Short notes on developments worth knowing about, even outside your own specific subfield — new datasets, shifting norms in the field, and techniques gaining traction.
For years, serious remote sensing work meant a capable desktop machine, locally stored imagery, and software like ENVI or ArcGIS running the classification. Platforms like Google Earth Engine changed the constraint: the imagery and the compute both live in the cloud, and a researcher can run a classification over an entire district from a browser tab. The practical effect is that the bottleneck has moved from "can my machine handle this" to "do I know which method to apply" — which is arguably a healthier bottleneck for a field to have.
It's still common to see land-cover classification accuracy reported as resubstitution accuracy — testing a classifier on the same points it trained on, which reliably overstates real-world accuracy. Reviewers and journals are increasingly expecting a genuine held-out test split instead. It's a small methodological change with an outsized effect on whether an accuracy number means anything — why validation method matters more than classifier choice.
Combining Landsat 5, 7, 8, and 9 into a single harmonized series now stretches back to 1984 — over four decades of consistent observation. That's long enough to distinguish a genuine multi-year trend from a single unusual year, which used to require either specialized preprocessing or a research group with existing access to curated archives. Run a harmonized 1984–present trend analysis in your browser →