GLACIERNET: A NOVEL CONVOLUTIONAL NEURAL NETWORK APPLICATION FOR DEBRIS-COVERED GLACIER MAPPING

The global temperature has been continuously increasing over the past decades. The effect of temperature increase can directly affect the health, dynamics, and processes of alpine glaciers. To understand these changes on hundreds of glaciers, one of the basic steps is to map glaciers either using the conventional method of field mapping or digital mapping using remote sensing techniques. However, the glacier mapping is not straightforward using either of these techniques, and it becomes much more difficult when it comes to mapping debris-covered glaciers. Either manual on-screen digitization using satellite data or detailed post-processing following semi-automated mapping is considered to be the best possible approach. Consequently, in this research, the convolutional neural network (CNN), which is a deep learning, feed-forward neural network, is applied to the Landsat era satellite images to automatically map debris-covered glaciers. The Landsat data of the network that are appropriately selected by considering the period, region, resolution, cloud cover, and bands, are pre-processed to fit the requirements of the CNN input. Based on the properties of input satellite data, CNN segmentation process, and empirical results, we designed a new CNN structure named as GlacierNet by appropriately choosing the type, number and size of layers, and encoder depth. The GlacierNet is trained by learning pre-processed satellite data and corresponding glacier boundaries that are available at the Global Land Ice Measurements from Space (GLIMS) database. During the GlacierNet mapping, the network has two main processes such as encoding and decoding. The encoding process extracts the features from input image data and thereby many small feature maps are generated. The decoding process fuses all the small feature maps that are from the encoding process to obtain size restored binary mapping results. After that, the mapping results are processed by physical and geological techniques to further improve the mapping accuracy. Our preliminary results indicate high accuracy in glacier mapping, a major step in developing a fully automated methodology for glacier mapping.