Evaluating a modified StarDist Implementation for Individual Tree Detection and Crown Delineation in heterogeneous Landscapes

Abstract. Individual tree detection and crown delineation (ITDCD) in dehesa landscapes is complicated by geometric distortions from steep terrain, varying tree densities, and the partly multi-crown 'broccoli-like' structure of holm and cork oaks. This study evaluates the usability of a modified StarDist deep learning model, which has recently shown effectiveness for ITDCD in Canadian forests. Moreover, this study develops a workflow transforming the original StarDist, designed for microscopy images, into an ITDCD solution, taking the georeferencing of geospatial data into account. The tile-wise organized ground truth dataset is created with the pretrained Tree Segmentation model available in the ArcGIS Living Atlas, combined with manual revision. Several augmentation methods are applied, resulting in 960 images, which are split into 85 % for training and 15 % for validation. Following the approach of the Canadian forest study, the StarDist implementation is modified by introducing a constraint to the probability loss function. Rather than computing loss across all pixels, the modified loss function considers only pixels explicitly annotated as objects, while background pixels are excluded. An additional dataset of 1,200 trees serves as ground truth for testing the prediction across the entire study area. Using an Intersection over Union of 0.5, this test demonstrates good performance (Accuracy: 87.50 %; F1-score: 0.85). The accuracy varies with tree density: in areas with sparse tree cover, nearly all tree crowns are detected; in moderately dense areas, a number of tree crowns are missed; whereas in very dense tree layers, the frequency of missed detections increases.