Edited by: Susana Araújo, ITQB-Universidade Nova de Lisboa, Portugal
Reviewed by: Cristina Cruz, University of Lisbon, Portugal; Jan F. Humplík, Palacký University Olomouc, Czech Republic; Jean-Luc Regnard, Montpellier SupAgro, France
*Correspondence: José L. Araus
This article was submitted to Crop Science and Horticulture, a section of the journal Frontiers in Plant Science
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Maize crop production is constrained worldwide by nitrogen (N) availability and particularly in poor tropical and subtropical soils. The development of affordable high-throughput crop monitoring and phenotyping techniques is key to improving maize cultivation under low-N fertilization. In this study several vegetation indices (VIs) derived from Red-Green-Blue (RGB) digital images at the leaf and canopy levels are proposed as low-cost tools for plant breeding and fertilization management. They were compared with the performance of the normalized difference vegetation index (NDVI) measured at ground level and from an aerial platform, as well as with leaf chlorophyll content (LCC) and other leaf composition and structural parameters at flowering stage. A set of 10 hybrids grown under five different nitrogen regimes and adequate water conditions were tested at the CIMMYT station of Harare (Zimbabwe). Grain yield and leaf N concentration across N fertilization levels were strongly predicted by most of these RGB indices (with
Low soil fertility, alongside drought and heat, is a major stress factor limiting crop productivity on a world scale (Stewart et al.,
Maize is the second most cultivated cereal worldwide and the most commonly cultivated cereal in Africa in terms of land area and production (FAO,
In this sense, two strategies are considered paramount for crop scientists: (i) breeding to improve varieties toward higher nutrient use efficiency and tolerance to nutrient-deficiency (ii) and appropriate fertilization management (Wezel et al.,
Remote proximal sensing technologies are being used currently for precise management of crops, whereas its potential application for field high throughput phenotyping has gathered increasing interest in recent years (Araus and Cairns,
As a low-cost alternative, vegetation indices derived from Red-Green-Blue (RGB) cameras have been employed for remote sensing assessment in field conditions, providing a wide-range of phenomic data about genotypic performance under different stress conditions and species, including water stress and foliar diseases in bread wheat, durum wheat and tritordeum and triticale (Casadesus et al.,
Information derived from plant samples may also be relevant for crop monitoring and phenotyping (Araus and Cairns,
The main goal of this study is to develop affordable easy-to-use new phenotyping tools that increase selection efficiency for grain yield and leaf N concentration under different N fertilization conditions in maize. To accomplish this objective, we compared the accuracy of field-spectroradiometer data vs. RGB-derived vegetation indices assessing GY and leaf N concentration in a set of ten maize hybrids grown in the field under five N-fertilizer levels. Firstly, we assessed the performance of these parameters for all the N-treatments together, and subsequently we dissected the correlations within each N-level for further discussion of phenotyping. Additionally, simple regression models were made for GY prediction and these models were tested and validated against the experimental yield of another trial. The performance of the leaf parameters N/LA, C/N, SLA, and δ13C and δ15N were also studied with the aim of relating these structural and compositional leaf properties with crop performance and phenotyping data. All RGB and UAV imagery were obtained at flowering stage in order to integrate the differences in crop performance from plant emergence to flowering stage, when the number of kernels per ear is determined.
Field trials were carried out at the Southern Africa regional station of CIMMYT (International Maize and Wheat Improvement Center) located in Harare (17°43′32″S, 31°00′59″E) where two field experiments were studied. Before sowing, soil pH, total soluble salts (TSS), nitrogen as nitrate (
Ten maize hybrids were sown, three of them were commercial hybrids (PAN7M-81, SC635, SC537) and the other seven were maize hybrids developed at CIMMYT (TH11894, TH127591, TH127053, TH127618, TH13466, CZHH1155, TH127004). These maize hybrids cover a big range of agronomical sensitivity to low nitrogen conditions. A split-plot arrangement in a randomized block design was set up and five nitrogen fertilization levels (0, 10, 20, 80, and 160 kg·ha−1 NH4NO3) were applied in both trials. Two and three replicates were set for the first and second trials, with 100 and 150 being the respective number of plots in each trial (trials S and P, respectively). A two-row border was sown between fertilization treatments and on the edges of the trial to prevent spatial variability.
