1. River Ice Classification from High-Angle Oblique Imagery Using Deep Neural Networks.
- Author
-
Hamill, D., Giovando, J., Engel, C., Rocks, J., and Daly, S.
- Abstract
This study investigated the use of deep convolutional neural networks (DCNN) for monitoring river ice imagery generated by remote cameras on the Pend Oreille River, Idaho. The cameras were installed at selected locations along the river to provide information on ice conditions to aid in the wintertime operation of the Albeni Falls Dam. Manually reviewing the imagery for the presence of ice is difficult for large volumes of imagery. The primary advantage of DCNNs compared to other machine learning algorithms is that features in the image can be directly inferred from the image data without any subsequent data transformations or derivative products. High-angle oblique imagery collected periodically (e.g. hourly) on the Pend Oreille River was used to repurpose a highly-efficient, pre-existing DCNN. A transfer learning methodology was employed to retrain and test the DCNN using ground training tiles derived from a weakly supervised conditional random field. The tiles represented river ice, snow, terrain, water, and vegetation. The model converged with an average testing accuracy of 95%, an average f1 score of 94%, and an ice classification accuracy of 82%. The operational effectiveness of the model was assessed by applying it to several years of winter imagery which contained a large spectrum of lighting and weather conditions. The resulting time series of predicted conditions was validated against manually sorted images. The DCNN classifier accurately predicted the daily presence of river ice with an f1 score of 76% when compared to the manual classification. The model presented in the study can aid winter operation engineering decisions of water managers by providing automated alerts of river ice conditions. [ABSTRACT FROM AUTHOR]
- Published
- 2019