Remote sensing
Mapping deforestation with Google Earth Engine and machine learning
Originally published by OLBGIS
Deforestation mapping is one of the clearest uses of satellite imagery and machine learning together. The question is simple: where was there forest, and where is it gone?
Google Earth Engine makes it practical for small teams. The imagery archive and the compute sit in the same place, so you can build and test a workflow without downloading terabytes.
Why Earth Engine
Earth Engine hosts the Landsat and Sentinel archives alongside tools for compositing, indices and classification. Because processing runs next to the data, a workflow that would take days of downloads can be prototyped in an afternoon, and rerun when new imagery arrives.
A workable approach
Start with cloud-masked composites for a before and an after period, chosen from the same season to limit false change. Add spectral indices such as NDVI as inputs, then train a classifier such as a random forest on labelled sample points.
Classify both periods, compare the maps, and compute area change. Validate on points the model never saw, ideally checked against higher-resolution imagery, and report accuracy per class rather than a single headline number.
Pitfalls to plan for
Persistent cloud cover, seasonal leaf loss and inconsistent definitions of what counts as forest account for most disappointing results. Published global products, such as the Hansen Global Forest Change dataset, make a useful baseline to compare against before you trust a custom model.