Seeds were sown during the wet season, on December 23th 2013, in two rows per plot; rows were 4 m long and 75 cm apart (6 m2/plot), with 17 planting points per row and 25 cm between plants within a row. All trials were homogeneously fertilized with 400 kg·ha−1 of super-phosphate and potassium oxide fertilizer (P2O5 14% and K2O 7%). Weather conditions throughout growing season were recorded with a weather station. The mean temperature was 18.9°C, mean humidity 81.2 and total rainfall during the crop period was 563.1 mm, therefore, preventing the water deficit in these rainfed conditions.
The trials were harvested on May 20th 2014. The central 3.5 m of each row was harvested discarding 2 plants at each end, thus the collected weight corresponded to 5.25 m2 (0.75 m apart × 2 rows × 3.5 m long). The cobs were threshed and the grains dried until they reached around 12% moisture, and then the grain from each plot was weighed. Grain yield (GY, Mg·ha−1) was calculated as follows: (X kg plot−1 × 10)/5.25 where X is the grain weight per plot.
The normalized difference vegetation index (NDVI) was calculated using the equation:
The NDVI of individual plots at ground level (NDVIground) was determined with a ground-based portable spectroradiometer with an active sensor (GreenSeeker handheld crop sensor, Trimble, USA). This equipment uses the spectral wavelengths 650–670 nm as the red band and 765–795 nm as the near infrared. The distance between the sensor and the plots was kept constant using a ladder, around 0.5–0.6 m above and perpendicular to the canopy. The whole areas of the two trials were measured from 12 to 14 h on March 3rd and 4th, 2014.
The aerial NDVI index (NDVIaerial) was obtained using a UAV-based remote sensing platform developed by Airelectronics (Montegancedo campus, Spain) in collaboration with the Crop Breeding Institute-Zimbabwe, CIMMYT, QuantaLab at the Institute for Sustainable Agriculture (IAS-CSIC, Spain) and the University of Barcelona. This aerial platform was equipped with a multispectral camera (ADC-Lite, Tetracam, Inc., Chatsworth, CA, US), which provides spectral images on the green, the red and the near-infrared bands, with a final ground resolution of 10 cm per pixel when flying at an object distance of 150 m. These bands are approximately equal to the Landsat Thematic Mapper (TM) bands TM2, TM3, and TM4, respectively, so that the spectral wavelengths from 630 to 690 nm represent the red band and 760 to 920 nm the near infrared band. The flight was conducted at an altitude of 150 m at midday on a sunny day when crops were around the flowering stage. The collected images covered 220 out of the total 250 plots, completely covering the block S trial (100 plots) and partially covering block P (120 of the total of 150 plots). Aerial images were subsequently corrected and calibrated with ImapQ (QuantaLab-IAS-CSIC, Cordoba, Spain) which converts images to radiance. Mosaicking and rectifying processes were applied with Autopano (Kolor SARL, Francin, France) by applying the image stitching technique (SIFT algorithm) in addition to a manual orthorectification from several checkpoints selected. NDVI values were finally extracted from the images using ENVI software (Exelis Visual Information Solutions, Boulder, Colorado, USA).
Vegetation indices derived from red-green-blue (RGB) images were evaluated at the plot and the single leaf level (RGBcanopy and RGBleaf indices, respectively; Figure
Subsequently, images were analyzed with the open source Breedpix 0.2 software (Casadesus et al.,
