Automatic Wheat Ear Counting Using Thermal Imagery
GENERAL DISCUSSION
2. Photosynthetic area of the canopy using image processing systems
Regarding RGB indexes, we have proposed the novel vegetation index u*v*A to estimate the green canopy area associated with the photosynthetic area
sensor size and price less than 100 €) achieved comparable results to those obtained with the Nikon D70 camera (6.1-megapixel resolution, 23.7 x 15.6 mm sensor size and price around 300 €) and a hand-held portable spectroradiometer (GreenSeeker, price around 500 €). This contributes to the move towards the use of low-cost and small sized sensors such as action cameras (GoPro) or mobile phone cameras. On the other hand, the approach of using the color calibration experiment with a color chart (the ColorChecker Passport, in Chapter 5) did not improve the results. At the begining of the experiment, we assumed that this step would improve the information provided by the RGB sensor and therefore we would have better data; nevertheless, the color information did not change, we had very high determination coefficient (R2 ≈ 0.98) between the RGB indexes
derived from the calibrated and un-calibrated images, at least for the acquired database of 4,140 images taken at ground level. Moreover, this calibrated images (when the RGB indexes where used for GY prediction) tented to perform somewhat poorer; this is maybe because we have only used one calibration image (taken at the beginning of each block), and single image doesn’t take into account fast light changes in the field which actually might force the images to be wrongly calibrated from a single color values. Besides the fact that under natural light conditions the RGB cameras have very few color errors related with color calibration (Penczek et al., 2014). Furthermore, the u*v*A index force the hue and chroma values to be dependent to the luminance (L*) and contribute to the reduction of the color perception problems caused by natural light changes; in that way, green pixels from the canopy can be better interpreted due to color
shadows. Therefore, we proposed to take more periodical calibration images intercalated during the plot image acquisition or; in a simpler way, to use the automatic setting of the camera under natural light conditions.
2.1. Phenological stage and data acquisition time for photosynthetic area of the canopy
The photosynthetic area trait can be estimated from seedling growth (GS 10) to late grain filling (GS 99). Although in general, the area indexes (u*v*A, GA, GGA) and Normalized Difference Vegetation Index (NDVI) followed a pattern similar to that of crop growth in both experimental stations; the NGRDI, TGI indexes only followed this pattern in one of them (Chapter 5). This may due to NGRDI and TGI are more related with the nutrient status and crop biomass; and chlorophyll concentration, respectively; than photosynthetic area estimation (Hunt et al., 2014, 2013, 2011, 2005; Jannoura et al., 2015). However, the NGRDI index achieved the highest determination coefficient value (at least for Aranjuez) when it was correlated with grain yield. In addition, it was achieved a date of measurement before u*v*A and GA, which may be due to the fact that this type of index can saturate (Kefauver et al., 2015), but NGRDI is not as prone to saturation (Elazab et al., 2016). This could be an important factor to take into account to improve the performance RGB indexes. Furthermore, the best performance of the RGB indexes (best correlation with grain yield) were observed when the canopy color started to shift from green to yellow, which correspond to the second half of grain filling (GS 75-79) under support irrigation and late
compared to RGB indexes, the use of modified color-infrared (CIR) cameras could be useful. A conventional RGB camera can be used to build a CIR camera removing the internal “hot mirror filter” to enable recoding the near infrared (NIR) information (Lehmann et al., 2017). This modification allows to acquired infrared data with much higher resolution per measurement than the individual points acquired by GreenSeeker device. Besides that, the two-dimensional information would still be available for image processing tasks avoiding parallax issues of multispectral (multi-lens) cameras (Jhan et al., 2017).
2.2. Photosynthetic area of the canopy trait and grain yield
The photosynthetic area of the canopy in field conditions, assessed by RGB-derived indices, has shown to be highly correlated with GY regardless the growing conditions (Chapter 5). We have compared Green Area (GA), Greener Area (GGA), Normalized Green Red Difference Index (NGRDI) and Triangular Greenness Index (TGI) (Casadesús et al., 2007; Hunt et al., 2014, 2013, 2011, 2005) with a novel photosynthetic area index (u*v*A) (Fernandez-Gallego et al., 2019c) based on the CIE L*u*v* color space (Robertson, 1977). Even though, using previous RGB indexes of the literature (GA, GGA, NGRDI, TGI) and the novel vegetation index (u*v*A) we have achieved almost the same or better relationship with GY; the best performance of the RGB indexes were achieved at the second half of grain filling (GS 75-79, R2 ≈ 0.6) under support irrigation and
late planting conditions, and heading (GS 55-57, R2 ≈ 0.5) under rainfed
grain yield, under normal planting conditions (support irrigation and rainfed); for several indexes measured at early stages, the value of the h2 x r
A2 product
(Falconer and Mackay, 1996) was higher than the h2 of the grain yield. This shows
that under normal planting conditions the indirect selection based on RGB indexes performs better than the direct selection based on the yield harvested (Richards et al., 2002). In the case of late planting condition, in several indexes and particularly at early stages, the value of the h2 x r
A2 product was lower than
the h2 of the grain yield. This may due to the fact that high temperature
accelerates the plant growth thus preventing the appearance of genotypic differences (Stone and Nicolas, 1995). Thus, under normal planting conditions at early stages, indirect selection may still have provided a higher throughput via measurement of a specific RGB vegetation index and the benefit of its lower cost, rather than having to wait until maturity to harvest the crop. The early stages seem to be the best phenological stages for saving in cost and time and increase the selection intensity. The cost per breeding plot of the whole harvesting process (until getting the yields) may go easily beyond € 15, plus the cost of maintaining the crop until maturity. By contrast the evaluation at ground level done by hand by walking through the plots takes a fraction of minute, which means that one worker provided with a camera may even record thousands of plots in a day (Fernandez-Gallego et al., 2019c).
3. Equipment considerations and future works