3  Machine Learning Methods for Sea Ice Mapping

A review of machine learning methods for sea ice mapping from traditional machine learning to deep neural networks.

3.1 A list of models used for sea ice mapping

3.1.1 Traditional Machine Learning

gray-level co-occurrence matrix (GLCM)

Gabor Filters

3.1.2 Deep Neural Networks

Samira splits methods into patch-based and pixel-based.

Pixel-based methods use encoder-decoder networks with skip connections

Patch-based methods use hierachical feature extraction methods

3.1.3 Patch-based classification

Assign a single label to one image region or patch. This is more computationally efficient.

Most have used convolutional neural networks (CNN).

Questions:

  1. Architecture
  2. Layers
  3. Patch size
  4. Number of training images
  5. Location of training images

Li et al [12] CNN from Gaofen-3 SAR images

Boulze et al [13] CNN

3D-CNN and Squeeze-Excitation (SE) -based networks (Hu et al. 2019) - SE-based networks focus on interdependencies between channels (rather than/as well as) spatial interdependencies and structures.

Han et al. (2019) computed the gray-level co-occurrence matrix (GLCM). Use K-nearest-neighbour clustering to identify unlabeled neighbouring that are used to increase/boost/ehance label samples. Spectral and spatial dimensions are reduced through a correlation based dimension reduction algorithm.