Abstract: A high performance anomaly detection model has been constructed to address the problem of sparse anomalous image data in the real world. A two stage framework anomaly detection model is built using only normal training data and a small amount of synthetic anomaly sample. First， a ResNet 18 encoder model is trained to extract representation by the pretext of classifying normal data and synthetic anomaly data. Then， a single classifier for anomaly images is built through modelling the distribution of normal data representations using Gaussian density estimation. GradCAM is applied to extend the model， enabling the anomaly detection model to locate anomaly regions without labels. Finally， experiments are conducted on a simulated anomaly detection dataset using real world images， demonstrating that the proposed algorithm can detect anomaly and provide location results in remote sensing images that are even difficult to recognize with human eyes.

Key words: anomaly detection, remote sensing, deep learning, convolutional neural network