Abstract
This study investigates the amplitude artefacts induced by sublook decomposition (SD) in terrain observation by progressive scans synthetic aperture radar (TOPSAR) imagery and their implications for ship classification. The work presents a dual contribution: first, a theoretical framework is developed to derive the mechanism by which TOPSAR's inherent phase modulation is transformed into periodic amplitude artefacts during the SD process. Second, a novel, lightweight convolutional neural network, LeNet-4SD, is proposed to effectively leverage SD data. The model's architecture is distinguished by a multi-input design that processes azimuth sublooks, range sublooks, and the single look complex (SLC) amplitude image, through parallel branches fused by a spectrum-based attention mechanism. An evaluation on the OpenSARShip dataset demonstrates that LeNet-4SD achieves classification performance comparable to significantly larger benchmark models while being over two orders of magnitude more computationally efficient. To interpret the model's behavior, Shapley additive explanations was used to quantify the contribution of each data source. The analysis reveals distinct, class-specific dependencies, highlighting the model's reliance on sublook diversity for container ships and on the high-resolution SLC for tankers. These findings underscore the importance of accounting for processing-induced artefacts and establish that custom, lightweight, multi-input models can be a more effective and efficient strategy for synthetic aperture radar applications than fine-tuning large, pretrained networks.