This thesis investigates the feasibility of applying image anomaly detection models with segmentation capabilities for monitoring forested terrain using top-down RGB images captured by unmanned aerial vehicles. The images are acquired from a dataset containing real-world images of Swedish forests. Eleven different models from the Anomalib Python library are tested, comprising four underlying architectures. The models are fine-tuned on normal images and evaluated using synthetic superimposed anomalies. Furthermore, a novel approach for evaluating the discriminative ability of image anomaly detection models in various lighting conditions is presented. The experiment results show that while adopting image anomaly detection models for terrain surveillance purposes is possible, limitations are encountered; some tested image anomaly detection models do not see any improvement from fine-tuning to the target domain, indicating that they are not suitable for the domain. Significant differences in discriminative ability, fine-tuning time, and prediction time were observed between the tested models, where no model scored the highest in the tested metrics, indicating that model selection should be use-case-dependent. No correlation is observed between discriminative ability and the time required for fine-tuning.