Using Land Cover Type of Upstream Area to Predict Pollution in the Chesapeake Bay Watershed
We wanted to put our combined skill set to use for an environmentally important goal. Since we all live in and enjoy the Chesapeake Bay watershed, this hackathon had particular meaning for us. Given our background in machine learning, statistics, and GIS, we thought we would try the Challenge 3, the machine learning option. It was important to us that the output could go beyond simply being a model to being a tool for scientific communication with our community. That is why we built an online interface for users to see how the actions within the watershed may affect pollution and improve water quality.
We iteratively created a workflow to prepare data for analysis, created and trained a machine learning model to predict nitrogen pollution at sampling sites based on land cover inputs, and built an online interface for public interaction.
We focused on nitrogen pollution because it is one of the main drivers of ecological impairment in the Chesapeake Bay. Nitrogen pollution is also well predicted by land cover, for which there is a recent USGS provided dataset on which we can build a model. There are also well known and established mitigation measures for nitrogen pollution, such as stream buffers and wetland catchments, which are easy for the public to understand and begin to implement to improve water quality.
In our data exploration we found a seasonal effect on nitrogen levels, with lower levels in the summer (Figure 1 and Figure 2). We also noticed that more samples, particularly in the bay itself, were taken in the summer.
Figure 1: The average nitrogen by station, broken out by month.
Figure 2: A boxplot of the nitrogen levels broken down by month across all locations. The means clearly change seasonally.
First, we extracted land cover within watersheds. We used the public National Land Cover Dataset (NLCD) 2016 Science product created by USGS to provide land cover data at a 30 meter resolution. We used the USGS watershed basin boundaries (which are named using hydrologic unit codes, called HUCs) as boundaries to distinguish which land cover drained into which waterways. The number of pixels of each land cover type and their percentage within the HUC boundary were compared to the water quality sampling data taken at stations within those boundaries using various statistical correlation tests and then input into a machine learning model (gradient boosting regression). We tested HUC sizes 8, 10, and 12, with 8 being larger regions and 12 being the smallest. HUC sizes are compared in Figure 3.
Figure 3: A comparison of hydrological unit code (HUC) sizes. HUC-8 is the largest, followed by HUC-10, with HUC-12 being the finest-grained.
At first, even using the smallest land cover grouping at the HUC 12 level of boundary, we did not find as strong a relationship between land cover and station sampling data as we expected, given the causal relationship between them recorded in peer-reviewed literature. We considered this may be because the water station samples were influenced by upstream water basins that we had not included.
To address this concern, we additionally extracted land cover in HUCs upstream of each sampling station. This was done by using the National Hydrography Dataset to create an iterative tool to identify which waterways flowed downstream to each sampling station. We used the selected set of streams (example in Figure 4) for each sampling station to identify which HUCs the streams crossed through on their way to the sampling station. We then used these grouped upstream HUCS to examine the land cover type amounts and percentages for the waterways flowing into the sampling stations.
Figure 4: All the streams feeding into a particular station with corresponding HUC12 regions.
Data analysis included Pearson correlations (see Figure 5) to decide which major groups should be included in the interface. Various machine learning models were tested, from which we selected gradient boosting regression was selected. The R^2 for these models (see Figure 6) showed that when all extractions were considered, the best fit was achieved. However, the two best fit models included the upstream tracing land cover extractions. The second best fit was only slightly lower, and achieved using only the upstream tracing that grouped influential HUC 12 watersheds. R^2 for models which explicitly look at comparing the usefulness of different spatial granularities for landcover also support that the upstream HUC data is most useful and that HUC8 to too large an area to be indicative of water quality (see Figure 7). This supported our hypothesis and general literature consensus that upstream land cover was as influential as the land cover immediately surrounding the sampling station. The feature importance graph for the model which only used the upstream landcover features (see Figure 8) indicates that cultivated crops, open water, and wetlands all play an important role in predicting nitrogen levels.
Figure 5: The correlation between land use types, and between land use types and nitrogen.
Figure 6: The R^2 values for different models. We see that adding more geographic region sizes improves performance. More interestingly, the second best model is the basci Latitude, Longitude, Month, and Year variables win the upstream land cover information.
Figure 7: The R^2 values for the different models. We can see from this that the basic variables (LLAMY) plus land use and upstream land use datasets performed the best.
Figure 8: Feature importance graph for gradient boosting regression with upstream landcoverage as input features.
One challenge was subsetting the land cover upstream from the stations to reflect contributing runoff sources as closely as possible. First, the hydrography data needed to be downloaded with an intact geometric system, meaning the lines representing streamflow had to be collected logically to indicate directionality. Then we had to learn to manipulate this data using the associated tools. The large number of stations meant this process must be automated. Models were created to batch process our stream selection and to extract land cover for those streams.
Further challenges arose in finding the most appropriate statistical analysis. We needed to measure complex correlations between a large number of land cover types and thousands of samples. This took a great deal of literature review and expertise to select and test appropriate models.
The data manipulation, figure creation, and modeling code lives in the model directory, and the data transformation scripts live in the eda directory.
The react and flask demo lives in the demo directory.
It requires running both the flask server (cd demo/server ; python app.py) and the React front-end (cd demo/client ; npm start) once you have the requirements installed.
