Determining broadacre crop area estimates through the use of multi-temporal modis satellite imagery for major Australian winter crops

Determining broadacre crop area estimates through the use of multi-temporal modis satellite imagery for major Australian winter crops

[Abstract]: Since early settlement, agriculture has been one of the main industries contributing to the livelihoods of most rural communities in Australia. The wheat grain industry is Australia’s second largest agricultural export commodity, with an average value of $3.5 billion per annum. Climate variability and change, higher input costs, and world commodity markets have put increased pressure on the sustainability of the grain industry. This has lead to an increasing demand for accurate, objective and near real-time crop production information by industry. To generate such production estimates, it is essential to determine crop area planted at the desired spatial and temporal scales. However, such information at regional scale is currently not available in Australia.

The aim of this study was to determine broadacre crop area estimates through the use of multi-temporal satellite imagery for major Australian winter crops. Specifically, the objectives were to: (i) assess the ability of a range of approaches to using multi-temporal Moderate Resolution Imaging Spectroradiometer (MODIS) imagery to estimate total end-of-season winter crop area; (ii) determine the discriminative ability of such remote sensing approaches in estimating planted area for wheat, barley and chickpea within a specific cropping season; (iii) develop and evaluate the methodology for determining the predictability of crop area estimates well before harvest; and (iv) validate the ability of multi-temporal MODIS approaches to determine the pre-harvest and end-of-season winter crop area estimates for different seasons and regions.

MODIS enhanced vegetation index (EVI) was used as a surrogate measure for crop canopy health and architecture, for two contiguous shires in the Darling Downs region of Queensland, Australia. Multi-temporal approaches comprising principal component analysis (PCA), harmonic analysis of time series (HANTS), multi-date MODIS EVI during the crop growth period (MEVI), and two curve fitting procedures (CF1, CF2) were derived and applied. These approaches were validated against the traditional single-date approach. Early-season crop area estimates were derived through the development and application of a metric, i.e. accumulation of consecutive 16-day EVI values greater than or equal to 500, at different periods before flowering. Using ground truth data, image classification was conducted by applying supervised (maximum likelihood) and unsupervised (K-means) classification algorithms. The percent correctly classified and kappa coefficient statistics from the error matrix were used to assess pixel-scale accuracy, while shire-scale accuracy was determined using the percent error (PE) statistic. A simple linear regression of actual shire-scale data against predicted data was used to assess accuracy across regions and seasons. Actual shire-scale data was acquired from government statistical reports for the period 2000, 2001, 2003 and 2004 for the Darling Downs, and 2005 and 2006 for the entire Queensland cropping region.

Results for 2003 and 2004 showed that multi-temporal HANTS, MEVI, CF1, CF2 and PCA methods achieved high overall accuracies ranging from 85% to 97% to discriminate between crops and non-crops. The accuracies for discriminating between specific crops at pixel scale were less, but still moderate, especially for wheat and barley (lowest at 57%). The HANTS approach had the smallest mean absolute percent error of 27% at shire-scale compared to other multi-temporal approaches. For early-season prediction, the 16-day EVI values greater than or equal to 500 metric showed high accuracy (94% to 98%) at a pixel scale and high R2 (0.96) for predicting total winter crop area planted.

The rigour of the HANTS and the 16-day EVI values greater than or equal to 500 approaches was assessed when extrapolating over the entire Queensland cropping region for the 2005 and 2006 season. The combined early-season estimate of July and August produced high accuracy at pixel and regional scales with percent error of 8.6% and 26% below the industry estimates for 2005 and 2006 season, respectively. These satellite-derived crop area estimates were available at least four months before harvest, and deemed that such information will be highly sought after by industry in managing their risk. In discriminating among crops at pixel and regional scale, the HANTS approach showed high accuracy. Specific area estimates for wheat, barley and chickpea were, respectively, 9.9%, -5.2% and 10.9% (for 2005) and -2.8%, -78% and 64% (for 2006). Closer investigation suggested that the higher error in 2006 area estimates for barley and chickpea has emanated from the industry figures collected by the government.

Area estimates of total winter crop, wheat, barley and chickpea resulted in R2 values of 0.92, 0.89, 0.82 and 0.52, when contrasted against the actual shire-scale data. A significantly high R2 (0.87) was achieved for total winter crop area estimates in Augusts across all shires for the 2006 season. Furthermore, the HANTS approach showed high accuracy in discriminating cropping area from non-cropping area and highlighted the need for accurate and up-to-date land use maps.

This thesis concluded that time-series MODIS EVI imagery can be applied successfully to firstly, determine end-of-season crop area estimates at shire scale. Secondly, capturing canopy green-up through a novel metric (i.e. 16-day EVI values greater than or equal to 500) can be utilised effectively to determine early-season crop area estimates well before harvest. Finally, the extrapolability of these approaches to determine total and specific winter crop area estimates showed good utility across larger areas and seasons. Hence, it is envisaged that this technology is transferable to different regions across Australia. The utility of the remote sensing techniques developed in this study will depend on the risk agri-industry operates at within their decision and operating regimes. Trade-off between risk and value will depend on the accuracy and timing of the disseminated crop production forecast.