Integrating camera imagery, crowdsourcing, and deep learning to improve high-frequency automated monitoring of snow at continental-to-global scales

Snow is important for local to global climate and surface hydrology, but spatial and temporal heterogeneity in the extent of snow cover make accurate, fine-scale mapping and monitoring of snow an enormous challenge. We took 184,453 daily near-surface images acquired by 133 automated cameras and processed them using crowdsourcing and deep learning to determine whether snow was present or absent in each image. We found that the crowd-sourced data had an accuracy of 99.1% when compared with expert evaluation of the same imagery. We then used the image classification to train a deep convolutional neural network via transfer learning, with accuracies of 92% to 98%, depending on the image set and training method. The majority of neural network errors were due to snow that was present not being detected. We used the results of the neural networks to validate the presence or absence of snow inferred from the MODIS satellite sensor and obtained similar results to those from other validation studies. This method of using automated sensors, crowdsourcing, and deep learning in combination produced an accurate high temporal dataset of snow presence across a continent. It holds broad potential for real-time large-scale acquisition and processing of ecological and environmental data in support of monitoring, management, and research objectives.