A basic understanding of rice physiology is essential to the success of remote sensing applications in rice-based agricultural systems. This knowledge can play a critical role in the planning stages of a remote sensing project (e.g., identifying optimum acquisition dates for the purchase of imagery) as well as in the final stages of analysis (e.g., aiding in the delineation of rice paddies or the estimation of growth stages) (Ribbes and Toan, 1999, Le Toan et al., 1997). The growing cycle of rice can be separated into two stages with respect to most analyses of remotely sensed data: vegetative and reproductive (Casanova, 1998, Ribbes and Toan, 1999).
The vegetative stage includes the part of the growth cycle where the plant develops and grows, starting after sowing and ending when the plants start to reproduce. This stage is characterised by a steady increase in plant height and biomass. The reproductive stage starts when the plant stops growing taller and ends after maturity and includes panicle and grain development (Ribbes and Toan, 1999). It may be beneficial at times to further split the reproductive stage into two categories: reproductive pre-heading and reproductive post-heading. Reproductive pre-heading defines the period from panicle primordia initiation to heading and post-heading refers to the period from heading to maturity (Casanova, 1998).
The length of the growth cycle of rice can vary from 3 to 6 months for different varieties (Casanova, 1998), and can also be categorised into two main groups for many remote sensing applications: tropical and temperate. Growth cycles for tropical rice varieties last about 110-120 days, while those of temperate varieties usually last around 140-150 days (Le Toan et al., 1997). However, this duration can vary based on cultivar. For example, short duration varieties have been bred with growth cycles less than 90 days (Senanayake et al., 1994). These differences in growth cycle length are due to differences in vegetative stage duration: the vegetative stage can be anywhere from 40 to 120 days in length (Senanayake et al., 1994).
Irrespective of cultivar, reproductive pre-heading duration is about 23-25 days, while reproductive post-heading duration lasts 30-35 days (Senanayake et al., 1994). During the reproductive stage, plant height and biomass typically remain stable at around 100 cm and 2000 gm-2, respectively (Ribbes and Toan, 1999). The vertical characteristics of the rice plant also change as the plants grow, with stem inclination decreasing and the leaf angle increasing (Ribbes and Toan, 1999).
For different locations, the timing of the growth cycle of rice varies depending on local climate, management, and cultivar planted. Consequently, the timing of rice for a particular site of interest should be known. Australian rice varieties have changed over the past 40 years with Caloro dominating the 1960's, Calrose in the 1970's to mid 1980's and Amaroo from the mid 1980's into the 1990's (Brennan et al., 1994). This has been associated with development of long-grain, fragrant and Spanish varieties to meet higher priced markets (Brennan et al., 1994). Short duration varieties are generally desirable because they can be competitive with weeds, require less pesticides and herbicides, utilise less irrigation water, and allow for double cropping in tropical environments (Khush, 1987). In a temperate environment like southern NSW, short duration varieties are also desirable because they can allow more leeway for sowing and harvesting between the limitations of cold springs and autumns (Reinke et al., 1994). Timing of rice in southern NSW is summarised as:
1. placed under permanent flood and aerially sown in late September/early November;
2. canopies emerging during late October/late November;
3. flowering by late January/early February;
4. de-watered in late February/March; and
5. harvested in March/May.
Other summer crops in NSW include corn, sorghum, and soybeans, while winter crops include wheat, barely, oats, and canola. Pasture is grown in both seasons. Citrus, stone fruits, and grapes are also grown in the area. Some of these crops may use remnant soil moisture after a flooded rice crop, whereas others are furrow or drip irrigated.
There is an interest in monitoring other crops than rice within the irrigation areas of southern NSW from remote sensing. Of these other crops, there is a good potential for remote monitoring of corn and soybeans. The spectral reflectance characteristics of corn and soybeans along the EMS are slightly different in shape and amplitude (Thenkabail et al., 2000) allowing for differentiation between these two crops (Badhwar et al., 1982). Multitemporal remote sensing data has been used to estimate soybean and corn crop characteristics such as yield, LAI, biomass, plant height (Thenkabail et al., 1994a, Thenkabail et al., 1994b), development stage (Badhwar and Henderson, 1985), and crop proportion (Badhwar, 1984b, Badhwar, 1984a). For single date imagery, however, the timing of image acquisition can greatly influence classification results since confusion between spectral signatures can occur due to differences in crop growth stages. That is, on the day of image acquisition, the two crops could look spectrally similar.
Moderately high correlations (from 0.7 to 0.85) have been reported between several soybean and corn crop characteristics when related to VI's. Soybean was correlated to standard NIR and red-based VI's , whereas corn crop characteristics were more highly correlated with VI's that include at least one MIR band (Thenkabail et al., 1994b, Thenkabail et al., 1994a). Mature soybean crops have higher reflectance in the NIR and lower reflectance in the red portions of the EMS than corn, resulting in detectably higher standard VI values for soybeans (Tucker et al., 1979). This means that three of the main summer crops (rice, corn, and soybeans) can potentially be discriminated from each other using remote sensing. However, to do this, more than one image throughout the growing season might be needed in order to take advantage of spectral differences due to the phenology of these crops.
Remote sensing based applications, then, will not only take advantage of both the characteristics and timing of growth cycle, but will also consider the spectral reflectance of different crops. Since rice is the focus of this report, the basic spectral patterns of rice must be understood. The reflectance from rice, like all green vegetation, can be summarised by a generalised vegetation response as seen in Figure 3. It is the differences in this basic vegetation response that allow discrimination between vegetation types. However, these vegetative-type responses are harder to differentiate between each other than a non vegetative-type response like soil or water. This is true since non vegetative-type features usually reveal drastically different response curves when compared to vegetation (Figure 3).
As irrigated rice fields are flooded, the spectral characteristics of water can be used to distinguish potential rice paddocks and provide an early estimate of rice area (Barrs and Prathapar, 1996, McCloy et al., 1987). Inaccuracies result, however, when this early estimate is not adjusted by a later image, which can aid in elimination of permanent water bodies and other irrigated crops from the classification (Barrs and Prathapar, 1996, McCloy et al., 1987). The visible and near infrared wavelength response of rice, once the vegetation starts to cover the water in flooded paddocks, is much the same as other crops (Martin and Heilman, 1986). However, rice was found to be more distinguishable from other crops due to its water absorption characteristics by including middle infrared (MIR) wavelengths in the crop discrimination(Martin and Heilman, 1986, Thenkabail et al., 1994b).
