Local and Hierarchical Feature Fusion for Registration of Satellite Images Using Residual Deep Neural Network Modified with Attention Mechanism and Depth-Wise Separable Convolution
Abstract
An accurate multi-sensor satellite image registration is a fundamental prerequisite for an optimal preprocessing step in a wide range of satellite image applications. This study proposes an innovative approach that integrates a residual deep neural network (RNN) model modified with an attention mechanism and depth-wise separable convolution (AM-DSC) layer, for hierarchical feature investigation and scale invariance feature transform (SIFT) for local features. These extracted features undergo dimensionality reduction through principal component analysis (PCA), followed by feature selection using the L2 Norm distance metric. Subsequently, random sample consequence (RANSAC) is adopted for image resampling and transformation. Considering the non-specific surface area captured, the algorithm has been evaluated with ten different datasets of synthetic aperture radar (SAR) and multispectral image pairs to demonstrate the accuracy of the proposed algorithm. This investigation was conducted using evaluation measures, including root mean square error (RMSE), RMSE cross-validation based on leave-one-out (RMSELOO) method, correctly matching ratio (CMR), peak signal-to-noise ratio (PSNR) and spectral similarity index measure (SSIM). The proposed methodology demonstrated superior performance, achieving a CMR of more than 40% and enhancing registration accuracy as measured by RMSE, PSNR, and SSIM.