THERMAL ANOMALY DETECTION BASED ON SALIENCY ANALYSIS FROM MULTIMODAL IMAGING SOURCES
- Institute of Photogrammetry and GeoInformation, Leibniz University Hannover, Germany
Keywords: thermal infrared imaging, thermal anomaly, saliency map, Dempster-Shafer theory, Random Forest, distributed heating systems
Abstract. Thermal anomaly detection has an important role in remote sensing. One of the most widely used instruments for this task is a Thermal InfraRed (TIR) camera. In this work, thermal anomaly detection is formulated as a salient region detection, which is motivated by the assumption that a hot region often attracts attention of the human eye in thermal infrared images. Using TIR and optical images together, our working hypothesis is defined in the following manner: a hot region that appears as a salient region only in the TIR image and not in the optical image is a thermal anomaly. This work presents a two-step classification method for thermal anomaly detection based on an information fusion of saliency maps derived from both, TIR and optical images. Information fusion, based on the Dempster-Shafer evidence theory, is used in the first phase to find the location of regions suspected to be thermal anomalies. This classification problem is formulated as a multi-class problem and is carried out in an unsupervised manner on a pixel level. In the following phase, classification is formulated as a binary region-based problem in order to differentiate between normal temperature variations and thermal anomalies, while Random Forest (RF) is chosen as the classifier. In the seconds phase, the classification results from the previous phase are used as features along with temperature information and height details, which are obtained from a Digital Surface Model (DSM). We tested the approach using a dataset, which was collected from a UAV with TIR and optical cameras for monitoring District Heating Systems (DHS). Despite some limitations outlined in the paper, the presented innovative method to identify thermal anomalies has achieved up to 98.7 percent overall accuracy.