It is also very important to understand the interaction of the reflectance of vegetation in key bandwidths of the EMS to relevant vegetation characteristics such as biomass or leaf area index (LAI). Since most studies relating vegetation to biomass or LAI use VI's of NIR and red EM portions, these interactions, specifically should be understood. NIR reflectance of rice is directly related to green biomass. It continues to increase from early tillering, where NIR reflectance is about 15%, to heading where it reaches a maximum of about 50% (Casanova, 1998). Post-heading NIR reflectance decreases to about 33% as rice crop green biomass decreases due to death and loss of leaves (Casanova, 1998). Red reflectance of rice is inversely related to green biomass, where it decreases from about 10% at emergence to 2% at flowering, and then increases to about 18% at maturity due to senescence (Casanova, 1998). Some VI's perform better than others, but in general, these interactions result in better correlations between VI's and biomass in the vegetative and pre-heading stages than in the post-heading stage, where VI's tend to saturate (Casanova, 1998).
The saturation of VI's results in a notable limitation to the usefulness of remote sensing in estimating crop biomass, LAI, and, therefore, yields in rice-based irrigated agricultural systems. This saturation is mainly due to the inability to detect the accumulation of biomass that takes place well after crop canopy cover reaches 100% (Thenkabail et al., 2000). After canopy closure, NIR reflectance continues to increase significantly, while red reflectance may only change slightly, resulting in minimal changes in a VI (Thenkabail et al., 2000). Since NIR is typically the numerator of a VI ratio, and red is the denominator, it takes an inordinately large increase in NIR to increase the overall VI after heading (Thenkabail et al., 2000). This is demonstrated in the saturation of NDVI above LAI's of 2.5 to 3 (Thenkabail et al., 2000) and fraction of intercepted photosynthetically active radiation (
) above 94% (Casanova, 1998). In irrigated agriculture, LAI's often exceed 3, which decreases the accuracy of yield measurements from remote sensing. However, this is somewhat offset by the fact that growth stage information received from remote sensing is often useful and that microwave remote sensing measurements can provide structural measurements after LAI's exceed 3.
The applications of remote sensing relevant to this discussion can be separated into 6 main categories as determined by the bulk of the current rice-based remote sensing literature. These are:
1. crop type identification;
2. crop area measurement;
3. crop yield;
4. crop damage;
5. water use/ moisture availability (ma) mapping; and
6. water use efficiency.
A discussion of each follows.
The most commonly practiced application in remote sensing of agriculture is mapping land cover to identify crop types. This process primarily uses the spectral information provided in the remotely sensed data to discriminate between perceived groupings of vegetative cover on the ground. The spatial (Atkinson and Lewis, 2000) and temporal information included in single date and time series data, respectively, usually play a secondary role, but can also aid in the classification procedure. Discrimination of crops is usually performed with `supervised' or `unsupervised' classifiers. The basic difference between these types of classification is the process by which the spectral characteristics of the different groupings are defined. Common clustering algorithms include maximum likelihood, minimum distance to mean, and parallel piped (Jensen, 1986).
Unsupervised classification relies upon a computer algorithm to define natural groupings of the spectral properties of the pixels in an image. After initial classification, the analyst attempts to assign the perceived groupings a posteriori (after the fact). Problems arising from using this methodology are related and include classification of meaningless groupings, idiosyncratic definition of initial number of classes, and subjectivity involved in combination of similar classes. However, if land cover classes are spectrally well separated, adequate results may be obtained.
Because the unsupervised classification algorithm will automatically output the often arbitrarily defined number of classes input to it, the resulting classes are frequently non-representative or meaningless. That is, the initial classification can be drastically different based on the number of classes the analyst originally requests. Some of the classes generated may have little meaning with respect to reality because they contain more than one functional or `on the ground' class. Others are too specific (i.e., several spectral classes form one functional class) and need to be recombined. Spectral classes that contain more than one functional class are frequently harder to deal with because the analyst will need to reprocess them. The recombination of overly specific classes can also be problematic since the process is often subjective. This subjectivity is regularly expressed by results that are not easily duplicated and non-standardised (often represented by abrupt land cover changes at management boundaries).
The advantage of an unsupervised approach, however, is that no a priori (before the fact) knowledge is needed. This can be beneficial when every crop type in a study site may not be known, or when attempting to discriminate between groups of crop vigour (Barrs and Prathapar, 1994). Unsupervised classification also avoids problems with the user biasing the classification with improper or poorly represented training data, which can be the case in supervised classification.
Supervised classification, which is the most common classification method in agricultural areas, requires a priori knowledge. It relies on the analyst to define the perceived groupings by identifying homogeneous areas, called training sites, from either in situ collection or directly from the image. These training sites are statistically analysed and then used to assign every pixel in the image to the group in which it has been determined a member. Problems may arise when defined training sites are a poor representation of the group's spectra. For example, training data with multiple modes in the training class histogram suggests that there are at least two different types of land cover within the training area. Also, positive spatial autocorrelation exists among pixels that are contiguous or close together (i.e., adjacent pixels have a high probability of having similar brightness values), which can cause a reduction in variance between adjacent pixels (Campbell, 1981). This reduction in variance can make large clumps of contiguous training pixels less representative of a particular cover type over the extent of an entire image and is why several single-pixel training sites located spatially apart from each other can result in better classifications than large clumps of contiguous training pixels (Medhavy et al., 1993, Campbell, 1981). Spatial autocorrelation also means that pixels in a remotely sensed image should not be thought of as entirely discrete features independent of their neighbours, but rather a set of continuos features influenced by their neighbours (Campbell, 1981).
Therefore, care must be taken to collect representative and non-autocorrelated training data. When such care is taken, classification for crop identification can be quite effective. Many authors report classification accuracies exceeding 90% for rice identification and other crops (Aplin et al., 1998, Barbosa et al., 1996, Le Hegarat-Mascle et al., 2000, McCloy et al., 1987, Medhavy et al., 1993, Kurosu et al., 1997, Panigrahy et al., 1999). However, the map user should be aware that not all accuracy assessments are equivalent.
In general, classification accuracy assessment involves sampling classes of the thematic map to summarise the amount of overlap between the classified map and what is present on the ground, although the performance of the classifier can also provide information on map accuracy (Richards, 1996). The `ground truth' information is usually gathered from existing maps, aerial photography, or field surveys using any number of sampling schemes (e.g., simple random sample, stratified random sample, or systematic sample to name a few). However, because of differences in the accuracy of the `ground truth' source and the sampling scheme used to gather the information, a certain amount of ambiguity is present in any classification accuracy summarisation.
