Injecting Domain Knowledge Into Deep Neural Networks for Tree Crown Delineation

Automated individual tree crown (ITC) delineation plays an important role in forest remote sensing. Accurate ITC benefits biomass estimation, allometry and species classification among other related tasks, all of which are used to monitor health make decisions management. In this paper, we introduce Neuro-Symbolic DeepForest, a convolutional neural network (CNN) based algorithm that uses neuro-symbolic framework inject domain knowledge (represented as rules written probabilistic soft logic) into CNN. We create encode concepts for competition, allometry, constrained growth, mean area, color. Our results show the model learns from annotated training data well under some conditions, injection improves performance affects bias. then analyze effects each rule on its aspects performance. find addition can improve F1 by much 4 F1-points, reduce KL-divergence between ground-truth predicted area distributions, aggregate error delineations.