Forest Cover Mapping

Forest cover mapping, what is the state of the art?

It’s been a long time since I first started working with satellite data for forest cover mapping. Like so many techy things, the scale and pace of change is almost unbelievable. I worked for a project that was arguably the first to map humid tropical forest worldwide at relatively high resoultion. We processed Landsat MSS and TM images using decidely old school techniques, what was state of the art at the time. Images were delivered on magenetic tape, not unlike the tape used for Led Zeppelin masters of the time, and unsupervised classification took hours, sometimes the whole night for a single scene.

Of course software like ClasLite and Google Earth engine came along later, much later. The idea of having the entire Landsate archive available for analysis from anywhere… this has been true for about 15 years, but still blows me away.

Recently, I’ve been diving into remote sensing products again (with an eye on Madagascar, more on this later). There are so many exciting things happening in recent years, and I’m just starting to catch up. A few things stand out:

• Sentinel, Sentinel-2 Leval 2A: freely available, relatively high resolution, orthorectified, radiometrically and atmospherically corrected. And there are cloud free composites*
• Machine learning: All new to me, but there are many Python implementations of neural network architectures like U-Net (PyTorch). At the end of the day, this “feels” similar to old-school supervised classification, but the underlying implementation is completely different

*The cloud free products for the area I’m interested in (and probably cloudy places everywhere) are a little dissapointing, so far…

So much more to say