Agronomic research suggests Australian farmers may increase wheat yields by 13 to 47% and reduce risk of crop failures by sowing crops earlier than is currently practiced. This study investigated the potential to use remote sensing to assist in evaluating changes in farmer sowing date practices in response to research recommendations, using study paddocks in the southern Australian cropping region. Vegetation Indices (VIs) derived from Landsat imagery were utilised by fitting a 5 parameter logistic curve (5PL) to the VIs from days 50 to 250 of each year, encompassing the crop green-up period. This method was applied to 10 co-located paddocks with sowing time records spanning years 1988 to 2014, and 17 dispersed paddocks with two years of sowing history. The method shows promise in detecting crop green-up, but only in years and locations where cloud cover does not prevent successful image capture at key crop growing times.
2.1
Paddock management records were obtained for 10 paddocks in central Victoria (‘Merriwa’; 143.738 °E, 35.829 °S) (Figure 1), with sowing dates and crop type records covering the period from 1988 to 2013. Boundaries of these paddocks were digitised to exclude patches of perennial vegetation, dams etc. These paddocks lie within an area of 4.4 x 1.4 km, with paddock areas ranging from 9.9 ha to 87.6 ha (mean 52.4 ha). To apply the analysis over a broad geographic extent, sowing dates and crop types in 2013 and 2014 were obtained for 17 paddocks from 13 farms distributed across Western Australia, South Australia and Victoria (‘NPS’ paddocks; Figure 1).
Except perhaps by using Synthetic Aperture Radar (SAR), or extremely high-resolution satellite imagery, detecting actual sowing operations or the disturbance of paddock soil surfaces that occur during sowing, seems unlikely. Furthermore these options are currently infeasible for assessment of trends due to lack of historical imagery with extensive coverage over the grain cropping regions of Australia. To monitor ‘sowing dates’ from space requires detecting land cover or surface change, from which an estimate of crop sowing date might be obtained. The most apparent change that might be detected is the early green-up of paddocks that occurs as crops emerge after sowing, using this as a surrogate for sowing date. To assess year-to-year trends in sowing dates and/or timing of paddock green-up requires selecting remote sensing data sources which:
capture the entire Australian broad acre cropping region have spatial resolution (pixel size) smaller than the typical land management unit (i.e. paddocks ranging from tens to a few hundred of hectares in area) capture spectral bands suitable for vegetation monitoring, including visible and NIR. provide regular revisits at intervals suitable for capturing the timing of green-up provide archival data to enable trends to be analysed over several years
The imagery sources that appear most suitable are MODIS (250 m pixels, daily revisit) and Landsat (25 m pixels, revisits every 8 to 16 days). The spatial resolution of Landsat being finer than the typical land management unit therefore appears to be the most suitable candidate. Three different Landsat satellites provide varied coverage frequency over the period of interest from 1988 to 2014 (Table 1).
The Australian Geoscience Data Cube (AGDC) is a collaboration between Geoscience Australia, CSIRO and the
National Computational Infrastructure (NCI) developed to improve the accessibility and usability of national
datasets. AGDC is an approach to storing, organising and analysing gridded geospatial information to enable
timeseries analysis to be undertaken efficiently. The AGDC hosts a geo-referenced, calibrated and standardised
timeseries of Landsat imagery from 1987 to the present, which can be analysed as spatio-temporal ‘stacks’ covering the
whole Australian continent. These images are corrected to ARG25
Many spectral indices have been applied to vegetation sensing using Landsat bands
3 Both Landsat 7 and Landsat 8 were being acquired by Geoscience Australia during this period.
recommended for use with low vegetation densities as is this case when sensing emerging crops. gNDVI has been
shown to be more sensitive to chlorophyll than the conventional red NDVI
Mathematical growth functions such as the generalised logistic function
( − ) (1 + ( ) ) (1)
To convert the discrete VI data to a crop growth curve calculated from the time-series data extracted for each paddock, a 5PL function was fitted to the VIs. Fitting was performed using a Python scipy.optimize.curve_fit routine which applies a non-linear least squares method (Levenberg-Marquardt algorithm) to fit a function to a data series. Successful curve fitting by the scipy.curve_fit routine benefits from ‘initial guess’ starting parameters, which can reduce time taken to find an optimal fit, and also improve the likelihood of finding a solution or fit for each paddock-year. Initial guess parameters were refined through iterative testing carried out on a subset of data using Python code which automatically refined the input parameters over the test areas for each different VI.
Cereal crops in south eastern Australia are not generally growing before day 100 of each year, and maximum greenness reached before day 250. During initial testing it was apparent that better fitting was achieved if a longer summer period (with a low crop cover baseline) was included, so the 5PL function was fitted to days 50 to 250 of each year. Curves cannot be fitted where there are fewer samples than parameters, so paddock-years with less than 5 Landsat data samples during this period were automatically excluded. 2.6
For crops to grow requires both sowing and adequate rainfall for growth. Adequate rainfall may occur before or
after sowing occurs. A general rule was used to determine the day of the year (DOY) of ‘opening-rains’
To assess the effectiveness of the different VIs, we compared the recorded sowing dates and the expected greenup dates with the VI estimated green-up dates. The estimated green-up DOY was determined using a fixed threshold of 10% above the baseline level for the ARG25 indices, and 300% for the PV index. Dates estimated from the fitted 5PL functions were excluded if they fell outside the range DOY 50 to DOY 250. 3 3.1
The retrieve_aoi_time_series.py tool took approximately 2 hours to extract each paddock time-series datasets on the Raijin supercomputer at NCI. The extracted time-series data covered 26 years from 1988 to 2014. The number of suitable quality Landsat images available between DOY 50 and 250 for the Merriwa paddocks varies due to cloud cover and the number of operational satellites (Table 1). This is reflected in Figure 2 showing that prior to 2000 between 3 and 10 images were available per year, except for 1998 with zero samples. From 2000 to 2014 between 9 and 28 samples were available per year in the AGDC for these paddocks. The number and distribution of Landsat samples over the key green-up period has important implications for the successful fitting of the 5PL function. Cloud cover reduces the number of observations, and gaps in the data at during the key green-up period means the curve-fitting cannot effectively capture the time at which green-up occurs (Figure 3).
