Code for generating Landsat reflectance time series, water chemistry matchups, and machine learning models to predict water chemistry for US lakes

Code and models associated with the generation of the LAGOS-US LANDSAT extension module, which contains 45,867,023 sets of reflectances from Landsat 5, 7, and 8 as well predicted water quality for chlorophyll, Secchi deph, true water color, dissolved organic carbon, total suspended solids, and turbidity for 136,977 lakes 4+ hectares in the conterminous US based on matchups to in situ data from LAGOS-US LIMNO. Contained here are the following files:

GEE_Extraction_Code.txt: code used to extract time series of Landsat reflectances for a list of LAGOS-US lakes using the shapefiles contained within the LAGOS-US LOCUS data module using Google Earth Engine. The code harmonizes among sensors, calculates pixel-specific band ratios, and exports a CSV of the time series for each lake. 

lagos_us_matchup.R: Creates a matchup dataset for a window of +/- 7 calendar days using LAGOS-US LIMNO data and the compiled CSVs generated by GEE.

lagos_us_rs_model.R: Suite of 40 random forest models run for water quality variable predictions, including testing the usage of log10-transformed response, band ratio usage, the removal of scenes with negative median reflectances, and the removal of scenes with >50% cloud cover. The selected final model for each water quality variable is used to make final predictions for the data module.

Each of the final model fitted objects is contained within this model and code product in an individual RData file:

Colort: true water color (pcu) - log10-transformed

CHL: chlorophyll (micrograms/L) - log10-transformed

NTU: turbidity (NTU) - log10-transformed

TSS: total suspended solids (mg/L) - log10-transformed

DOC: dissolved organic carbon (mg/L) - log10-transformed

Secchi: Secchi depth (m) - untransformed

lagos_us_pred_obs.R: Script that calculates and plots predicted vs in situ output from each of the six water quality variable models and generates plots of mean absolute percentage error (MAPE) by range of each water quality variable.

The addendum files associated with the lagos_us_tt_full_compare.R script contain Generalized Random Tessellation Stratified (GRTS) test and train data splits as well as a comparison of the original full data random forest model to a comparable model using test and train splits. The predictions from these data are also included in the EDI repository associated with this data product.