Certainly, the resulting accuracy assessment will be influenced by the reliability of the `ground truth' source. Validation data might be less reliable than the classification itself, yet is used as `ground truth' because it has been traditionally the standard by which management decisions have been made. For example, agricultural census' of large areas in rural-based countries may be extremely generalised, yet are often used to `ground truth' a classification of fine spatial scale (e.g., 30 metre) satellite imagery. In this case, the classified map is probably more accurate than the `ground truth' making unclear as to what the discrepancies between the classified map and the `ground truth' are attributed to.
Another source of confusion introduced by classification accuracies is the way in which the accuracies are reported. It is advisable to differentiate between what is loosely known as producer's accuracy and user's accuracy (Story and Congalton, 1986). Producer's accuracy, here, is a strict comparison of the number correctly classified on the map versus the number of that same class determined by the `ground truth' (i.e., per cent correct). This measurement only considers errors of omission (`ground truth' points that were left out in the classification). User's accuracy, on the other hand, considers both errors of omission and commission (i.e., those that were misclassified by being wrongly omitted as well as wrongly included) (Story and Congalton, 1986, Congalton, 1991). User's accuracy is usually lower than producer's accuracy, which is why producer's accuracy is usually reported. This can be potentially problematic when comparing results of classifications between different studies, especially when classification accuracy methodologies are not explained well in publications or reports (which is often the case).
To achieve high classification accuracy, multitemporal (Kurosu et al., 1997, Panigrahy et al., 1997), multi-sensor (Le Hegarat-Mascle et al., 2000, Okamoto and Kawashima, 1999) or GIS (Aplin et al., 1998) data is often integrated into the classification procedure. However, incorporating these data into a classification does not necessarily improve all crop type accuracies. For example, Barbosa et al. (1996) attained higher accuracies for several crop types using single date imagery as opposed to multitemporal imagery. These results depended on both the timing of the single date image as well as the accuracy assessment methodology used (Barbosa et al., 1996). For example, rice crop identification using one single date spring image produced equivalent or higher accuracy than the multitemporal classification (Barbosa et al., 1996). This specific example could be related to the ease by which the water in flooded rice fields can be separated from other areas at the time of image acquisition. In contrast, some authors believe that the early (spring) estimate of rice crops can be improved by use of a later image, which can aid in elimination of permanent water bodies and other irrigated crops from the classification (Barrs and Prathapar, 1996, McCloy et al., 1987). The `law of diminishing returns' is most likely at work in this situation; slight increases in the accuracy of rice crop identification may not be worth the cost of purchasing another remotely sensed image unless the impact of the assessment on management is relatively high or potentially costly (Pax-Lenney and Woodcock, 1997b).
The performances of different sensors for crop identification have been tested over varied geographic areas and crop types. These most commonly include broadband optical (e.g., LANDSAT, SPOT, and AVHRR), and microwave (e.g., ERS, RADARSAT, JERS) used alone or in combination in the form of either single or multiple date imagery. In general, multitemporal microwave imagery results in roughly equal classification accuracies as single date optical imagery when specifically considering rice (i.e., about 90%) (Kurosu et al., 1997) or other crops (Le Hegarat-Mascle et al., 2000). Because of the complementarity of the data, the fusion of both optical and microwave imagery has resulted in higher overall and individual crop classification accuracies (> 95%) than either produce alone (Le Hegarat-Mascle et al., 2000).
Higher crop classification accuracies have also been achieved by the combination of GIS and remotely sensed data (Aplin et al., 1998). This method of classification depends on accurate GIS data and is sensitive to missing boundary lines. By using GIS data, Aplin (1998) achieved greater classification accuracies in `constrained' areas having well-defined boundaries (e.g., agricultural fields and urban areas). Conversely, the inclusion of GIS data decreased classification accuracy in `unconstrained' areas or those areas having poorly defined boundaries (e.g., grassland and bare soil areas) (Aplin et al., 1998). Similarly, others have improved classification accuracy using the contextual or landscape information inherent in the imagery (Moody, 1997, Stuckens et al., 2000). These techniques require no a priori knowledge, and therefore, are more flexible because they do not rely on an updated and accurate GIS layer. However, the user has more control when using GIS data, which can result in very accurate results.
Finally, hyperspectral remote sensing allows for the classification of imagery using many narrow bands along the EMS. These narrow bandwidths permit the use of very specific spectral characteristics for optimising the categorisation of agricultural properties such as biomass and leaf area index (Thenkabail et al., 2000) as well as classification of crops and crop stress (Lelong et al., 1998). The narrow bands in hyperspectral data are often ordered sequentially along a wide bandwidth (e.g., every 10 nm from 470 - 2500 nm) providing continuos spectral information over that EM portion. This continuos nature of hyperspectral data can provide very valuable information that can increase classification accuracy and is why this data source will be frequently used in the future. However, the tremendous spectral information comes at a cost; the datasets are much larger than broadband datasets and thus are harder to maintain in terms of data management and processing. Also, in multi-temporal analysis, accurate atmospheric correction of the hyperspectral data will be more critical than for broadband data.
Crop area measurement is a very common practice in agriculture. Remote sensing is often used for this purpose because of its strengths in regard to spatial extent, temporal density, relative low costs, and potential for rapid assessment of spatial features. Many of the same issues concerning crop type identification also affect crop area measurement from remotely sensed data. This is because crop type identification is a necessary first step to area estimation. In many cases, though, crop type identification is more concerned with classifying all crop types from each other, where area estimation often is concerned with only a few target crops. In either case, these two applications are frequently performed in sequence: first crop identification and then area estimation. There are a few issues that are not exclusively related, but tend to more specifically pertain to crop area estimation, including positional accuracy, mixed pixels and pixel size, and a mismatch between individual and overall accuracies of the results.
Positional accuracy, here, can be defined as the difference in the position of a feature on a map compared to the feature's real world or `true' position. As such, the position of boundary lines on the map, for instance, are most likely not where they are in the real world, but are more accurately represented as a belt or swath around that boundary line on the map. This swath contains the `true boundary line' and has a width that is inversely related to the scale of the source (Van Niel and McVicar, 2000). For example, as the scale of the source gets smaller (area representation gets larger) the width of the swath around the line generally represents a larger distance, making the positional accuracy decrease (Van Niel and McVicar, 2000). In other words, as the swath gets wider, the relative certainty of the position of the `real boundary' gets smaller and the error likely increases. Therefore, when attempting to estimate crop areas accurately, considerable thought should be given to achieving high positional accuracy of crop boundary lines if GIS data are used in conjunction with remote sensing. Ground validation points are extremely important, in this case, to quantify positional accuracy and thus spatial uncertainty. Knowledge of spatial uncertainty is essential in making appropriate managerial decisions on crop area measurements (Van Niel and McVicar, 2000). For example, if crop area estimation is performed in an attempt to monitor year-to-year land use, it is extremely important to know how much potential error is included in the estimate solely due to uncertainties. With this knowledge, the manager has a better idea if the year-to-year difference in area is reliably measured from the remotely sensed data or if it is `absorbed' by the inaccuracies of the dataset itself.
