Leveraging Remote Sensing and Deep Learning to Estimate Suspended Sediment Concentration
Abstract
Continuous measurement and monitoring of suspended sediment concentration (SSC) in rivers are essential to study the impacts of climate and anthropogenic influences on erosion and sediment transport processes. We use Landsat satellite imagery to develop a SSC model in Google Earth Engine (GEE) that allows users to produce heatmaps of SSC for the Snake and Columbia Rivers in the Northwest US. We conducted a laboratory experiment to determine relationships between multispectral image bands and SSC, and then used satellite imagery to estimate SSC at large scales.
We implemented a laboratory experiment with a multispectral sensor to test the reflectance signature of varying levels of SSC at two water temperatures (19 °C and 24°C). We implemented image segmentation architectures of U-Net and adopted transfer learning using Efficient-Net to train an SSC model. We report the best results obtained from fine-tuned Efficient-Net b3 with an average coefficient of determination (R2) of 0.96 for 10-fold cross-validation and an R2 score of 0.95 in unseen test data. We found that measurements of water temperature are important in model accuracy, with implications for real-world models in a warming climate. Further, we used the AquaSat data that reports SSC in US rivers and associated Landsat recorded imagery (mean and standard deviation values of bands) to train a SSC model for the Northwest US rivers. We tested different models, including Random Forest, XGBoost, and different architectures of deep learning and Encoder-Decoder. We report an R2 of 0.71 in unseen test data from our deep learning model (Encoder-Decoder). We found that reflected light from NIR, blue, and red bands, respectively, were most important for predicting SSC. Also in both laboratory and using AquaSat data, we found that model uncertainty increases in higher values of SSC. Ultimately, our model will support a GEE app that will provide spatiotemporally continuous SSC maps for rivers in the US Northwest from 1984-present. This app can assist in aquatic habitat preservation and restoration, water transportation, and water resources planning which all depend on high-resolution and high-quality sediment distribution information.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMIN12B0263M