Authors: Anna C. Cardall, Riley Chad Hales, Kaylee B. Tanner, Gustavious P. Williams, Kel N. Markert – Brigham Young University
Title: Google Earth Engine Notebooks for remote sensing models to generate coincident samples, perform feature engineering, and generate models using L1 regularization
Abstract: We present two Colab notebooks that accept a CSV file with sample date, locations, and values. The 1st notebook queries Google Earth Engine (GEE) to find the satellite image closest in time and space to the sample and extracts the band values from the satellite image. The notebook returns a file with the date, location, measured value, time offset (in hours) from the sample to the satellite image, and band values. The second notebook accepts a csv file in the format provided by the first notebook and uses Least Absolute Shrinkage and Selection Operator (LASSO or L1) regularization to generate an explainable model. The user can choose feature engineering options such as band differences, band ratios, band logs, and band inverses to generate more than 90 potential model parameters. L1 regularization can be used to search this large parameter space and select only a few features to use for the model. This allows researchers to analyze the selected model parameters and characterize physical meaning to the model and parameters. The current notebooks are designed to use 40 years of Landsat data, but can be easily modified to use other datasets, sensors, or satellites available on GEE. The Colab notebooks are available at https://github.com/BYU-Hydroinformatics/ee-wq-lasso.