Pixel size of the remotely sensed data also affects positional accuracy of boundary lines in crop area estimates, and therefore should be considered for its appropriateness to a particular application. One prominent issue is the relation of the pixel size to the paddock size (or feature element) being measured (Woodcock and Strahler, 1987, Pax-Lenney and Woodcock, 1997a). Although for large areas AVHRR data is very attractive due to its spatial extent and high repeat cycle, the 1 km pixel size is often disproportionately larger than paddock sizes. This may not be an unsurmountable problem when monitoring areas with many contiguous smaller fields of the same crop type (Quarmby et al., 1993b), but at the same time, must be considered. In an attempt to either compensate for mismatched pixel-to-field sizes or to increase the accuracy of area estimations, the spectral characteristics of impure or `mixed' pixels can be `unmixed' by linear mixture modelling.
In linear mixture modelling, the analyst assumes that any mixed pixel's spectral signature is made up of a combination of pure spectral signatures of all the separate land cover types contained in that pixel in proportion to the area of which that cover type is found in the pixel (van Leeuwen et al., 1997, Maas, 2000, Quarmby et al., 1992). The spectral signatures resulting from a pixel containing only one cover type (pure pixel) are known as end members. Proper unmixing of mixed pixels relies upon the identification of good end member spectra (Quarmby et al., 1992). Different combinations and proportions of the end members are combined to best match the signature of the mixed pixel, supposedly allowing for a better estimate of crop areas. Problems can arise when end members are either not a good representation of a particular land cover class, or when a land cover class' end member is not collected. Even when numerous and representative end members are gathered, results can still be spurious due to confusion caused by mixed pixel spectra being explained by multiple possibilities of end member combinations. Also, green crops can produce very similar end members at certain times of the year, resulting in erroneous unmixing in agricultural systems.
As with identification of crop types, ground validation is also very important in area estimation. Traditional estimates of crop area have come from census data over rather large areas, and are usually rather gross and usually contain little specific spatial context. This means that there is often no explicit locational knowledge contained in the `ground truth'. This leads to very ambiguous results where overall values may be closely related, but the analyst often has no idea what sort of spatial variation exists between the `ground truth' and the estimates from the remotely sensed data, leaving the comparison less meaningful. It is important, then, to make sure that the `ground truth' data is defensibly more reliable than the observation as well as spatially explicit, wherever possible.
Often, there exists in the validation of area estimations, a mismatch between reported overall and individual accuracies. Many authors represent areal accuracy as an overall summation of several paddocks or districts. This tends to reveal extremely accurate results, often exceeding 99% (McCloy et al., 1987, Quarmby et al., 1992, Fang et al., 1998, Okamoto and Kawashima, 1999). However, this can be misleading because there often exists large areal errors in the individual paddocks or districts that were added together to generate this statistic. Unfortunately, this mismatch between overall and individual areal accuracy is rarely discussed, revealing a perceived dichotomy in the results.
The discrepancy between the overall accuracy and individual district- or paddock-level accuracy is due to errors of underestimation and overestimation cancelling each other out (Van Niel and McVicar, 2000). That is, when several areas are added together, the overestimates are often almost exactly offset by the underestimates, resulting in a good overall estimation of area even though individual error is much larger (McCloy et al., 1987, Quarmby et al., 1992, Fang et al., 1998, Okamoto and Kawashima, 1999).
In a study at the CIA in southern NSW, this individual areal error was seen to drop considerably when just two areas were added together, and continued to approach zero as more areas were summed (Van Niel and McVicar, 2000). This phenomenon can be exploited for management purposes, but care must be taken to ensure that proper interpretation of areal accuracy is achieved for specific management situations (Van Niel and McVicar, 2000). That is, it is only appropriate to use the summed or overall areal accuracy when the overall area is the specific concern.
Crop yield forecasts can greatly influence farm-level management decisions, such as fertiliser applications and water delivery, as well as provide a means for farm income assessment. Consequently, individual farmers and district-level land managers show great interest in producing rapid and accurate estimates of crop yield, both locally and regionally. In the past, the standard yield estimation procedure included the analysis of crop cuttings at randomly sampled ground plots during harvest (Murthy et al., 1996), or meteorological regression models using rainfall and past yield data (Karimi and Siddique, 1992). These methods often produce results that are either not timely nor spatially explicit. Though still used, these methods are being replaced by estimation of crop yields using remote sensing because of its ability to produce results quickly and spatially. Using this technology, it was found that spatially meaningful estimates of yield can be made as early as 1 to 3 months prior to harvest (Quarmby et al., 1993b, Rasmussen, 1997), thus impacting management reaction time to yield forecasts.
Positive correlations exist between measurements of LAI (Curran et al., 1992, Tucker, 1979, Nemani and Running, 1989a, McVicar et al., 1996a, McVicar et al., 1996b, McVicar et al., 1996c), plant condition (Sellers, 1985), and VI's. Based on this relationship, and assuming that LAI (or biomass) is related to yield, there are 3 main approaches used to forecast yield. These are: 1) correlate yield with NDVI (Maselli et al., 1992, Smith et al., 1995); 2) correlate
yield with integration under the NDVI curve, denoted ∫NDVI (Benedetti and Rossini, 1993, Quarmby et al., 1993a, Rasmussen, 1997, Rasmussen, 1998a, Rasmussen, 1998b, Pinter Jr. et al., 1981, Honghui et al., 1999); or 3) simulate yield with crop models (sometimes using remotely sensed inputs) (Rosenthal et al., 1998, Inoue et al., 1998, Maas, 1988). For a detailed review of crop yield modelling using remote sensing, see McVicar and Jupp (1998). A brief summary of all three approaches follows.
Both regressions of VI and ∫NDVI studies generally attempt to estimate yield using remotely sensed data alone. A direct comparison of VI response to yield can result in highly correlated regression equations. For example, Harrison et al. (1984) achieved correlation coefficients (r) as high as 0.9 for calrose rice and 0.79 for Inga at the CIA using the greenness ratio of MSS7/MSS5. This study developed the yield model based on imagery acquired when the crop was at booting stage, the time when rice displays green leafy surfaces (Harrison et al., 1984). Promising correlation coefficients were also shown between reflective bands of TM imagery and harvested grain yield (up to 0.92 for band 7) (Tennakoon et al., 1992). These relationships can be used to estimate rice yield at district levels with reasonable success. For example, Patel et al. (1991) was able to estimate district rice yield with an accuracy ranging from 86% to 98%.
