Available at www.sciencedirect.com journal homepage: www.elsevier.com/locate/issn/15375110 Research Paper: PA—Precision Agriculture Non-rigid registration of hyperspectral imagery for analysis of agronomic scenes Hector Erives a,Ã , Glenn J. Fitzgerald b , Thomas R. Clarke c a Electrical Engineering Department, New Mexico Tech, Socorro, NM 87801, USA b Department of Primary Industries, Horsham,Victoria 3401, Australia c US Arid Land Agricultural Research Center, Maricopa, AZ 85239, USA article info Article history: Received 13 January 2007 Received in revised form 11 June 2007 Accepted 15 August 2007 Available online 17 October 2007 Analysis of remote-sensing imagery usually entails the registration of images from multiple different wavelengths. Even though a staring instrument has the advantage of readily producing coherent spectral images, often these images still need some form of band-to-band registration to correct for instrument movement, pixel shift caused by vibration at different acquisition times or optical distortion introduced by difference in the optics. In this study, a method is reported to register hyperspectral images. The method consists of partitioning the pair of registering images (from different wavelengths) into multiple overlapping regions of interest (ROIs) and finding registration errors in each of these areas using the standard phase correlation (PC). The registration errors from the ROIs are then used to compute a geometric transformation, which is applied to the entire image to correct for non-rigid image registration errors that are mostly due to optical effects. This technique was tested on hyperspectral imagery acquired by a portable hyperspectral- tunable imaging system developed by the US Arid-Land Agricultural Research Center (USALARC) for ground and remote-sensing applications that include vegetative analysis for precision agriculture. & 2007 IAgrE. Published by Elsevier Ltd. All rights reserved. 1. Introduction Image registration is the process of aligning two or more images of the same scene taken at different times, from different view points, by different sensors or at different wavelengths. The misalignment of the images may come from different sources, including sensors that are flown at different times over the same scene and images acquired at different ground resolutions and instruments with more than one optical component. In the case of spectral images, the registration process consists of designating a reference or base image, and aligning images acquired at other wave- lengths with respect to a reference waveband. Registration of images is a first and important step in remote-sensing applications that require multi-spectral, multi-temporal observations. There are two main methods that have been extensively referenced in the literature such as area- and feature-based methods. The area-based match- ing methods are among the most popular image registration approaches (Shimizu & Okutomi, 2006; Coulter et al., 2003; Hong & Zhang, 2005). These methods consist of computing the correlation of all the pixels within a region of interest ARTICLE IN PRESS 1537-5110/$ - see front matter & 2007 IAgrE. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.biosystemseng.2007.08.006 Ã Corresponding author. E-mail addresses: erives@ee.nmt.edu (H. Erives), Glenn.Fitzgerald@dpi.vic.gov.au (G.J. Fitzgerald), TClarke@uswcl.ars.ag.gov (T.R. Clarke). BIOSYSTEMS ENGINEERING 98 (2007) 267– 275