A Systematic Review of Individual Tree Crown Detection and Delineation with Convolutional Neural Networks (CNN)

• Published in: Current forestry reports Vol. 9; no. 3; pp. 149 - 170
• Main Authors: Zhao, Haotian, Morgenroth, Justin, Pearse, Grant, Schindler, Jan
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.06.2023; Springer Nature B.V
• Subjects: Accuracy; Artificial neural networks; Data integration; Deep learning; Delineation; Earth and Environmental Science; Ecology; Environment; Environment models; Environmental Management; Forest management; Forestry; Forestry Management; forests; Machine learning; model validation; Nature Conservation; Neural networks; Remote sensing; Remote Sensing (J Suarez; Section Editor; Sustainable Development; Systematic review; Topical Collection on Remote Sensing; Training; tree crown; trees; Deep learning; Instance segmentation; Crown delineation; Tree detection; Forestry; Object detection; Remote sensing
• ISSN: 2198-6436; 2198-6436
• DOI: 10.1007/s40725-023-00184-3

Purpose of Review Crown detection and measurement at the individual tree level provide detailed information for accurate forest management. To efficiently acquire such information, approaches to conduct individual tree detection and crown delineation (ITDCD) using remotely sensed data have been proposed. In recent years, deep learning, specifically convolutional neural networks (CNN), has shown potential in this field. This article provides a systematic review of the studies that used CNN for ITDCD and identifies major trends and research gaps across six perspectives: accuracy assessment methods, data types, platforms and resolutions, forest environments, CNN models, and training strategies and techniques. Recent Findings CNN models were mostly applied to high-resolution red–green–blue (RGB) images. When compared with other state-of-the-art approaches, CNN models showed significant improvements in accuracy. One study reported an increase in detection accuracy of over 11%, while two studies reported increases in F1-score of over 16%. However, model performance varied across different forest environments and data types. Several factors including data scarcity, model selection, and training approaches affected ITDCD results. Summary Future studies could (1) explore data fusion approaches to take advantage of the characteristics of different types of remote sensing data, (2) further improve data efficiency with customised sample approaches and synthetic samples, (3) explore the potential of smaller CNN models and compare their learning efficiency with commonly used models, and (4) evaluate impacts of pre-training and parameter tunings.