Most methods for hyperspectral anomaly detection (HAD) construct profiles of background pixels and identify pixels unconformable to the profiles as anomalies. Recently, isolation forest-based algorithms were introduced into HAD, which identifies anomalies from the background without background modeling. The path length is used as a metric to estimate the anomaly degree of a pixel, but it is not flexible and straightforward. This paper introduces the half-space tree (HS-tree) method from the theory of mass estimation into HAD and proposes a metric involving mass information and tree depth to measure the anomaly degree for each pixel. More specifically, the proposed HS-tree-based detection method consists of three main steps. First, the key spectral-spatial features are extracted using the principal component analysis and the extended morphological attribute profile methods. Then, the ensemble of HS-trees are trained using different randomly selected subsamples from the feature map. Finally, each instance in the feature map traverses through each HS-tree and the anomaly scores are computed as the final detection map. Compared with conventional methods, the experimental results on four real hyperspectral datasets demonstrate the competitiveness of our method in terms of accuracy and efficiency.