Sensing lakes and streams using machine learning
Catchment satellite imagery - Sentinel 2
The catchment is important
The topographic catchment is the area delivering surface water to a lake or stream. This area can be delineated using digital elevation models. One of my earlier projects project showed that stream water CO2 could be predicted from catchment characteristics. This analysis relied on several features including average elevation and slope and land cover distribution. The use of machine learning algorithms improved the predictive accuracy. More recently, I have shown that large collections of geospatial features at the catchment, buffer, and lake level can be used to predict water quality in lakes as well paper. Therefore, adding further data sources and improved modeling techniques should improvement predictions in both lake and stream ecosystems.
Leveraging neural networks and satellite imagery
So far, the predictive models have relied on traditional ‘tabular’ data that could be viewed in a normal spreadsheet. However, using the flexibility of neural networks this kind of data can be combined with image data in a single model. Image data may capture the notion of land cover and landscape configuration in ways that cannot be reduced to single numbers. Freely available satellite data such as that from the Sentinel 2 satellite provide imagery at high resolution (10 m). My earlier projects described above have shown that topography is important, suggesting that the addition of digital elevation models is beneficial. Fusion of these data sources could improve our ability to predict general stream and lake chemistry by using modern machine learning techniques. If successful, the ability to predict surface water chemistry for new sites, lower the burden associated with manual sampling and improve management of aquatic habitats.
