. .  , 2003, . 24, . 4, 885–890Calculating environmental moisture for per-field discrimination ofrice cropsT. G. VAN NIEL*, T. R. McVICARCSIRO Land and Water, PO Box 1666, Canberra, ACT 2601, AustraliaH. FANG and S. LIANGDepartment of Geography, University of Maryland, College Park, MD 20742,USA(Received 10 December 2001; in final form 20 June 2002)Abstract. The accuracies of rice classifications determined from density slicesof broadband moisture indices were compared to results from a standard super-vised technique using six reflective Enhanced Thematic Mapper plus (ETM+)bands. Index-based methods resulted in higher accuracies early in the growingseason when background moisture differences were at a maximum. Analysis ofdepth of ETM+ band 5 resulted in the highest accuracy over the growing season(97.74%). This was more accurate than the highest supervised classificationaccuracy (95.81%), demonstrating the usefulness of spectral feature selection ofmoisture for classifying rice.1. IntroductionDiscriminating rice from other crops to a high level of accuracy is commonplaceusing various clustering algorithms, namely supervised classifiers (e.g. Barbosa et al.1996). In structured environments, the addition of contextual information canincrease accuracy over per-pixel classifications (e.g. per-field classification, Aplin et al.1999). For the specific case of highly regulated agricultural systems where within-crop phenological differences are small, it is also true that a single spectral featuremay capture most between-crop variability. Isolating the times when this occurscould (i) simplify the classification procedure considerably, and (ii) possibly increaseaccuracy further by eliminating superfluous information. In this study, we test theability of spectral features related to ‘environmental’ moisture content to discriminaterice and non-rice crops. Environmental, as defined here, describes the mixture ofplant canopy and background reflectance, and is used to differentiate the currentstudy from leaf-scale studies.In the following discussion, formulae are adapted to the mean wavelength(MWL, weighted by bandpass function) of Enhanced Thematic Mapper plus(ETM+) bands. Three spectral features relating to environmental moisture areused, the Normalized Difference Infrared Index (NDII=(r835nm−r1650nm)/*Corresponding author; e-mail: [email protected] Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online © 2003 Taylor & Francis Ltdhttp://www.tandf.co.uk/journalsDOI: 10.1080/0143116021000009921T . G. Van Niel et al.886(r835nm+r1650nm), Hardisky et al. 1983), the Moisture Stress Index (MSI=r1650nm/r835nm, Hunt and Rock 1989), and the depth of ETM+ band 5 (D1650nm=1−R∞, adapted from Kokaly and Clark 1999). Depth of bands within absorptionfeatures has been used to estimate various plant-based properties from hyperspectralremote sensing. Continuum-removed reflectance (R∞) is determined for every channelwithin the absorption feature by dividing reflectance of the channel by the value ofa reference, or continuum line forming a ‘ceiling’ above the entire absorption feature(see Kokaly and Clark 1999). In its most simple case, band-depth analysis requiresthree bands, two defining the ends of the continuum line and one between the endpoints to be related to the continuum line. This three-point simplification permitsthe application of band-depth analysis on broadband imagery. For this study, con-tinua end points are established at the MWL of ETM+ bands 4 (835 nm) and 7(2208 nm), allowing for the calculation of the continuum-removed reflectance for theMWL of band 5 (1650 nm) using the following equationR∞=r1650nm(r835nm(1−c))+(r2208nmc)(1)where the constant c represents the relative position of band 5 to, in this case, band 7c=1650−8352208−835#0.59359 (2)Because r1650nmchanges more rapidly with increasing water than r835nmandr2208nm(Gausman et al. 1978, Tucker 1980), R∞ and therefore D1650nmare related toenvironmental moisture. Figure 1 illustrates this relationship, where D1650nmis greaterfor flooded rice compared to intermittently irrigated soybeans. The distinction betweenrice and soybeans is maximized prior to canopy closure (figure 1), where backgroundwater and soil signals, respectively, have an opposing affect on D1650nm. Temporalvariations due to different growth rates and management practices (e.g. date ofplanting) result in different surfaces, affecting the amount of spectral separabilitybetween rice and non-rice crops. In this case, discrimination between rice and soybeansis enhanced early in the growing season with respect to environmental moisture.2. Study siteColeambally Irrigation Area (CIA), New South Wales, is approximately 95 000 hain size, comprising over 500 farms; fields are large (up to 70 ha) and well maintained.Primary summer crops are rice, maize and soybeans. Timing of generalized surfacesignals is summarized across the bottom of figure 2. Rice is sown in early Octoberand starts senescing in late February. Maize is sown about the same time as rice,yet matures faster, so it is often well into senescence by the end of February. Soybeansare sown later, emerging from the soil in early December with senescence startingin early–mid March. All three crops are harvested from March–April. As opposedto flooded rice, maize and soybeans are only intermittently irrigated.3. Methods and discussion of results3.1. Image pre-processingTwelve cloud-free ETM+ images were acquired between October 2000 andMarch 2001. Fine spatial resolution digital aerial photographs (1.5 m pixel size),acquired in early January 2001, were used to geo-reference the ETM+ images.Atmospheric correction was then applied to all ETM+ images. First, surfaceRemote Sensing L etters 887Figure 1. Continuum removed reflectance (R∞) and depth of ETM+ band 5 (D1650nm) foropen and closed canopies of rice and soybeans. These ground spectra were collectedat the CIA using an Analytical Spectral Devices (ASD) radiometer.reflectance was calculated for two images when simultaneous ground reflectancewas measured (3 February and 7 March). Image reflectance was calculated usingMODTRAN4 simulations of three atmospheric parameters (path radiance, transmit-ted flux and spherical albedo) for various atmospheric visibility values. Optimumvisibility for clear sky conditions was calculated (50 km) for these two