data jungle adventures

We report on the tree, and vegetation detection in “no tree, no me” and also describe ambitions to detect canopy and individual trees. Recently, advances were made in ‘vulgarising’ deep learning and making it accessible in an applied urban planning and ecology context. DeepForest, a python package for predicting individual tree crowns from RGB imagery, was published with extensive documentation for non-ML-techy audiences.

It contains rebuild models, sample data and utility functions (yes, there is a cropping and projection function) that everyone working on spatial data and ML dreams of, with decent results. It may be limited by ‘only’ using RGB data making the package particularly suitable for UAV data and omitting some bands of coarser multi-spectral data.

We trialled the package in a desert environment and asked it to learn about Acaia saplings and tree pits for a construction monitoring task in 5 epochs. We only had 230 sampling points and created annotations by buffering point data (yes, lazy, very lazy). While we still got a box recall of 91 using a score threshold of 0.45, it still predicted mature trees without the ‘easy to learn’ pits.

The virgin script does not differ much from the ‘get started’ section in the doc lives here.

Future adventures will include detecting tree height using stereo and tri-stereo satellite imagery from the Pleiades or other sources. And here, it is interesting how the satellite image capture modes evolved. The capacity for gathering information and supporting use cases is increasing from mono to stereo and tri-stereo modes.

A tri-stereo satellite has three eyes, forward, at-nadir, and backward. It allows an increasing capturing capacity of urban areas, a higher accuracy of 3D models, detecting heights of buildings, et more. We are heading towards more adventures on trees and the three-eyes satellites.