Mapping Banana Plantations from Object-oriented Classification of SPOT-5 Imagery

The objectives of this research were to develop and evaluate an approach for object-oriented mapping of banana plantations from SPOT-5 imagery, and to compare these results to banana plantations manually delineated from high spatial resolution airborne imagery. Cultivated areas were first identified through large spatial scale mapping using spectral and elevation data. Within the cultivated areas, separation of banana plantations and other land-cover classes increased when including image co-occurrence texture measures and context relationships in addition to spectral information. The results showed that a pixel size of 2.5 m was required to accurately identify the row structure within banana plantations, which enabled object-based separation from other crops based on texture information. The user’s and producer’s accuracies for mapping banana plantations increased from 73 percent and 77 percent, respectively, to 94 percent and 93 percent after post-classification visual editing. The results indicate that the data and processing techniques used offer a reliable approach for mapping banana plants and other plantation crops. Introduction Remote sensing has been used extensively for crop yield estimation (Horie et al., 1992; Lobell et al., 2003; Singh et al., 2002; Sun, 2000), crop management (Pinter et al., 2003; Yang and Anderson, 1996), and precision farming (Basso et al., 2001; Robert, 1997). Other approaches have focused on a combination of crop models and the use of remote sensing (Basso et al., 2001; Doraiswamy et al., 2003; Moulin et al., 1998). Many of these approaches have used empirical models and spectral vegetation indices to predict crop variability, biomass, and yield (Tucker et al., 1980; Wiegand et al., 1991). Remote sensing has been used for mapping the extent of grain crops and to a limited PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Sep t embe r 2009 1069 Kasper Johansen and Stuart Phinn are with the Centre for Remote Sensing and Spatial Information Science, School of Geography, Planning and Environmental Management, The University of Queensland Brisbane, QLD, 4072, Australia ([email protected]). Christian Witte is with the Department of Environmental and Resource Management, QScape Building, 80 Meiers Road, Indooroopilly, QLD 4068, Australia. Seonaid Philip and Lisa Newton are with the Department of Natural Resources and Water, P.O. Box 2116, Mareeba, QLD, Australia. Photogrammetric Engineering & Remote Sensing Vol. 75, No. 9, September 2009, pp. 1069–1081. 0099-1112/09/7509–1069/$3.00/0 © 2009 American Society for Photogrammetry and Remote Sensing Mapping Banana Plantations from Object-oriented Classification of SPOT-5 Imagery Kasper Johansen, Stuart Phinn, Christian Witte, Seonaid Philip, and Lisa Newton extent, horticulture. However, there has been limited application of high spatial resolution image data because of the heterogeneity of the fields to be mapped in traditional per-pixel image classification (Navalgund et al., 1991; Shrivastava and Gebelein, 2007; Tennakoon et al., 1992; Yadav et al., 2002). The segmentation of image pixels into homogenous objects has been explored in several studies through clustering routines and region-growing algorithms (e.g., Haralick and Shapiro, 1985; Ryherd and Woodcock, 1996). The concept of segmentation is based on the theory of spatial scale in remote sensing described by Woodcock and Strahler (1987) who showed that the local variance of digital image data in relation to the spatial resolution can be used for selecting the appropriate image scale for mapping individual land-cover features. Image data of the Earth’s surface can be divided into homogenous objects at a number of different spatial scales, which are interrelated in a hierarchy, where large objects consist of several smaller objects (Burnett and Blaschke, 2003; Muller, 1997). Wu (1999) and Hay et al. (2003) explored different multi-scale image segmentation methods and found image objects to be hierarchically structured, scale dependent, and with interactions between image components. Object-oriented image classification typically consists of three main steps: (a) image segmentation, (b) development of an image object hierarchy based on training objects, and (c) classification (Benz et al., 2004; Blaschke and Hay, 2001; Flanders et al., 2003). Object-oriented image classification is based on the assumption that image objects provide a more appropriate scale to map environmental features at multiple spatial scales and more relevant information than individual pixels (Gamanya et al., 2007). The advantage of using objectoriented image analysis is the capability to define criteria for image objects at set scales using spectral reflectance characteristics, as well as within and between object texture, shapes of features, context relationships, and ancillary spatial data of different spatial resolution consisting of both thematic and continuous data values (Bock et al., 2005). Recently, object-oriented image classification has been used more extensively due to improvements of object-oriented segmentation and classification routines such as Definiens Professional 5 and Definiens Developer 7 (Benz et al., 2004; Definiens,