Opportunities and Challenges using Observations and Simulated Data for Machine-Learning-Based Retrievals of Water Quality
Abstract
Due to the large data volume of hyperspectral observations, the computational cost of traditional retrieval methods, such as optimal estimation, can be too high for many applications. Faster retrieval methods are being evaluated for operational use, many of which use machine learning (ML) for prediction and uncertainty quantification. Optimization and validation of traditional techniques leverage observed reflectances from remote sensing instruments and spatiotemporally collocated in-situ measurements. This approach has proven successful, however the number of samples in these matchup datasets is insufficient for training, validation, and testing more complex ML models, such as neural networks.
In this work, we assess the feasibility of developing ML based retrievals using simulated Hyperspectral Imager for the Coastal Ocean (HICO) reflectance data and associated water quality parameters. In order to demonstrate operational readiness, the model performance is validated with both coincident reflectance/in situ matchups and simulated data. Here, we present the benefits and lessons learned when transitioning from training on simulated data to predicting on observed reflectances.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMGC41A..02L