Likewise, ∫NDVI studies have produced similar results with relation to rice yield. Rice yields have been estimated within 3% of official values for 3 consecutive years in one study (Quarmby et al., 1993b), while correlation coefficients have been reported as high as 0.93 for millet yield in another (Rasmussen, 1992). However, the ∫NDVI method is highly sensitive to remotely sensed data availability, which can result in gross yield underestimates when portions of the growing season are missing (Quarmby et al., 1993b). Until just recently, the only operational satellite sensor with a repeat cycle sufficient for ∫NDVI studies of crop yield has been NOAA AVHRR (Rasmussen, 1992). The most notable additions to this list are the TERRA (ASTER and MODIS sensors), and Ikonos satellites both launched in 1999 (Table 1). Anticipated sensors could also impact the usefulness of ∫NDVI studies by increasing the number of satellites with relevant repeat cycles. These include Quickbird-1, and Orbview 3 and 4 (Table 1). The definition of the proper time interval for ∫NDVI studies has been disputed, but shown to only be significant when integrated over the reproductive stage (for millet) (Rasmussen, 1992). For rice, where LAI usually exceeds 5, VI's saturate, limiting the usefulness of this approach, and suggesting that different methods might achieve better results. To date, the ∫NDVI method has mostly been used in dryland farming. In an attempt to achieve highly accurate yield estimates, many models have been developed which aim to simulate plant growth. Such plant growth models need to be validated to ensure that they meet a desired level of predictive capability. Plant growth models can be validated by comparing simulated and in situ measurements of vegetation parameters (e.g., biomass or yield). However, plant growth models may also be validated by comparing simulated reflectance with that measured by remote sensing. This can be achieved by inverting either empirically-based relationships, or physically-based radiative transfer models. This means that a plant growth model estimate of cover can be converted, through a radiative transfer model, to an estimate of reflectance, or recalibrated through the estimation of key crop variables like LAI or above ground biomass (Inoue et al., 1998, Fischer et al., 1996, Fischer et al., 1997, Maas, 1988). This recalibration of crop variables links well to simple growth models because they can be directly related to, a key variable in most of these models (Inoue et al., 1998).
Because the radiative transfer model approach can be complicated and necessitates many input variables, the recalibration of simple models is often preferable (Inoue et al., 1998). Recalibration is also, very often, a better method for yield estimation than either using the remotely sensed data as direct input to simulation models, or ∫NDVI because it is not as limited by the availability of remote sensing data (Inoue et al., 1998). The recalibration approach using remote sensing has also been shown to drastically improve non-calibrated simulation models (Inoue et al., 1998).
Of great concern in agricultural systems is the loss of productivity due to crop damage and its negative impact on meeting the increasing demands for food globally (Fox Strand, 2000). Rice damage and subsequent yield reduction can occur for many reasons including, various pathogens, insects, and weeds and often possesses a complex interaction with cropping practices (Ennaffah et al., 1997, Islam and Karim, 1997, Savary et al., 1997). Cropping practices such as differences in water depth and fertiliser applications have been shown to affect crop production (Anbumozhi et al., 1998). Likewise, irrigation with wastewater can cause heavy metal contamination in soils, also resulting in crop damage and yield reduction (Cao and Hu, 2000). Natural variables, such as climate, can also cause damage from either seasonal temperature or precipitation (drought) variations, although irrigated systems are generally sensitive to cold temperature events only. In southern NSW, depth of ponded water is used to regulate plant temperature. The rice panicle is kept under water, where during the night it remains warmer than the surrounding air temperature. Rice leaf blast epidemics may be related to temperature as well, leading to model simulations of yield losses due to current global warming trends (Luo et al., 1998). Crop damage research using remote sensing can be simplified into two different categories: those that measure the source of damage (direct measurement), and those that measure the effect of the source of damage (indirect measurement). In certain situations, the source of crop damage can be measured directly from remote sensing. These studies include assessment of salinisation, weed infestations, and waterlogging.
Saline soils have been shown to have higher reflectance in both the visible and NIR wavelengths than non-saline soils from both ground and satellite radiometric measurements (Rao et al., 1995) allowing for mapping of salinisation (Sharma and Bhargava, 1988, Dwivedi, 1992, Wiegand et al., 1996, Yu-liang, 1996). It should be noted, though, that successful mapping of salinisation results only when salt scalds are clearly visible on the soil surface. This is not the case for all salt affected soils, which means that salinisation is not always easily mapped from remote sensing. Detection of water from dry land is uncomplicated spectrally leading to the classification of potentially waterlogged areas (Sharma and Bhargava, 1988). However, these areas should be integrated with an accurate land use map as permanent water bodies and natural wetlands could quite easily be included in the classification otherwise. Ground-based electromagnetic (EM-31) surveys (Hume et al., 1999) and airborne radiometrics (Cook et al., 1996) have also proven useful for mapping soil attributes.
Weed infestations can also be directly measured with remote sensing if the weeds are prevalent prior to crop emergence (Lamb, 2000). In this case, the living vegetation (weeds) is easily identifiable from the background soil. These infestations are also detectable after crop emergence if their spectra are significantly different from the background soil and crop cover (Lamb et al., 1999). It was found that weed detection requires high spatial resolution and is sensitive to spatial registration and weed density, which can all affect the accuracy of the results (Lamb et al., 1999). This spatial resolution requirement limits the usefulness of most current satellite platforms for weed detection and is why most of these projects make use of airborne systems.
In most cases, however, it is impossible to measure the source of the damage directly using remote sensing. For example, the measurement of ducks or insects in a rice paddock is not feasible, whereas, the effects of these pests can potentially be detected through loss of above ground biomass or yield. In this sense, crop damage is almost always associated with crop yield, and therefore, there is much overlap with the previous section. This means that any of the multiple techniques that have been developed for estimating yield from remotely sensed data, can then be associated with losses due to various forms of crop damage. In specific cases, however, the indirect measurement is not necessarily associated with monitoring yield or biomass losses from a particular event (e.g., an insect infestation), but rather is an attempt to detect possible plant stress (Lelong et al., 1998) over a less discrete period of time.
