Efficient edge-assisted remote sensing video analysis with gradient-weighted class activation mapping for adaptive region-of-interest extraction

To address the high bandwidth and computational demands of object detection in remote sensing videos, we present EVA-GARE, an edge-assisted video analysis framework that coordinates resource-constrained acquisition devices with ground-edge servers. The device side uses a Grad-class activation map-based adaptive region-of-interest (RoI) extraction mechanism for preliminary processing, while an adaptive quality controller and RoI integrator reduce the amount of data transmitted to the edge server. The edge server performs detection on the received data and provides network-state feedback for subsequent processing. Experiments using a device-side YOLOv5s and an edge-side YOLOv7-E6 were conducted on the evaluated remote-sensing datasets and video sequences in a controlled laboratory testbed. In these experiments, EVA-GARE reduced the transmitted data volume by ~80%, while the peak F1 score observed in the evaluated video sequences reached 0.87.