Characterizing the Impact of Common Electro-Optical Sensor Anomalies on Maritime Image Classifiers

As the number of Earth observation imaging systems has surged, the potential applications for satellite imagery have expanded significantly, from enhancing maritime security to monitoring environmental changes. However, the growing volume and complexity of data pose new challenges for detecting and classifying objects, such as vessels, using Deep Learning (DL)-based computer vision models. One critical issue is the impact of sensor anomalies-unexpected variations in sensor performance-that can severely affect model accuracy. Previous work has demonstrated that image classification architectures can have large performance variations when sensor anomalies are introduced. As EO sensor anomalies may occur throughout the life-cycle of space-based sensors, it is important to characterize performance impacts across a range of models and investigate whether techniques may be applied to increase overall model robustness. This paper addresses this challenge by investigating how sensor anomalies influence image classification performance across various model architectures and complexities. We find that sensor anomalies can notably degrade classification accuracy. To mitigate this, we introduce and evaluate two approaches aimed at enhancing model robustness: training models with domain-specific pre-trained weights and incorporating augmented, anomalous data into the training process. Our results show that both methods effectively improve model performance, with augmented data providing a more robust solution. Overall, further work is needed to understand how image classification models may be made more robust so that practical applications using these techniques may be reliable.