It is well established that physiological stress in crops is related to incident reflectance from crop canopies and that this altered reflectance can be detected with remote sensing (Knipling, 1970, Thenkabail et al., 1994b). These changes in reflectance are prominent in both the visible and NIR regions of the EMS. An increase in leaf reflectivity in the visible light portion as a response to stress is caused by the metabolic sensitivity of chlorophyll resulting in less efficient absorption of light (increased reflection) (Knipling, 1970). NIR reflectance of leaves is more variable at the onset of stress (Knipling, 1970), but has been shown to increase steadily as stress continues in the form of dehydration (Gausman, 1974). NIR also looses distinctiveness in its spectral signature, especially in the water absorption regions, as the curve flattens (Gausman, 1974). These individual leaf responses are often seen at the crop level as losses in leaf area or foliar density in conjunction with increases in shadow and non-foliar surfaces (Knipling, 1970) if the stress continues long enough. Stresses due to dehydration will not generally be a problem in the irrigated areas of NSW, but will affect the surrounding dry-land agricultural systems. Also, the irrigated crops should have similar physiological/spectral responses to other stresses (e.g., disease). In either case, the major objective of such crop damage detection is the link to management decisions in a timely and cost effective manner in order to operate agricultural systems more efficiently
This is why so much of the current crop damage literature concentrates on precision agriculture. Precision agriculture is a process that heavily relies on observations describing within-paddock variability in order to implement dynamic management decisions that may result in higher crop production (Cook and Bramley, 1998). Another key component of precision agriculture is its reliance on technology such as Global Positioning Systems (GPS) and Variable Application Technology (VAT) (Cook and Bramley, 1998). These technologies, along with remote sensing, make the implementation of the procedure possible.
Critical spatial input to precision agriculture decision support systems include crop yield and soil properties. Several studies have shown how to generate these spatially variable inputs through analysis of ground samples (Bailey et al., 2001, Gandah et al., 2000), however, most prove to be both time and cost consuming, especially over large areas (Viscarra Rossel and McBratney, 1998). These limitations have made remote sensing an attractive alternate source of this spatially variable observational data (Moran et al., 1997, Lamb, 2000).
In general terms, the main benefits to the use of remote sensing in precision agriculture are associated with its strengths in mapping both time and space. Seasonal (temporal) and spatial variability of environmental characteristics, mainly crop and soil conditions (Moran et al., 1997), can be mapped efficiently with remote sensing. These data can then be used to implement timely management actions as the crops develop. Current limitations to the use of remote sensing in precision agriculture are mainly related to technology. Current satellite systems have fixed spectral bands, too coarse spatial resolutions, inadequate repeat cycles, and long delivery times (Moran et al., 1997). Airborne sensors overcome some of the spectral and spatial resolution problems associated with satellites, but these systems generate additional problems with radiometric and geometric correction making it hard to monitor large areas (Moran et al., 1997) as well as adding difficulty and time to image processing. Both systems (satellite and airborne) also tend to accumulate high costs associated with the large time-series of inter annual data that has to be acquired. Moran et al. (1997) discuss in detail the potential role of remote sensing in precision agriculture.
Benefits of precision agriculture not necessarily associated with remote sensing are increased efficiency, reduced risk of environmental hazards, and improved process control (Cook and Bramley, 1998). The limitations include cost and complexity of the system, training of users, and delivery of appropriate input data in a timely manner. As such, the potential usefulness of remote sensing in crop damage management, specifically for implementation in precision agriculture is high. However, the limitations of precision agriculture currently make it non-operational at a regional scale. As seen from various paddock-level studies, the process has merit, but is not yet routinely viable for most farmers or land managers. This is likely to change in the near future as some of these limitations disappear due to advances in technology and research methodologies.
Knowing crop water use, both temporally and spatially, in irrigated areas allows water delivery to match agricultural demands. Crop water use can be determined either by crop specific empirical models or use of process based models. Both of these approaches require access to ground based meteorological data, usually with a daily time step. To perform either modelling approach over large irrigation areas will require access to a network of meteorological stations with a suitable spatial density and extent to characterise the spatial variation in observed meteorological variables.
A commonly used method to estimate crop water use is application of the Food and Agriculture Organisation (FAO) Guidelines for prediction of crop water use (Smith et al., 1991, Doorenbos and Pruitt, 1977, Frere and Popov, 1979). Class A Pan evaporation can be measured if such facilities exist and are properly maintained, or can be estimated from commonly observed meteorological data. The FAO method requires modification of crop coefficients; it does not require remotely sensed data. However, knowledge of different land uses are required to extend this approach spatially. Remote sensing is seen to be a timely and cost effective way to provide these maps to GIS models annually. A derivate of this method have been used by Kirk et al. (1999) to provide estimates of crop water use for farm-level irrigation water use efficiency for irrigation areas in South Australia. The location of paddocks was obtained from a GIS data base of paddock level cropping.
Many other water balance/plant growth models have been developed which require meteorological data, and soil and generic plant parameters to successfully run. These models can range from empirical to process based, however, as with the FAO approach, suitable maps of land use is required to take these estimates into the spatial domain. Given the issue of advection in irrigated environments (Humphreys et al., 1994a) the areal weighted up-scaling approach using GIS strata may introduce large errors in the regional crop water use calculation.
Regional remotely sensed Ts observations can be used in a resistance energy balance model (REBM, Monteith and Unsworth, 1990) to provide estimates of the H. This allows the _E to be estimated as the residual, obtained by rearranging Equ'n 1. REBMs require effective models for the resistances and other terms (including Rn and G). Kustas and Norman (1996) review the use of remote sensing to estimate _E. REBMs describe the fluxes between soil and plant canopies and the atmosphere in terms of `resistances'. The models are `closed' by specified meteorological inputs defined at a reference height above the surface and some assumptions or models defined at or below the earth's surface. To solve a REBM at the time Ts is observed, several surface and near surface meteorological variables (either measured or estimated) are required. These include air temperature (Ta), relative humidity (h) or vapor pressure (ea), solar radiation (Rs) and wind speed (u) at some reference height above the surface and over a time limit of sufficient length for equilibrium to be assumed. The feasibility and effectiveness of providing such data when only standard daily meteorological data are available has been demonstrated (McVicar and Jupp, 1999a). REBMs also need surface information such as albedo and LAI (or percent vegetation cover) which may be obtained either from reflective remotely sensed data or in situ measurements.
During detailed experiments, when specific time-of-day meteorological data are collected, progress in REBM parameter formulation has been made. However, such experiments are limited to short times for small areas (e.g. Kustas et al., 1989, Kustas et al., 1990, Zhang et al., 1995, Raupach et al., 1997, Flerchinger et al., 1998). For operational regional irrigation water supply management, obtaining REBM estimates of fluxes at isolated points is of minimal use. Either model inputs or model outputs need to be spatially interpolated to the entire region of interest. The issue of `interpolate then calculate (IC)' or `calculate then interpolate (CI)' has received attention to spatially estimate moisture deficit (Stein et al., 1991), moisture availability (McVicar and Jupp, 1999b, McVicar and Jupp, 2000), global solar radiation (Bechini et al., 2000), and soil properties (Heuvelink and Pebesma, 1999, Bosma et al., 1994).
