In recent work, there have been numerous approaches to 2D empirical wavelet transforms (EWT), which create 2D filter banks based on the information present in an image. These implementations of the EWT have shown promising results regarding image processing and computer vision tasks but their construction makes assumptions about the geometry of the spectrum. In this thesis, the author proposes a 2D empirical wavelet transform that creates filters whose shapes are fully adaptive to the information present in the signal. This is done by the application of the watershed transform, which defines a contour which separates select markers by the path of highest separation. The author then demonstrates the usefulness of such a transform by applying it to a texture segmentation computer vision problem.