Deep Learning of River Width Changes from Ultra-High-Resolution Commercial Satellite Imagery
Abstract
Many of the worlds regions most sensitive to changes in river discharge due to climate change are geographically remote (e.g., the Arctic) and/or economically under-resourced (e.g., in much of the semi-arid zones). As a result, field data of these impacts are sparse. Remote sensing from satellite (and drone) imagery offers a scalable solution to vastly increase such data availability. Indeed, several satellites have been used to map local and global water distributions over time. Until recently, the resolution of publicly available satellite data (e.g., 30 m for Landsat) only allowed the tracking of relatively large changes in large bodies of water. Moreover, early approaches classified water based on pixel-level criteria that are satellite specific, require tuning, and post-processing of misclassified non-water objects (such as man-made structures). In this work, we use commercial satellite imagery with panchromatic and multispectral resolutions of ~ 50 cm to 2 m. Multiple fully convolutional neural networks are used to identify only rivers and their changes in width over time. These networks can not only learn, e.g., entropies and highly non-linear relationships between spectral bands (in multispectral images), but also the morphologies of objects such as rivers. As a result, they offer a more robust and universal classifier. Our trained models for multispectral images (using only RGB and NIR bands) show excellent performance without any parameter tuning (F1-scores of ~92%), which allows us to use the predictions from this model to construct a larger training set for overlapping panchromatic images. Trained models for panchromatic-only images are unsurprisingly less accurate but together with standard post-processing techniques can still provide adequate river detection, especially given their much larger spatial and temporal resolutions. Our results demonstrate the power of a deep learning framework that can be readily applied to many applications of interest to the hydrology community.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H42G..04M