We at Dianthus are delighted to announce the unveiling of our upgraded and second AI model, specially designed for comprehensive automated forest stand delineation. We have reengineered the Deep Learning Mask R-CNN Instance Segmentation model from Meta AI’s Detectron2 library. Through the adaptation of the initial 3-band (RGB) image interpretation routines, we’ve made it possible to incorporate an unlimited number of informational channels/bands for the model.
In this particular model, we’ve incorporated information channels such as NIR, Red, Green channels from aerial images, laser-measured tree height, Sentinel-2 VNIR (1 & 2) and Red images, NDVI derived images, laser-measured terrain slope, laser-generated water accumulation raster, and black & white aerial photographs from 1960, for training and using the model. The training process utilized roughly 800,000 manually delineated forest stands sourced from all across Sweden.
The selected 11 information channels were chosen meticulously to mimic human forest delineation procedures as much as possible. The human delineation process takes into account a broad array of biological, economical, silvicultural, environmental, and aesthetic factors. Using these 11 carefully selected information channels, our AI model is able to closely approximate the manual delineation.
Aside from the data utilized by the segmentation model, we also incorporate prior knowledge concerning aspects like agricultural land, marshland, environmental considerations, roads, lakes, and streams, thereby enhancing the quality of our delineation even further.
Our AI model is deployed for use in a C++ binary and can function in both CPU and GPU environments. The resulting forest stand delineation is utilized for our fully automated forest management plans, which are accessible via our commercial REST-API skogskartor.
We extend our gratitude to Sveaskog and Södra Skogsägarna for contributing crucial data and insight during the development phase.
