EchoCNN-Denoiser: a reservoir computing inspired deep learning model for enhanced synthetic aperture radar image despeckling

The occurrence of speckle noise in synthetic aperture radar (SAR) images significantly impacts the accurate extraction of essential information for remote sensing applications. To address this issue, a denoising model, named EchoCNN-Denoiser, is proposed that utilizes the combined strengths of convolutional neural networks (CNN) and reservoir computing (RC) to effectively minimize speckle noise and improve the quality of SAR images. RC offers a method that employs the temporal processing capabilities of recurrent neural network (RNN) without the complexity of training them. We utilized CNN for spatial feature extraction from SAR images, and echo state network (ESN), a concept of RC, is utilized for capturing temporal dependencies, resulting in a robust framework for SAR image denoising. CNN extracts hierarchical features using convolution and pooling operations and the ESN transforms these features into a higher-dimensional space through a sparsely connected reservoir. To optimize performance, the reservoir-to-output weights are trained with parameters such as spectral radius, input scaling, and sparsity. Through extensive experiments on real and virtual SAR datasets, the proposed technique demonstrates superior performance in noise reduction compared with existing methods. The model is also validated using multiple evaluation techniques. Initially, our model is compared against traditional filters and other deep learning denoising methods to evaluate its relative performance which is tested using image quality metrics such as peak signal-to-noise ratio, structural similarity index measure, equivalent number of looks, perception-based image quality evaluator and blind/referenceless image spatial quality evaluator, and perceptual index, achieving scores of 30.01, 0.90, 9.48, 27.22, 41.65, and 34.44, respectively. Following that, layer-wise relevance propagation (LRP) is applied to gain a deeper understanding of the model predictions. The heatmaps obtained from LRP visualization confirm that EchoCNN-Denoiser effectively preserves essential image structures while reducing noise. Further, a paired t-test is conducted to statistically assess the effectiveness of the model. Finally, an ablation study is performed to evaluate the contribution of each component to the overall performance of the model. These validations demonstrate the effectiveness of EchoCNN-Denoiser in producing high-quality, despeckled SAR images. As noise reduction is the primary step of SAR image processing, the proposed technique has several applications in remote sensing, by significantly enhancing the SAR image quality.