Assessment of thermal noise impact on sea ice classification using Sentinel-1 images and U-Net

Sea ice poses a significant threat to high-latitude navigation and offshore operations. Accurate and timely classification of sea ice is crucial for ensuring the safety of maritime activities in polar regions. Synthetic aperture radar (SAR) is widely used for sea ice classification due to its high resolution and all-weather observation capability. However, the Sentinel-1 extra-wide (EW) swath mode images, which are commonly used to monitor sea ice in the polar region, exhibit thermal noise in the cross-polarization images, and it is thought to affect the accuracy of sea ice classification models. In this study, we used Sentinel-1 EW mode images and a deep learning (DL) model, U-Net, to investigate the impact of thermal noise on sea ice classification. Sensitivity experiments were conducted for the U-Net and the comparison models, such as support vector machine (SVM), random forest (RF), and convolutional neural network (CNN), with or without using a denoising method for cross-polarization images. Both co-polarization and cross-polarization images were used to train these models. The experimental results indicate that SVM, RF, CNN, and U-Net achieved classification accuracies of 67.98%, 77.96%, 86.49%, and 90.00% respectively, using undenoised images. The classification accuracies improved to 71.69%, 80.75%, 86.65%, and 90.73% respectively after the denoising method was used. The SVM and RF models show an increase in accuracy of about 3%, while the CNN and U-Net models show an improvement of less than 1%, suggesting that CNN and U-Net are more tolerant to noise when used for sea ice classification.