A Novel Complex Encoder Transformation MLMVN Framework for Marine Vessels Detection Using SAR Images

The active data acquisition mechanism of Synthetic Aperture Radar (SAR) sets it apart from passive Earth observation sensors. Processing SAR data in the complex domain, rather than the real-number domain, is more intricate but essential to minimize information loss. This article introduces a novel Complex Encoder and Transformation Multi-Layered Multivalued Neuron (CET-MLMVN) framework that preserves the phase information of complex data. In this framework, the inputs, outputs, and weights are all complex numbers, and a non-gradient approach is used for weight updates. The study demonstrates how processing data in the complex domain reduces information loss, contributing to improved accuracy. The framework has been extensively analyzed concerning hyperparameters such as batch normalization, network depth, and activation functions and has been benchmarked against other architectures. It was tested for marine vessel detection using data from two spaceborne sensors (PALSAR-1, PALSAR-2), achieving overall accuracies of 99.25 and 98.99%, respectively. The CET-MLMVN framework significantly outperforms other methods when applied to the coherency matrix [T3] and its subchannels.