Toward Polar Sea-Ice Classification using Color-based Segmentation and Auto-labeling of Sentinel-2 Imagery to Train an Efficient Deep Learning Model

TL;DR

This paper presents a color-based segmentation and auto-labeling approach for Sentinel-2 satellite images to train a deep learning model for classifying polar sea ice, achieving high accuracy with minimal manual labeling.

Contribution

The study introduces an automatic labeling method for satellite imagery and demonstrates its effectiveness in training a U-Net model for sea ice classification.

Findings

Auto-labeled data achieves 90.18% accuracy in classification.

Filtering clouds and shadows boosts accuracy to over 98%.

Auto-labeled training yields comparable results to manual labeling.

Abstract

Global warming is an urgent issue that is generating catastrophic environmental changes, such as the melting of sea ice and glaciers, particularly in the polar regions. The melting pattern and retreat of polar sea ice cover is an essential indicator of global warming. The Sentinel-2 satellite (S2) captures high-resolution optical imagery over the polar regions. This research aims at developing a robust and effective system for classifying polar sea ice as thick or snow-covered, young or thin, or open water using S2 images. A key challenge is the lack of labeled S2 training data to serve as the ground truth. We demonstrate a method with high precision to segment and automatically label the S2 images based on suitably determined color thresholds and employ these auto-labeled data to train a U-Net machine model (a fully convolutional neural network), yielding good classification accuracy. Evaluation results over S2 data from the polar summer season in the Ross Sea region of the Antarctic show that the U-Net model trained on auto-labeled data has an accuracy of 90.18% over the original S2 images, whereas the U-Net model trained on manually labeled data has an accuracy of 91.39%. Filtering out the thin clouds and shadows from the S2 images further improves U-Net's accuracy, respectively, to 98.97% for auto-labeled and 98.40% for manually labeled training datasets.