Snowfall intensity estimation based on convolutional neural networks using CCTV data

초록

"This study proposes a model for estimating snowfall intensity using Convolutional Neural Networks (CNNs) based on CCTV data. CCTV data from the Cloud Physics Observatory Station were synchronized with PARSIVEL measurements, defining an image-value framework where CCTV data served as inputs and PARSIVEL snowfall intensity as outputs. A k-Nearest Neighbor (kNN) algorithm was employed to extract snowfall particle information by calculating pixel differences between frames and dynamically updating foreground and background models. The data were cropped to a 640 × 640 pixel region of interest (ROI), excluding non-snowfall regions to enhance training efficiency. The CNNs-based image-value model was trained on 80% of 95,374 data points, with the remaining 20% reserved for testing. Performance evaluation involved random sampling of 500 test samples and analysis using five performance metrics. Results showed that the snowfall intensity estimated by image-value model closely matched PARSIVEL observation data, achieving the highest performance rating in two out of five metrics. This study demonstrates the effectiveness of integrating CNNs with generative models to enhance the accuracy of snowfall intensity estimation. Furthermore, it emphasizes that the incorporation of particle distribution data can facilitate more sophisticated and precise snowfall characterization. © 2025 Korea Water Resources Association.

키워드