Predict just about anything with Google Earth Engine. Part I

Sampling strategies and tools: what you need to work through before applying machine learning or AI

Would you like to be able to develop and prepare the data you need to pose, explore, and answer the most pressing and complex questions in your field of research? This course concerns itself with one of the most demanding and least covered parts of developing a predictive model for precision agriculture, or just about anything: sampling.

What you’ll learn

  • How to use Google Earth Engine (GEE) to sample and pre-screen predictor variables to develop spatially explicit predictive models.

Course Content

  • Getting started –> 3 lectures • 1hr 25min.
  • Compute and sample covariates and explore their predictive power on Y –> 3 lectures • 1hr 7min.
  • Advanced tools for data aggregation and extraction. –> 3 lectures • 1hr 20min.

Predict just about anything with Google Earth Engine. Part I

Requirements

  • High school level algebra.
  • University level statistics.
  • Basics of spatial analysis.
  • Basics of Google Earth Engine.

Would you like to be able to develop and prepare the data you need to pose, explore, and answer the most pressing and complex questions in your field of research? This course concerns itself with one of the most demanding and least covered parts of developing a predictive model for precision agriculture, or just about anything: sampling.

When studying machine learning through video tutorials you normally access somebody’s dataset and learn how to apply algorithms. But how were those neat datasets created? This course details how to use and adapt to your unique needs some tools I developed to sample just about any spatially explicit variable through the Google Earth Engine Platform. This course is biased in favor of herbaceous crops but the tools presented are flexible enough to be adapted to your research interests.

In this course you will learn about a complete workflow to identify and extract covariates with predictive power:

· One practical solution to cloud and shadows filtering.

· Filtering a collection.

· Mapping a function over a collection.

· Applying masks to images and image collections.

· Composite RGB images.

· Land Cover Land Use (LCLU) Classification.

· Stratified and balanced sampling strategies.

· Split training and validation dataset.

· Results visualization.

· Add a legend.

· Classification accuracy assessment.

· Computing area in hectares for each LCLU class.

· Exporting LCLU raster data as assets for reuse.

· Convert raster to vector data and export as assets for reuse.

· Build time series aggregating spatially and over time windows.

· Convert Sentinel 1 SAR data from linear to decibel.

· Compute soil moisture from Sentinel 1.

· Infer NDVI from Gravimetry through Ordinary Least Squares regression.

· Compute long term statistics using balanced samples.

· Identify the lowest/highest performing pixels on the property for corrective purposes.

· Screen covariates for predictive purposes.

· Tools to build increasingly elaborate and complex time series.

· Aggregate data from different datasets at different time granules.

· Create parcels/paddocks/pixels for meaningful aggregation.

· Export the data as .CSV