Table 2 demonstrates the range of suitability of the different VIs tested, and highlights that estimation of the expected green-up date is more successful than estimation of an actual sowing date. However, none of the VIs performed well when all paddock-years of valid fitted 5PL functions were included. By restricting analysis to valid fits for paddock-years with more than 15 samples during DOY 50 to 250, the PV performed best, especially when compared against the expected green-up DOY rather than the sowing date. This number of fits reduced successful analysis to only 7 years (2004-7, 11, 13, 14). When examining valid fits for paddock-years with at least 20 samples between DOY 50 to 250, all VIs performed better but PV and OSAVI performed especially well. However, this reduced analysis to only 4 separate years (2004, 2005, 2007, 2013). Again, the fits of estimated green-up DOY against expected green-up DOY were better than against actual sowing DOY.
The Merriwa paddocks provided a long time-series, but with co-located paddocks. The NPS paddocks were therefore included to test the methods using geographically distributed locations. The PV index which gave the best fit against the Merriwa paddocks, was fitted against recorded sowing dates. Using all fitted paddock-years (17 paddock-years) gave a poor result (R² = 0.02), selecting only those with more than 15 samples (14 paddock years) achieved an R² of 0.03, and with more than 20 samples an R² of 0.55 but only captured 6 of 17 paddock years. 3.3
Sowing of Merriwa Pastoral paddocks has started an average of 0.8 days earlier each year (1988 to 2013). The end of sowing has occurred 0.3 days earlier each year, suggesting that the duration of sowing has also increased over this time by approximately 0.5 days per year. Opening rains on the other hand, appear to be occurring 0.56 days later per year over that period.
In most cases the trends in VI estimated green-up dates were negative (results not shown), suggesting the VI method could detect crop green-up occurring earlier each year, in line with the recorded sowing dates and expected green-up dates. However, these comparisons resulted in low R² values and the slope of trends do not consistently match. This suggests that the 5PL estimates are not an accurate predictor of trends in actual sowing dates. 4
We have evaluated a method to use Landsat imagery to estimate crop green-up by fitting a logistic growth
function to VIs. This investigation has included a group of co-located paddocks with long history of crop
management, as well as several widely distributed paddocks with shorter sowing records. Determining sowing date
on the basis of crop emergence or green-up to is complicated by the reliance of crop growth upon early season
rainfall which can be spatially variable. Furthermore, the practice of dry-sowing is recommended as another way of
increasing crop yields
The predictive value of VIs is dependent on an adequate number of remotely sensed samples (i.e. cloud free
satellite image coverage) during the critical crop green-up period. Landsat revisit intervals of 8 to 16 days means
this has been an important constraint in this study. More frequent satellite acquisitions at the key time would
increase the likelihood of cloud free images around the time of early crop growth, and would improve detection of
crop green-up. The incorporation of satellite imagery from other sources such as MODIS, Sentinel 2 might partially
alleviate this issue. MODIS has been used to map crop timing at global scale
This application of remote sensing to detecting crop green-up has also tested the sensitivity of remote sensing at
the resolution of sources such as Landsat to detect very low levels of vegetation cover (or LAI). There are a number
of factors which affect the sensitivity of remotely sensed data to crop vegetation, which include leaf optical
properties, atmospheric perturbations, solar zenith angle and leaf inclination angle
Scaling up this assessment of crop green-up to national scale would require access to mapped paddock boundaries and assessment of which crops have been grown each year. For most of the grain cropping zone, paddock boundaries and crop types are not widely available. Additionally, the high computational requirements and long times taken to perform this analysis are also a constraint on scaling up to national coverage. The input of specialised programming expertise would be needed to improve performance of the analysis across larger extents.
The ability of this approach to monitor trends in actual sowing dates practiced by farmers appears limited to years and locations where cloud cover does not prevent successful image capture at key crop growing times, and the application of the method is unlikely to deliver reliable and consistent results across broad areas. However, the method applied here does shows promise in detecting crop green-up, and could potentially be applied to other phenological studies. With access to additional compatible remotely sensed imagery, the method could be used to assess historic and ongoing changes in timing of crop emergence, which is a function of both farmer management, and climatic influences such as timing of opening rains. Furthermore, this study is an early demonstration of the value of the AGDC infrastructure in enabling access to national coverage of time-series remotely sensed data.
This work was funded by the GRDC project CSP00178 and used the National Computational Infrastructure National Facility in Canberra, Australia, which is supported by the Australian Commonwealth Government. Landsat data were accessed through the Australian Geoscience Data Cube project which is a collaboration between Geoscience Australia, CSIRO and NCI. Merriwa Pastoral and GRDC National Paddock Initiative (BWD00025) contributed paddock data. Dr Rob Woodcock (CSIRO) provided helpful assistance regarding AGDC data analysis.