The goal of this repository is to run a shiny app server which outputs predictions of the lake level for lakes in the Yahara Watershed in Southern Wisconsin. It is currently implemented for the main lakes in the chain: Mendota, Monona, Waubesa, and Kegonsa.

##yahara_flash The code to run the shiny app is in the folder yahara_flash. The app itself is currently hosted at https://ecotheory.shinyapps.io/yahara_flash/. The app mostly loads data and plots it.

The lake level data is retrieved from the USGS server https://waterservices.usgs.gov/nwis/iv/?format=rdb&sites= Site keys: site_keys = c("05428000", "05429000","425715089164700", "05429485")

Historical rain data comes from NASA with the R library nasapower. Coordinates: c(43.0930, -89.3727)

The precipitation forecast data uses the R library openmeteo. Same coordinates.

There are two different modeling schemes used to make forecast: General Additive Mixed Models (GAMM) and Recurrent Neural Networks (RNN).

The fitted GAMMs are stored in the variable "lakeGAMsfull.var" and the GAMM forecast is made on the shiny server.

The fitted RNNs are not stored on the shiny server (see below). The forecasts made by these models are loaded from a file stored here in this github repository at: "https://raw.githubusercontent.com/jusinowicz/yahara_flashiness/master/yahara_flash_local/lakemodel_",n,"_forecast.csv" where n corresponds to the lakes in N to S order (Mendota, Monona, Waubesa, Kegonsa).

#yahara_flash_localv2 There is a second version of the app in the folder yahara_flash_local which performs the actual updates of the forecasts. The ML model, which is currently built with both Convolutional Neural Neworks (CNN) and Long Short Term Memory (LSTM) networks, can take 10-15 minutes to run and so it does not make sense to host it on the shiny server itself. The models are run locally on another server, and then both of the LSTMs and the forecasts are updated here.

The fitted RNNs are stored in their respective folders.

The forecasts are in lakemodel_n_forecast.csv, where n has the same lake indexing as above.

#data This is where the historical and forecasted data sets get stored.

#flashines2016 This is the original code that matches the publication Usinowicz et al. 2016.

This code is essentially designed to take basic USGS .csv data with lake level and precipitation, in combination with any other user-defined covariates (e.g. impervious surface) to investigate flashiness.

This code should be viewed as a collection of useful routines, as opposed to a single coherent program. To summarize the types of routines:

1. Data processing: taking individual column vectors of the dependenet and various independent variables, removing winter months and leap days, producing lagged covariates (i.e. of precip and lake level), combining into a single object, and making a time series object.

2. Analysis of heteroskedasticity: non-stationarity of the variance is used as a first test for changing flashiness.

3. Fitting of GAM models: the role of various covariates in driving flashiness is investigated through the use of GAM models which allow flexible, non-parametric fits. Model fitting is a very interactive process, which basically requires adding and dropping covariates in a nested way while tracking changes in AIC. Thus this section should serve largely as an example of how to do this, as opposed to an automated tool.

4. Proportion of variance explained: the GAM model can be analyzed for each period of differing flashiness levels to determine what proportion of the variance is explained by each variable or set of variables. This is used to argue that flashiness is being more or less driven by any of these variables, relative to others.

5. Partial flood analysis with moving window: Standard USGS methods exist for calculating the probability that the stage level of any waterway will exceed certain thresholds in a given period of time. Given that the variance in lake-level is non-stationary in Mendota, we adapt the standard approach with a moving-window that simply applies the standard approaches to a limited 10-window. In this way, we can track how the probability of certain thresholds increases or decreases through time. For exampl, we can se that the probabiliyt of a 100-year flood increases through time.

6. Visualization: plotting the fitted GAM models in various ways to ease interpretation, and visually confirming predicted lake level values relative to reality.