To avoid running a REBM at every point in the agricultural system, a number of methods have been developed. The methods are based on the link of the energy balance and water use. For example, leaf temperature can increase when the leaf is green, as stomatal closure to minimise water loss by transpiration results in a decreased _E. At the same time, due to the requirement of the energy balance, H usually increases.
Idso et al. (1981), Jackson (1982) and Jackson et al. (1981, 1983) pioneered methods using daytime Ts for assessing crop health and establishing irrigation scheduling at the field scale. These techniques, culminated in the development of the Crop Water Stress Index (CWSI) which has the form CWSI = 1 - mad, where the subscript d means the daily integral. The Normalised Difference Temperature Index (NDTI) directly maps ma regionally. The high spatial density present in remotely sensed data is exploited by using this data as a covariate to spatially interpolate the NDTI in a CI approach (McVicar and Jupp, 1999b, McVicar and Jupp, 2000).
A number of approaches based on the negative correlation between Ts and NDVI (denoted Ts-NDVI) to provide a measure of environmental stress have developed. Goetz (1997) reported that the negative correlation between Ts and NDVI, observed at several scales (25 m2 to 1.2 km2), was largely due to changes in vegetation cover and soil moisture. For complete canopies, the slope of the Ts versus NDVI plot has been related to canopy resistance (Sellers, 1987, Hope, 1988, Nemani and Running, 1989b). Nemani et al. (1993) found the slope of Ts versus NDVI plot to be negatively correlated to a crop-moisture index. These empirical relationships defined from the slope of Ts-NDVI plots need to be acquired over a range of NDVI and Ts conditions, to allow the `warm edge' to be calculated with any confidence. It also should be noted that the slope of Ts-NDVI plots is empirically related to ma. For two different images, with different meteorological conditions, resulting in different atmospheric resistances, the relationship between the slope of Ts-NDVI plot and ma can be non-unique.
Advances are currently being made that use the Ts versus NDVI plot combined with process based understanding to provide a more mechanistic interpretation of the remotely sensed data. There are two methods currently being put forward. The first, a progression of the slope of the Ts versus NDVI plot approach, describes the data as falling into a `triangle' (Carlson et al., 1990, Carlson et al., 1994, Price, 1990, Gillies and Carslon, 1995, Gillies et al., 1997). The second, the Vegetation Index / Temperature Trapezoid (VITT) promotes the idea of data falling into a trapezoid (Moran et al., 1994, Moran et al., 1996, Yang et al., 1997). The VITT is an evolution of the CWSI. The unifying feature for both the `triangle' and the `trapezoid' is the appearance of the `warm edge', the line between the dry bare soil point and the fully vegetated point (for the trapezoid its the fully vegetated stressed point). For any point, if both the vegetation cover and Ts can be measured, then ma can be empirically defined for that image. The top side of the `trapezoid' may collapse, forming a `triangle', as a function of the remotely sensed spatial scale and the timing of acquisition during the growth cycle.
The main limiting factor for use of thermal remote sensing for operational management of irrigated areas is the repeat characteristics of remotely sensed data with suitable spatial resolution. Another difficultly in many irrigated rice growing areas is the measurement and specification of lateral energy flow (advection) from surrounding non-irrigated areas, which is usually not represented in REBMs. For example, in an advective environment, Dunin (1991, pp 45) states "the discrepancy (between lysimeter and Bowen ratio values of _Ea) reached 80 Wm-2 by noon and persisted in excess of 100 Wm-2 throughout the afternoon". This illustrates the difficulty in estimating _Ea using models that don't include components for lateral flow of energy (as in most REBMs).
With population growth comes a decrease in living space (or global land area per capita) and thus, increased competition for land and water resources (Lund and Iremonger, 2000). These higher demands on water and land, among other things, result in a need for more efficient use of the resources. Assessing and improving Water Use Efficiency (WUE) in agricultural systems, then, will become exceedingly important as the demand for food production on these limited resources continues to increase. Water savings from these systems, in particular, could affect the regional and global water balances as the area of land placed under agriculture is expected to increase considerably in the near future (Lund and Iremonger, 2000).
WUE of the irrigated agricultural lands of southern NSW has become increasingly relevant, as the Snowy Water Inquiry into the environmental issues associated with the corporatisation of the Snowy Mountains Hydro-electric Scheme was initiated in 1998. Ninety-nine percent of the runoff from the Snowy Mountains is diverted inland to generate hydroelectric power, and irrigation water for agriculture (Gale, 1999). The Scheme provides an average of 1200 Gl per year to the Murray River and 1210 Gl per year to the Murrumbidgee (Gale, 1999). This water is used in the production of approximately A$1.5 billion per year worth of irrigated agricultural products (Gale, 1999).
Benefits of the scheme include electricity, increased urban water supply, recreation, dilution of salinity in the Murray River, stock watering, displacement of an estimated 5 million tonnes per year of carbon dioxide emissions to the atmosphere (the amount that would be produced if the same quantity of electricity was generated by coal-fired thermal power stations) (Gale, 1999), as well as providing the water for, among other crops, a 1.5 million tonnes of yield per year rice industry. Environmental impacts from the changes in flow and discharge regimes across these rivers include reduction in channel size, increases in groundwater levels, salinisation, erosion of river banks, changes in wildlife habitat, weed invasion, and a reduction of water quality (Gale, 1999).
The Snowy Water Inquiry was established by the Commonwealth, New South Wales and Victorian governments to determine how to best balance the environmental, electricity and irrigation interests of the region (Gale, 1999). Environmental issues of the inquiry included impacts on rivers, river channels, ecology, the lower Snowy River, and environmental flows (Gale, 1999). Economic issues included impacts on electricity generation, impacts on agriculture, and impacts on recreation and tourism (Gale, 1999). Social issues include the alteration of social values of heritage, culture, and community (Gale, 1999).
The final report delivered on 23 October 1998 presented 23 options for resolution, of which, composite option D was supported. Option D involves increasing the flow of the Snowy River below Jindabyne Dam to 15% of the pre-Scheme mean discharge (Gale, 1999). This option was chosen because it would provide environmental gains at minimal cost to agricultural, hydroelectric, and different government agencies (Gale, 1999). Whether the environmental gains of this option would be significant (Seddon, 1999, Erskine et al., 1999), or cost to agriculture minimal (Watson, 1999, Hoare, 1999) has been questioned. Also, it is questionable whether reducing one environmental problem is worth adding to another as increasing environmental flows in the Snowy River will translate into greater greenhouse emissions from subsequent thermal-powered electricity generation (Seaton, 1999).
