Fig 1.
Stereo camera pair arrangement.
Schematic showing the physical arrangement of the stereo pair of RGB cameras. Camera height above ground level, H, was fixed at 190 cm. The distance between camera apertures was fixed at G = 20 cm. The cameras used were of the make Canon EOS 60D with a focal length of 18 mm and resolution of approximately 0.04 cm per pixel.
Fig 2.
Components in the height estimation pipeline.
(a) Example single camera images of a field plot (Plot 52 of 60—variety: Gregory and treatment: Fertilized), taken at three different time points (from l. to r.: 23/09/2016, 11/10/2016 and 28/10/2016). (b) Corresponding depth maps using the disparity mapping technique described in the paper. (c) Frequency histograms showing leaf pixel height distributions on these days. Note that the vertical axis gives the number of plant pixels in the image at a particular height (horizontal axis). Colour coded vertical arrows indicate the locations of the common percentage threshold (98%) used to determine canopy height. The red arrow indicates the height of the spike distribution present on day 28/10/2016. Note also that a common rectangular region was considered in obtaining the data.
Fig 3.
A whole-of-data comparison of automated height estimation and manual measurement.
Manual measurements of canopy height for the field trial of 60 plots taken on four occasions are compared with the canopy height estimated using depth map estimation from stereo images. The solid line is the line of best fit to the data which, having a slope of 0.9973, is close to the ideal 45° line. Note that manual measurements as reported are the result of averages of several observations sampled across the plot, and therefore are prone to some degree of variation. Data points with different colors are from the four different days on 23/09/2016 (green squares), 27/09/2016 (yellow triangles), 11/10/2016 (blue diamonds) and 14/10/2016 (red circles).
Fig 4.
Graphical comparison of canopy heights for different days.
Graphs of manual measurements (dashed lines) and the proposed depth map estimations (solid lines) of canopy height for the complete data set comprising 60 plots and 4 days of results. The four colors represent the data for the four different days: 23/09/2016 (green), 27/09/2016 (yellow), 11/10/2016 (blue) and 14/10/2016 (red).
Fig 5.
Graphical comparison of canopy heights for a self-consistency study.
Graphs of canopy height for the complete data set as per Fig 4, except that manual measurements and theoretical estimations are separated for clarity of discussion. (a) Plant heights as obtained by manual measurement; (b) plant heights obtained by stereo matching. As above, the four colors represent data from the four imaging days: 23/09/2016 (green), 27/09/2016 (yellow), 11/10/2016 (blue) and 14/10/2016 (red).
Fig 6.
The effect of treatment on height distribution.
Demonstration of the effect of treatment on canopy height distribution as registered by differences in frequency histograms of leaf pixel height for the four days of imaging: 23/09/2016 (green), 27/09/2016 (yellow), 11/10/2016 (blue) and 14/10/2016 (red). The Australian wheat cultivar Drysdale was chosen for study due to its pronounced dependence on canopy fertilization. The solid lines for each respective day refer to averages of the three replicate fertilized Drysdale plots (error bars show variation over the three repeats). The dashed lines for each respective day refer to averages of the three replicate un-fertilized Drysdale plots (error bars show variation over the corresponding three repeats).
Fig 7.
Comparison of canopy development for two varieties: Drysdale and Mace.
Demonstration of the different rates and extents of development of canopy for two Australian cultivars as captured by their leaf pixel height frequency histograms, as a function of time. The colour coding again refer to the four days of imaging: 23/09/2016 (green), 27/09/2016 (yellow), 11/10/2016 (blue) and 14/10/2016 (red). The Australian wheat cultivars Drysdale and Mace were chosen as these exhibited considerable differences in mean canopy height. Only data from fertilized plots is included in the analysis. The solid lines for each respective day depict averages of the three replicate fertilized Drysdale plots (error bars show variation over the three repeats). The dashed lines for each respective day depict averages of the three replicate fertilized Mace plots (error bars show variation over the corresponding three repeats). Note that in this study in contrast to Drysdale (Fig 6), Mace did not show any significant variation with fertilizer.
Fig 8.
The overall growth behaviour of all 10 wheat varieties under two different fertilizer treatments.
Y: fertilized (solid lines) and N: without fertilizer (dashed lines).
Fig 9.
Comparison of plant development as a function of time in terms of medium canopy height under the two fertilizer treatments for all varieties.
Y: fertilized (solid lines) and N: without fertilizer (dashed lines).
Fig 10.
Normalization of leaf height frequency histograms.
Comparison of frequency histograms between un-normalized and normalized canopy height distributions. Normalization, allowing comparisons with different hardware configurations, involves a conversion to leaf pixel distribution (un-normalized, blue line and left vertical axis) to leaf area distribution (normalized, black line and right vertical axis).
Fig 11.
Image rectification and depth map.
Top: A stereo image pair taken of plot 57 on day 26/08/2016. The rectified image from the left camera is shown in (a) and the rectified image from the right camera is shown in (b). Bottom (Fig (c)): The resulting depth map.
Fig 12.
Comparison of height estimation techniques.
A single time stamp (27/09/2016) comparison of different methods of canopy height estimation for the 60 plot set of ten wheat varieties, two treatments and three replicates. The solid red line represents the heights estimated by the proposed automated method based on stereo images and derived depth maps. The dashed line depicts heights obtained by manual measurement, while the dotted line represents canopy heights estimated from images that have been calibrated using a more conventional approach that relies on additional reference data.