Fig. 6: Model interpretability with Grad-CAM results (2016–2019).
From: A human-machine collaborative approach measures economic development using satellite imagery
(A) Grad-CAM heatmap visualization over a grid image of the Ryongyon region from 2016 to 2019, indicating the reclaimed land as being a critical factor in the change of economic development scores. The pixels that contain dramatic gradient change appear brighter in the visualization. (B) Grad-CAM heatmap over three areas of interest: Samjiyon development project, Kalma tourist project, and Nyongbyun nuclear scientific research center. All three sites show growth in economic development and the heatmap highlights the key areas contributed to the score change. The model can detect subtle changes in roofing and road network in the Nyongbyun nuclear site that would be otherwise hard to detect with human eyes. Satellite images in the figure contain modified Copernicus Sentinel data [2016, 2017, 2018, 2019].
