Anomaly Detection in Image Streams with Explainable AI

With the increasing availability of high-quality image streams in various applications, anomaly detection of image streams is a field of study that is growing in popularity. This study proposed a novel framework for the early detection of anomalous behaviours present in image streams. The class imbalance problem, the interdependency between the images with regard to time, the lack of available labeled images, a data-driven anomalous threshold, and the trade-off of accuracy while increasing interpretability in the black-box optimization methods are some of the key aspects concerned during the model building process. By extracting meaningful features from image streams, the proposed approach quantifies the anomalous behaviour that may have occurred over some time. Such frameworks are vital in many application domains. In particular, calamities such as forest disturbances, the illegal extraction of sand from river channels, and land cover changes due to droughts and bushfires can be detected and mitigated, and other situations such as agricultural monitoring, including vegetation health and urban development assessments, can be provided with detection, possible prevention, and other assistance required through the proposed system. In this work, we define an anomaly as an observation that is very unlikely given the forecast distribution. The experimental design consists of three main modules: anomaly detection with machine learning, anomaly detection with deep learning, and explainable AI. Both the conventional machine learning and deep learning approaches for anomaly detection consist of three main components: computer vision, univariate time series forecasting, and an unsupervised anomaly detection component. In the conventional machine learning approach, Gabor Wavelet feature extraction, edge detected feature extraction, first- order features, and Gray-level Co-Occurrence Matrix feature extraction methods and Principal Component Analysis for dimensionality reduction is used for computer vision, and the Autoregressive Integrated Moving Average model is used for univariate time series forecasting, while the deep learning module uses Conventional Neural Network and Long Short Term Memory models for computer vision and time series forecasting, respectively. Both modules use box plot analysis and extreme value theory-based anomaly threshold calculation methods for the unsupervised anomaly detection component. The extreme value theory-based approach is proposed to calculate a data-driven anomalous threshold with valid probabilistic interpretations and thereby mitigate unrealistic assumptions made on the original distribution of the data. The forecast residuals are used for the anomalous threshold calculation process. Then, the forecast errors are compared against the extreme value theory-based data-driven anomalous threshold to determine the anomalies present in the image stream. Even though the pre-trained Convolutions Neural Networks based feature extractors play a major role in directing the overall anomaly detection module towards its success or failure, it is very much a black-box operation, and therefore an explainable AI module is used to explain the deep learning-based anomaly detection models. In this module using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) methods, high importance is given to explaining the feature extraction and their contribution to the final outcome of the anomaly detection module by evaluating the validity of the importance given to the extracted features by the anomaly detection models. The proposed framework is based on two main assumptions: anomalies show a significant deviation from the typical behaviour of a given system and a representative dataset of the system???s typical behaviour is available to define a model for the typical behavior of the image streams generated by a given system. With different data challenges, the performance of the designed frameworks is assessed using real data related to satellite images of an area affected by deforestation. One motivation for this application is that deforestation is a serious environmental issue in Sri Lanka, and an intelligent system that can detect any anomalous activity in forest cover would be beneficial. The performance evaluation suggested that employing a conventional machine learning approach to anomaly detection in image streams is more beneficial when the dataset available is smaller in size while utilizing a deep learning approach to anomaly detection in image streams is beneficial when a larger dataset is available. The extreme value theory-based anomalous threshold calculation method outperforms the conventional exploratory data analysis-based anomalous threshold calculation. The explainable AI module consists of a post-hoc, model agnostic, and local explanation that increases the trustworthiness of predictions given by the deep learning-based module. It further assists in optimizing its hyperparameters by giving insights into the internal workings of the black-box model. The proposed framework also inspires effective data visualization tools and thereby allows decision-makers to explore and easily understand detected anomalies using a combination of graphical and numerical methods.