Abstract: Traditional algorithm for matching multi-sensor images are mostly based on region information, which is sensitive to image rotation. This paper presents a new rotate-invariant feature named Gradient Radius Angle Pyramid Histogram(GRAPH), that can be used in matching multi-sensor images of arbitrary angle. We calculate the Gradient Radius Angle(GRA) based on normalized gradient vector and radius direction vector. The GRA is invariant to image rotation and robust to image noise, and illumination change. To get sparse description, we use circle region histogram to represent the statistic distribution of GRA. Pyramid Histogram(PH) is utilized to replace simple histogram for better distinctive capability. Experimental results show that the GRAPH feature can distinguish images of different scenes and accommodates to image rotation. Multi-sensor images are reliably matched based on GRAPH feature.