Remote sensing data includes images captured by satellite-based sensors that encode geospatial information. Image registration, an essential preprocessing step, involves aligning satellite images with geometrically precise reference images. Co-registering multi-temporal datasets to reference images is a prerequisite for reliable spatiotemporal analysis and related applications. In this paper, we present a novel satellite image registration method using Spatial Transformer Networks (STNs), where a localization network learns affine transformation parameters between the given input and reference image, and a grid generator-sampler produces the transformed image. The proposed model is capable of learning complex features/patterns despite the spatiotemporal changes. We evaluate our model on a custom multi-temporal ResourceSAT-2/2A LISS-3 dataset using two training approaches. Approach 1 involves applying the loss function after the grid generator-sampler stage, simulating practical scenarios where transformation parameters are unknown. Approach 2 applies the loss function after the parameter retrieval stage, leveraging known parameters from our custom dataset. In Approach 1, we evaluated four distinct models using various hyperparameter configurations and conducted a comparative analysis. The results, based on normalized cross-correlation (NCC) loss computed between the transformed and reference images using the Adam optimizer, were promising. In Approach 2, the NCC loss with the Adam optimizer was leveraged, but here it measures the error between the predicted and the known affine parameters. Our proposed method yields sub-pixel-level accurate image registration. We performed a comparative analysis of the results from Approaches 1 and 2 against existing methods to evaluate their relative performance.
Read more