The outcome of this inquiry will, no doubt, require irrigation managers to be more efficient with the water that is supplied to them. This increased WUE may occur through a combination of technology and changing management practices, and will most likely necessitate an analysis of WUE at different scales. Regional-scale analysis of WUE would be most helpful in defining gross efficiency relationships and help target areas that could potentially benefit from changes in management practice (McVicar et al., 2000). With this analysis, the baseline WUE could be determined and possibly whether or not plant breeding since the 1980's or 1990's has influenced WUE on a regional level. As such, regional WUE will be the focus of the following discussion.
Stanhill (1986) defined water use efficiency (WUE) both hydrologically and physiologically. Hydrological WUE is the ratio of evapotranspiration to the water potentially available for plant growth. It is expressed as a unitless percentage or fraction (0-1). Physiological WUE is a measure of the amount of plant growth for a given volume of water. The definition of physiological WUE can be defined for different measures of `plant growth' and `volume of water'. Turner (1986) notes that care is needed when defining WUE. For example, in some studies `plant growth' has been measured in units of net biomass (including roots) (Ritchie, 1983, Tanner and Sinclair, 1983, Turner, 1997) or crop yield (Tanner and Sinclair, 1983, Turner, 1997). In various previous studies the units of `volume of water' is either the total transpiration (Tanner and Sinclair, 1983, Turner, 1997, French and Schultz, 1984a, French and Schultz, 1984b), total evapotranspiration (Ritchie, 1983, Tanner and Sinclair, 1983, Turner, 1997), total water input (Sinclair et al., 1984), or the amount of precipitation plus initial soil water at the time of sowing (French and Schultz, 1984a, French and Schultz, 1984b).
Sinclair et al. (1984) introduced different time scales for several definitions of WUE. The temporal scales ranged from an instant, through daily, to a growing season. Intrinsically linked to a range in temporal scale is a range of spatial scale (Table 2), which can extend from a single leaf, through canopy, to field, farm level, and regional assessment. The scales are linked: for example, leaf WUE (in the order of 10's of cm2) will usually be measured over a short time (e.g., a second to daily). On the other hand, farm-level and regional WUE (in the order of 10's to 1000's of km2) will usually be measured over a longer time (e.g., a growing season). To date, there has been very little research focussing on farm level (Kirk et al., 1999, Tuong and Bhuiyan, 1999) or regional assessment (Schuepp et al., 1987, McVicar et al., 2000) of WUE.
Table 2: Different spatial scales of physiological WUE, corresponding temporal scales, and some common units in which they are measured.
Spatial scale |
Temporal scale |
Units |
single leaf (10's of cm2) |
second - day - growth stage |
mgCO2/gH2O or _mol/mmol |
canopy (1 to 10's of m2) |
second - day - growth stage |
mgCO2/gH2O or _mol/mmol |
field (100's of m2 to 10's of ha) |
day - growth stage - season |
mg/gH2O or kg/ha.mm |
farm (1 to 100's of ha) |
growth stage - season - year |
kg/ha.mm or tonne/km2.TL |
region (10's to 1000's of km2) |
season - year |
kg/ha.mm or tonne/km2.TL |
Single-leaf WUE is commonly defined as the net CO2 uptake per unit of transpiration. On a continuous basis - that is, at any instant within a day - it is expressed as the ratio of leaf net photosynthetic rate to leaf transpiration rate, or at the daily time-step it is expressed as the ratio of daytime CO2 uptake to daytime transpiration. Canopy (or community) WUE is commonly defined as the ratio of the net CO2 assimilation of crop canopy to crop canopy transpiration - that is, the ratio of the canopy CO2 flux to the H2O flux for canopy transpiration. Canopy WUE can be expressed continuously and at a daily time-scale, as above, and can also be calculated for specific growth stages. Field WUE can be defined as the ratio of grain yield per unit of water, hence the units would be kg/ha.mm; however the `plant growth' and `water' terms need to be explicitly defined. Regional WUE has similar definitions to field WUE except it applies to a larger area.
There are three main approaches available to assessing WUE regionally:
1. Remote sensing, which can estimate both evapotranspiration and CO2 exchange of large areas at specific times of day, can be used to present regional WUE estimates at specific times (Schuepp et al., 1987). This requires access to much ancillary ground based meteorological data and presents difficulties in extending this from the field sites to regions. There are also difficulties in temporally extending the data from specific times to entire growing seasons;
2. Using remotely sensed based estimates of yield and of _E. However, as has been discussed previously there is difficulty in estimating yield in irrigated environments where LAI is high. Methods exist for providing daily _E maps from specific time-of-day Ts observations, however, reliable estimates of _E are difficult to obtain in highly advective irrigated agricultural environments. Taking isolated daily observations to estimate growing season _E would again require access to ground based meteorological data; and
3. Regional databases of yield, precipitation, irrigation and initial soil water can be developed allowing an `input-output' (Zoebl, 2000) definition of regional WUE (McVicar et al., 2000). `Input' is the water available over the crop growing season and `output' is the yield. This approach is well suited to spatial assessment of regional WUE. Coupling this approach with canopy and field-scale process understanding will allow identification of the regional data bases that are required to develop` a more process constrained regional WUE estimation.
For many of the variables that influence WUE, data are usually not regionally available, and hence they cannot be included in the development of regional WUE estimates. Factors varying both spatially and temporally include:
1. crop varieties, which includes plant breeding (Khush, 1987, Brennan et al., 1994, Brennan et al., 1997) and genetic modifications;
2. soil conditions (Christen and Skehan, 1999, So and Ringrose-Voase, 2000), including soil erosion, sodicity, salinisation, and waterlogging;
3. climate change (Loaiciga et al., 1996), including precipitation patterns and CO2 concentration (Hunsaker et al., 2000);
4. agricultural practices, including the use of fertilisers (Anbumozhi et al., 1998), irrigation management (Pereira, 1999), crop rotation (So and Ringrose-Voase, 2000), planting density, and the use of mulch (Tolk et al., 1999) to reduce soil evaporation.
These interactions are complex, and largely unknown at the regional scale of southern NSW, making absolute measures of WUE difficult. The `input-output' GIS definition of regional WUE, however, is most suited to the analysis of relative trends, both spatially and temporally. Such a GIS would rely on access to data recorded by irrigation companies and the RiceGrower's Cooperative, all who are industry participants in the CRC for sustainable rice production.
