Classification of Sea Ice Types in Sentinel-1 SAR Data Using Convolutional Neural Networks

A new algorithm for classification of sea ice types on Sentinel-1 Synthetic Aperture Radar (SAR) data using a convolutional neural network (CNN) is presented. The CNN is trained on reference ice charts produced by human experts and compared with an existing machine learning algorithm based on texture features and random forest classifier. The CNN is trained on two datasets in 2018 and 2020 for retrieval of four classes: ice free, young ice, first-year ice and old ice. The accuracy of our classification is 90.5% for the 2018-dataset and 91.6% for the 2020-dataset. The uncertainty is a bit higher for young ice (85%/76% accuracy in 2018/2020) and first-year ice (86%/84% accuracy in 2018/2020). Our algorithm outperforms the existing random forest product for each ice type. It has also proved to be more efficient in computing time and less sensitive to the noise in SAR data. The code is publicly available.

Download Full-text

Convolutional Neural Networks for Classification of Sea Ice Types in Sentinel-1 SAR Data

10.5194/egusphere-egu21-10580 ◽

2021 ◽

Author(s):

Anton Korosov ◽

Hugo Boulze ◽

Julien Brajard

Keyword(s):

Random Forest ◽

Sea Ice ◽

Classification Accuracy ◽

Thermal Noise ◽

Noise Removal ◽

The Arctic ◽

First Year ◽

Sar Data ◽

The Impact

A new algorithm for classification of sea ice types on Sentinel-1 Synthetic Aperture Radar (SAR) data using a convolutional neural network (CNN) is presented.&#160; The CNN is trained on reference ice charts produced by human experts and compared with an existing machine learning algorithm based on texture features and random forest classifier. The CNN is trained on a dataset from winter 2020 for retrieval of four classes: ice free, young ice, first-year ice and old ice. The accuracy of our classification is 91.6%. The error is a bit higher for young ice (76%) and first-year ice (84%). Our algorithm outperforms the existing random forest product for each ice type. It has also proved to be more efficient in computing time and less sensitive to the noise in SAR data.&#160;Our study demonstrates that CNN can be successfully applied for classification of sea ice types in SAR data. The algorithm is applied in small sub-images extracted from a SAR image after preprocessing including thermal noise removal. Validation shows that the errors are mostly attributed to coarse resolution of ice charts or misclassification of training data by human experts.&#160;Several sensitivity experiments were conducted for testing the impact of CNN architecture, hyperparameters, training parameters and data preprocessing on accuracy. It was shown that a CNN with three convolutional layers, two max-pool layers and three hidden dense layers can be applied to a sub-image with size 50 x 50 pixels for achieving the best results. It was also shown that a CNN can be applied to SAR data without thermal noise removal on the preprocessing step. Understandably, the classification accuracy decreases to 89% but remains reasonable.&#160;The main advantages of the new algorithm are the ability to classify several ice types, higher classification accuracy for each ice type and higher speed of processing than in the previous studies. The relative simplicity of the algorithm (both texture analysis and classification are performed by CNN) is also a benefit. In addition to providing ice type labels, the algorithm also derives the probability of belonging to a class. Uncertainty of the method can be derived from these probabilities and used in the assimilation of ice type in numerical models.&#160; Given the high accuracy and processing speed, the CNN-based algorithm is included in the Copernicus Marine Environment Monitoring Service (CMEMS) for operational sea ice type retrieval for generating ice charts in the Arctic Ocean. It is already released as an open source software and available on Github: https://github.com/nansencenter/s1_icetype_cnn.

Download Full-text

Extraction of Sea Ice Cover by Sentinel-1 SAR Based on SVM with Unsupervised Generation of Training Data

10.20944/preprints202005.0336.v1 ◽

2020 ◽

Author(s):

Xiaoming Li ◽

Yan Sun ◽

Qiang Zhang

Keyword(s):

Machine Learning ◽

Sea Ice ◽

Learning Algorithm ◽

Texture Features ◽

Open Water ◽

Ice Cover ◽

Training Data ◽

Support Vector ◽

Training Samples

In this paper, we focus on developing a novel method to extract sea ice cover (i.e., discrimination/classification of sea ice and open water) using Sentinel-1 (S1) cross-polarization (vertical-horizontal, VH or horizontal-vertical, HV) data in extra wide (EW) swath mode based on the machine learning algorithm support vector machine (SVM). The classification basis includes the S1 radar backscatter coefficients and texture features that are calculated from S1 data using the gray level co-occurrence matrix (GLCM). Different from previous methods where appropriate samples are manually selected to train the SVM to classify sea ice and open water, we proposed a method of unsupervised generation of the training samples based on two GLCM texture features, i.e. entropy and homogeneity, that have contrasting characteristics on sea ice and open water. We eliminate the most uncertainty of selecting training samples in machine learning and achieve automatic classification of sea ice and open water by using S1 EW data. The comparison shows good agreement between the SAR-derived sea ice cover using the proposed method and a visual inspection, of which the accuracy reaches approximately 90% - 95% based on a few cases. Besides this, compared with the analyzed sea ice cover data Ice Mapping System (IMS) based on 728 S1 EW images, the accuracy of extracted sea ice cover by using S1 data is more than 80%.

Download Full-text

Sea-ice type classification from ERS-1 SAR data based on grey level and texture information

Polar Record ◽

10.1017/s0032247400013632 ◽

1995 ◽

Vol 31 (177) ◽

pp. 135-146 ◽

Cited By ~ 11

Author(s):

D.M. Smith ◽

E.C. Barrett ◽

J.C. Scott

Keyword(s):

Sea Ice ◽

Open Water ◽

Speckle Noise ◽

First Year ◽

Grey Level ◽

Texture Information ◽

Texture Parameters ◽

Sar Data ◽

Speckle Noise Reduction

AbstractThis paper describes the development of a practical algorithm for the classification of sea-ice types from ERS-1 synthetic aperture radar (SAR) data. The algorithm was based on a combination of grey level and texture information in order to overcome ambiguous grey level values of different ice types. The problem of calculating texture parameters for windows containing more than one ice type was overcome by first segmenting the image so that only pixels from the same segment were included in the calculation of the texture measure. The segmentation procedure was based on the iterative application of a speckle noise reduction filter, and was thus crucially dependent on the ability of such a filter to smooth out noise without destroying edges and fine features. In order to achieve this, a modification to the sigma filter of Lee (1983b) was developed; it out-performed the sigma filter for a model problem. Two ERS-1 SAR scenes of the marginal ice zone east of Spitsbergen in March 1992 were analysed by calculating values of grey level and range for different ice types contained within raw data extracts. Although the grey levels of some of the ice types overlapped, most of the ambiguity was removed through the additional use of range. It was also necessary to test for the wave-like appearance of open water. The classification scheme was demonstrated to identify correctly most of the grease/new ice, first-year ice, multiyear ice, rough ice, pancake ice, and open water in the two SAR scenes, although there was some misclassification of open water as first-year ice.

Download Full-text