Multiclass Classification of Global Water Bodies in ReaLSAT

Keywords: Deep Learning, Earth Sciences, Spatio-Temporal Data Challenges
Abstract Type: Guided Poster Abstract

Abstract

ReaLSAT is a global dataset that contains the location and surface area variations of 681,137 bodies of water from 1984 to 2015. While the dataset does not include a separate label for each body of water, classifying them into their respective types has several benefits. First, it helps the study of the impact of human actions and climate change on freshwater availability. Next, information presented as discrete bodies of water is more useful for policy makers as this will help answer causal questions related to urbanization and water shortage. Lastly, morphological data can help to evaluate the flows of sediment and water that influence aquatic and riverside habitats. The high dimensional and spatial-temporal nature of remotely sensed bodies of water makes this classification problem extremely challenging.
The objective of this poster is to do multiclass classification using transfer learning on a pre-trained convolutional neural network on a small labelled portion of the dataset. Motivated by the large size and primarily unlabeled nature of the dataset an analysis is also done on finding hidden structure in the labelled dataset using unsupervised clustering. This poster presents the preliminary data analysis of the ReaLSAT dataset and discuses the results and challenges of both of these methods.