Exploration of artificial intelligence approaches to Earth observing remote sensing

Abstract

Freshwater ecosystems such as lakes and reservoirs are important natural resources providing services such as drinking water, hydropower, fisheries, and support to regional economies. Traditionally, water quantity and quality changes have been identified from manual sampling programs and, more recently, autonomous sensor platforms. However, drivers of change frequently originate from the surrounding watershed. Optical remote sensing provides an opportunity to characterize water quality and the watershed simultaneously. However, from a data-driven perspective, the key landscape features, their inter-relations, and appropriate satellite data products are poorly constrained. The potential of artificial intelligence (AI)-derived insights gained from multi-channel satellite sensors are enormous. We present an unsupervised data-driven workflow to detect and describe landscape features in remote sensed multispectral images of lake watersheds in the continental United States. Our approach employs computer vision and machine learning techniques to efficiently segment size-varying multiband images and learn the spectra of the optimal number of segment classes for a dataset. These classes are used to generate a reduced dimensionality representation for a given watershed image encapsulating the most distinct features of that image. Downstream applications include land cover class detection, time-series analysis of environmental features and remotely sensing of water quality indicators.