Due to their role in certain essential forest processes, dead trees are an interesting object of study within the environmental and forest sciences. This paper describes an active learning-based approach to detecting individual standing dead trees, known as snags, from ALS point clouds and aerial color infrared imagery. We first segment individual trees within the 3D point cloud and subsequently find an approximate bounding polygon for each tree within the image. We utilize these polygons to extract features based on the pixel intensity values in the visible and infrared bands, which forms the basis for classifying the associated trees as either dead or living. We define a two-step scheme of selecting a small subset of training examples from a large initially unlabeled set of objects. In the first step, a greedy approximation of the kernelized feature matrix is conducted, yielding a smaller pool of the most representative objects. We then perform active learning on this moderate-sized pool, using expected error reduction as the basic method. We explore how the use of semi-supervised classifiers with minimum entropy regularizers can benefit the learning process. Based on validation with reference data manually labeled on images from the Bavarian Forest National Park, our method attains an overall accuracy of up to 89% with less than 100 training examples, which corresponds to 10% of the pre-selected data pool.     «

Due to their role in certain essential forest processes, dead trees are an interesting object of study within the environmental and forest sciences. This paper describes an active learning-based approach to detecting individual standing dead trees, known as snags, from ALS point clouds and aerial color infrared imagery. We first segment individual trees within the 3D point cloud and subsequently find an approximate bounding polygon for each tree within the image. We utilize these polygons to ext...     »