles (UAVs) have made significant contributions to landslide emergency response operations due to their precise and flexible imaging capabilities. However, the conventional workflow of generating orthophotos from UAV imagery and subsequent interpretation often exceeds the critical 72-h rescue window. To address this challenge, this article presents an onboard landslide extraction method for original UAV images utilizing a convolutional neural network (CNN). Given the abundance of overlapping images, the CNN is trained on labeled orthophotos. To minimize discrepancies between orthophotos and original UAV images, the proposed method integrates local and nonlocal features. Built upon the ResNet architecture, the method incorporates modules for extracting both shallow and deep features, enabling effective fusion through self-learning. This approach mitigates the issue of accuracy degradation caused by variations between training and testing data. Furthermore, considering the necessity of deploying the CNN-based landslide extraction model on low-power embedded platforms to achieve onboard landslide extraction, this article introduces a quantitative model compression technique. Specifically, the model’s weight and activation value data precision are linearly mapped from 32-bit floating-point type to 8-bit integer type, guided by relative entropy minimization. This results in substantial reductions in memory access and computational complexity during model inference. Experimental results demonstrate that the proposed method yields outstanding extraction performance on both the Jiuzhaigou and Bijie landslide datasets. The time taken for extracting landslides from a single
$6000 \times 4000$
pixel UAV image is reduced from 109.47 to 4.75 s, which is less than the 5.13-s interval between camera shots, thereby achieving onboard landslide extraction.