ROTATION-INVARIANT SELF-SIMILARITY DESCRIPTOR FOR MULTI-TEMPORAL REMOTE SENSING IMAGE REGISTRATION

In this paper, a novel approach for the registration of multi-sensor remote sensing images with substantial time differences is proposed. The proposed method consists of four main steps. First, robust image features are extracted using the well-known UR-SURF (uniform robust-speeded up robust features) algorithm. Second, the feature descriptors are generated using a novel method based on self-similarity measure, named RISS (rotation invariant self-similarity). The RISS descriptor is an inherent rotation-invariant descriptor based on the gradient orientation histogram of correlation values and is very resistant against illumination differences. Third, the outlier rejection process is performed based on a simple improvement of graph transform matching, named LWGTM (localized weighted graph transformation matching). Finally, the estimation of the transformation model and the rectification process are done using TPS (thin-plate spline) model and the bilinear interpolation method. Five multi-sensor remote sensing image pairs with relatively long years of time difference are used for evaluation. The results indicate the capability of the proposed method for reliable remote sensing image registration. The average recall, precision, the number of extracted matched points and the average registration accuracy of the proposed method are about 31.6, 39.5, 4940, and 1.8 pixels, respectively.

Automated summary

This paper presents a novel approach for the registration of multi-sensor remote sensing images with substantial time differences. The proposed method shows significant advantages in feature extraction, feature descriptor generation, outlier rejection, and transformation model estimation, and it has been demonstrated to be reliable through experiments.