UAV-Based Snow Cover Detection Using Deep Learning and Geospatial Visualization
Abstract
Snow disaster management represents a critical domain that requires accurate, real-time monitoring of snow-covered areas. In this regard, the detection of snow-covered surfaces from Unmanned Aerial Vehicles (UAVs) constitutes a key component of this task. Deep learning has demonstrated strong performance in aerial image classification; however, the deployment of these models on drone hardware remains constrained by several factors, including Size, Weight, and Power (SWaP) limitations. This paper proposes a robust end-to-end framework for UAV-based snow cover detection using geotagged drone imagery acquired over the Jbel Saghro mountain range in Morocco. The study compares eight EfficientNet architectures (B0 through B7) and MobileNetV2 against a Support Vector Machine (SVM) baseline. The models were evaluated using standard classification metrics and hardware-aware indicators, in particular, parameter count (in millions), computational complexity (GFLOPs), inference latency (milliseconds), and processing speed (frames per second, FPS), to simulate real-world UAV edge-computing constraints. The optimal model was subsequently embedded into a geospatial visualization pipeline using ArcGIS 10.8.2 within Power BI to generate interactive heatmaps that superimpose the spatial density of snow-covered observations onto road networks and topographic features. Experimental results on a dataset of 9307 annotated UAV images demonstrate that EfficientNet-B0 achieves an accuracy of 99.21% and an AUC of 0.99 while processing images at over 100 FPS, making it the optimal architecture for real-time onboard drone processing. This geospatial framework provides actionable insights for disaster management, road maintenance operations, and early warning systems in mountainous regions.