Efficient tree mapping through deep distance transform (DDT) learning

Abstract

Trees provide essential ecosystem services in urban areas, rural landscapes and forests. Individual tree information can inform forest and risk modelling, health studies and decision-making in public and non-governmental sectors. The increase in available remote sensing data and advances in automated object detection makes it feasible to map trees over large areas in unprecedented detail. Deep learning-based instance segmentation methods have thereby become the state-of-the-art in tree crown delineations tasks from aerial ortho-photography. Many of these methods are based on one- and two-stage detector frameworks such as Mask-RCNN and YOLO, which were developed focussing on speed and accuracy against common benchmark datasets. Another class of object detectors is based on encoder-decoder networks such as UNet which offer easy integration into existing workflows and high accuracy even in complex forest scenes in regional and national tree studies. While previous methods had to combine multi-model and multi-task outputs to create decision surfaces, we developed an efficient UNet-based modelling approach which focusses solely on learning the distance transforms of tree objects as cost surface for watershed segmentation. Our algorithm achieves superior instance segmentation across native forest, rural and urban environments in Aotearoa New Zealand, with an overall F1 score of 0.53 — 0.18 for small, 0.45 for medium and 0.67 for large crowns — surpassing previous approaches while decreasing modelling complexity, enabling fast and large-scale tree mapping.