Deep learning for individual tree detection with noisy labels

Automated detection with deep learning opens the way for large-scale mapping of individual trees from aerial or satellite imagery. Convolutional neural networks offer unprecedented performance, under the condition that numerous and accurate labels are available to train and evaluate networks. Those two conditions are difficult to meet in the context of tree mapping, due to the high variability of tree shapes, species and environments, and to the lack of unambiguous ground truth data. Consequently, models learn on noisy data, do not reach optimality, and the errors seen during training are propagated to the predictions.

Here, we characterize and address the different types of noise in individual tree labels, notably comission/omission errors and positional errors. We propose a new method for tree detection, with an additional degree of freedom to account for annotation errors. We train and evaluate models on two large-scale datasets of aerial images in Denmark and France with manual annotations. Our approach, along with model ensembling, is able to learn from noisy point annotations and generalizes well to new areas, including dense forests.