Approaches for optimal automated individual tree crown detection in regenerating coniferous forests

Automated tree detection provides a means to acquire information on tree abundance and spatial distribution, both of which are critical for evaluating the status of regenerating forests. It is also often a precursor to automated tree delineation, which typically utilizes image data surrounding a detected crown point. However, obtaining consistently accurate detection results has proven difficult owing to errors associated with image scale. In this paper, four approaches that reduce this scale dependence are evaluated, including (1) determination of optimum global image smoothing to apply predetection, (2) determination of optimum local image smoothing to apply predetection, (3) determination of the optimal local window size for use in the detection algorithm, and (4) post-detection merging of initially defined crown segments. Each approach was applied to three datasets acquired by different sensors and with different regenerating forest conditions. A common local maximum tree detection algorithm was implemented for approaches 1-3, and a watershed segmentation algorithm was applied in approach 4. Detection accuracy was evaluated using standardized methods. The highest accuracies for each dataset were obtained with approaches based on local scale representations where the regenerating structure favored such approaches. However, more consistent accuracies across all datasets were obtained with the optimum global scale approach. Post-detection merging of adjacent crown segments produced the poorest results. Error sources and the advantages and disadvantages of each approach are discussed in terms of developing more operational methods for automated tree detection in regenerating forests.