Comparison of machine learning algorithms based on Sentinel-1A data to detect icebergs

• Main Authors: XIAO Xiangwen, SHEN Xiaoyi, KE Changqing, ZHOU Xinghua
• Format: Article in Journal/Newspaper
• Language: Chinese
• Published: Surveying and Mapping Press 2020
• Subjects: iceberg; machine learning; sentinel-1a; sar; Mathematical geography. Cartography; GA1-1776; Arctic; Greenland; Iceberg*
• Online Access: https://doi.org/10.11947/j.AGCS.2020.20190174; https://doaj.org/article/3641b8957a014819a6859fb07856b9e3

Iceberg detection is of great significance for marine environmental monitoring and safe sailing of vessels. It is an important part of the construction of the Arctic channel and the exploitation of the Arctic. Iceberg detection using synthetic aperture radar (SAR) images has unique advantages. Many machine learning algorithms can be used in the recognition of icebergs in SAR images. In order to maximize the performance of machine learning algorithms, it is necessary to evaluate different machine learning algorithms and their matching feature and feature standardization methods, so as to select the optimal iceberg detection process method. Therefore, based on Sentinel-1A SAR image, this paper uses a variety of machine learning methods, a variety of feature combinations and a variety of feature standardization methods for iceberg detection, and compares the performance differences of each process method. Machine learning algorithms include Bayes classifier (Bayes), back propagation neural network (BPNN), linear discriminant analysis (LDA), random forest (RF) and support vector machine (SVM); feature standardization methods include Min-max standardization, Z-score standardization and log function standardization; data sets are comprised of 969 iceberg and non-iceberg samples with 12 SAR image features, located mainly on the east coast of Greenland. The classification result is measured by the area under the receiver operating characteristic (ROC) curve (AUC). The results show that the AUC value of RF with the best configuration is the highest, reaching 0.945, which is 0.09 higher than worst Bayes. In terms of detection rate, under the case of 80% iceberg recall rate, the non-iceberg recall rate of RF is 92.6%, which is the best, 1.4% higher than the second BPNN, 2.6% higher than the worst Bayes; under the case of 90% iceberg recall rate, the non-iceberg recall rate of BPNN is 87.4%, 0.8% higher than the second RF and 2.7% higher than the worst Bayes. The above results show that it is very important to select the ...