Evaluation of Dependency between Network Size and Accuracy of Autoencoders for On-board Anomaly Detection on Remote Sensing Data

Abstract

In this master thesis the dependency between network size and accuracy of autoencoders for anomaly detection is investigated. A pipeline comprising hyperparameter tuning to find and train suitable models, testing and pruning of autoencoders is developed. Linear and convolutional autoencoders are considered. The reconstruction error of autoencoders trained to reconstruct non-anomalous data is used to detect anomalies. Experiments are conducted on three different remote sensing datasets, including a hyperspectral dataset, with maritime and forestry scenes to study the effect that a change of spectral dimension and different background and anomaly classes have on the detection results. Each tested autoencoder is pruned in an iterative process of structured pruning and retraining, increasing the sparsity of the autoencoder at each iteration. Simulated annealing is used to find sparsity distributions where layers with more parameters are pruned at a higher rate. In the datasets considered the autoencoders outperform classical approaches like the DCT and RX in most cases, and perform at-par otherwise. The pruning experiments show that linear models can be pruned at high rates without any performance loss, while the accuracy of convolutional models drop at higher pruning rates.