Deep convolutional neural network with random field model for lake ice mapping from Sentinel-1 imagery
- 14 November 2021
- journal article
- research article
- Published by Taylor & Francis in International Journal of Remote Sensing
- Vol. 42 (24) , 9351-9375
Abstract
Timely information on lake ice cover conditions is critical in support of commercial shipping, winter-road transportation, and winter leisure activities such as snowmobiling and ice fishing. Monitoring of lake ice extent and ice phenology (i.e. dates associated with freeze-up/break-up and ice duration) is also valuable for improving numerical weather prediction (NWP) and for climate monitoring. The availability of free synthetic aperture radar (SAR) data from the European Space Agency’s Sentinel-1 A/B constellation provides an unprecedented opportunity to develop operational algorithms for lake ice cover mapping at a temporal frequency not available until now from SAR missions (ca. every 1–5 days depending on latitude). For NWP, rapid and accurate mapping of ice cover and open water areas in lake-rich regions is required. However, the classification of SAR imagery using traditional machine learning (ML) approaches is challenging due to the large data volume generated by imaging systems such as Sentinel-1 SAR as well as the complexity of radar signatures as a function of sensor and target characteristics. For lakes specifically, radar backscatter can vary greatly with incidence angle, polarization and surface properties (e.g. calm and wind-roughed open water, new thin ice, surface roughness due to the presence of pressure ridges, ice type, melting of ice and on-ice snow). To address the challenge of large-scale lake ice mapping, we investigate the GPU-boosted deep neural network approach that is efficient at handling big complex data. In this paper, we design a novel maximum a posterior (MAP) approach combining a convolutional neural network (CNN) and conditional random field (CRF) for better addressing the challenges of operational lake ice mapping. The proposed approach is tested on 17 Sentinel-1 dual-polarization (VV and VH) SAR images, where eight are used for training and validation and nine used for out-of-sample testing. To identify the optimal network architecture, an independent validation set is used to evaluate the performance of a series of six CNNs with increasing model complexity. The best model overcomes the noise effect in Sentinel-1 SAR imagery and the lake ice signature ambiguity issue; it achieves average classification accuracies of 97.10% and 97.14% for open water and ice, respectively, on the validation set. Moreover, the best model outperforms the last-ranked model by about 2% in terms of mean overall accuracy (OA), demonstrating the improvement and importance of model selection. The use of CRF can consistently improve the CNN by reducing the artefacts and noise effect in ice maps, outperforming the CNN model when used alone by about 3% in terms of mean OA on the validation set. The proposed CNN-CRF approach also achieves high accuracy on the nine test scenes, achieving a mean OA of 93.10%, demonstrating strong generalization capability that is important for SAR lake ice mapping.Keywords
This publication has 43 references indexed in Scilit:
- Sentinels for science: Potential of Sentinel-1, -2, and -3 missions for scientific observations of ocean, cryosphere, and landRemote Sensing of Environment, 2012
- GMES Sentinel-1 missionRemote Sensing of Environment, 2012
- The response and role of ice cover in lake-climate interactionsProgress in Physical Geography: Earth and Environment, 2010
- River-ice break-up/freeze-up: a review of climatic drivers, historical trends and future predictionsAnnals of Glaciology, 2007
- Indicators of Climate Warming in Minnesota: Lake ICE Covers and Snowmelt RunoffClimatic Change, 2006
- Recent trends in Canadian lake ice coverHydrological Processes, 2006
- Fast approximate energy minimization via graph cutsIEEE Transactions on Pattern Analysis and Machine Intelligence, 2001
- Comparison and fusion of co‐occurrence, Gabor and MRF texture features for classification of SAR sea‐ice imageryAtmosphere-Ocean, 2001
- Texture analysis of SAR sea ice imagery using gray level co-occurrence matricesIEEE Transactions on Geoscience and Remote Sensing, 1999
- A finite mixtures algorithm for finding proportions in SAR imagesIEEE Transactions on Image Processing, 1995