The leaf portions in the RGBleaf indices were also used the subsequent measures. Firstly, immediately before being scanned, a handheld spectroradiometer developed for leaf chlorophyll measurements (Minolta SPAD-502, Spectrum Technologies Inc, Plainfield, IL, USA) was used to measure the index related to leaf chlorophyll content (LCC). Four measurements were made for each leaf segment. Secondly, the leaves were oven dried at 70°C for 24 h and the dry weight was measured. Then the specific leaf area (SLA) was calculated using the equation
Finally, dry leaves were ground to a fine powder and 0.7–0.9 mg of leaf dry matter from each plot was weighed and sealed into tin capsules. Stable carbon (13C/12C) and nitrogen (15N/14N) isotope ratios as well as the leaf N and C concentrations (%) were measured using an elemental analyser (Flash 1112 EA; Thermo Finnigan, Bremen, Germany) coupled with an isotope ratio mass spectrometer (Delta C IRMS, Thermo Finnigan) operating in a continuous flow mode. Samples were loaded into a sampler and analyzed. Measurements were conducted at the Scientific Facilities of the University of Barcelona. Isotopic values were expressed as a composition notation (δ) as follow:
where “sample” refers to plant material and “standard” to international secondary standards of known 13C/12C ratios (IAEA CH7 polyethylene foil, IAEA CH6 sucrose and USGS 40 L-glutamic acid) calibrated against Vienna Pee Dee Belemnite calcium carbonate with an analytical precision (standard deviation) of 0.15%0. The same δ notation was used for the 15N/14N ratio expression but with the standard referring to air. For nitrogen, international isotope secondary standards IAEA N1, IAEA N2, IAEA NO3, and USGS 40 were used with a precision of 0.3%0. Further, the C/N ratio was obtained from these analyses and total nitrogen concentration per unit leaf area (N/LA) was calculated with the formula:
where
Statistical analyses were conducted using SPSS 21 (IBM SPSS Statistics 21, Inc., Chicago, IL, USA). Multiple variance analyses, the multiple comparison Duncan
Significant differences in GY between genotypes and nitrogen-fertilization levels were observed in this study (Table
GY | < 0.001 | < 0.001 | 0.884 | ||
NDVIaerial | 0.509 | < 0.001 | 0.745 | ||
NDVIground | < 0.001 | < 0.001 | 0.500 | ||
hue | 0.032 | < 0.001 | 0.997 | ||
a* | 0.006 | 0.125 | 0.727 | ||
b* | 0.590 | < 0.001 | 0.997 | ||
u* | 0.012 | < 0.001 | 0.801 | ||
v* | 0.567 | < 0.001 | 0.993 | ||
GA | 0.002 | < 0.001 | 0.708 | ||
GGA | < 0.001 | < 0.001 | 0.564 | ||
hue | 0.412 | < 0.001 | 0.817 | ||
a* | 0.364 | < 0.001 | 0.882 | ||
b* | 0.580 | < 0.001 | 0.999 | ||
u* | 0.310 | < 0.001 | 0.687 | ||
v* | 0.787 | < 0.001 | 0.999 | ||
GA | 0.042 | < 0.001 | 0.604 | ||
GGA | 0.289 | < 0.001 | 0.225 | ||
LCC | 0.939 | < 0.001 | 0.973 | ||
Leaf %N | 0.026 | < 0.001 | 0.999 | ||
N/LA | 0.347 | < 0.001 | 0.972 | ||
C/N | 0.120 | < 0.001 | 1.000 | ||
δ15N | 0.375 | < 0.001 | 1.000 | ||
δ13C | < 0.001 | < 0.001 | 0.937 | ||
SLA | 0.822 | 0.004 | 0.89 |
PAN 7M-81 | 5.03 b | 2.203 ab | 111.24 ab | 20.23 a | 21.07 b | 4.53 b | –11.7 ab | 41.53 a |
TH11894 | 5.12 b | 2.196 ab | 101.75 a | 21.48 a | 20.77 b | 3.78 ab | –11.49 cd | 42.02 a |
TH127591 | 5.14 b | 2.312 abc | 117.87 ab | 20.41 a | 19.75 ab | 4.04 ab | –11.79 a | 43.06 a |
TH127053 | 5.7 bc | 2.32 abc | 118.29 ab | 19.56 a | 19.75 ab | 4.18 ab | –11.55 bcd | 41.13 a |
SC635 | 3.52 a | 2.506 bc | 125.31 ab | 19.98 a | 18.38 ab | 3.66 ab | –11.77 a | 41.84 a |
TH127618 | 5.38 b | 2.621 c | 130.12 ab | 20.54 a | 17.27 a | 3.64 ab | –11.63 abc | 43.56 a |
TH13466 | 5.75 bc | 2.4 abc | 117.62 ab | 20.91 a | 18.96 ab | 3.76 ab | –11.45 d | 42.57 a |
CZH1155 | 4.63 b | 2.526 bc | 133.84 b | 18.9 a | 18.48 ab | 3.19 a | –11.64 abc | 42.28 a |
TH127004 | 6.49 c | 2.138 a | 113.71 ab | 19.05 a | 21.35 b | 3.87 ab | –11.71 ab | 41.76 a |
SC537 | 4.67 b | 2.38 abc | 121.71 ab | 20.32 a | 19.65 ab | 3.79 ab | –11.7 ab | 42.39 a |
0 | 3.13 a | 1.887 a | 107.62 a | 17.82 a | 23.75 d | 5.612 e | –11.584 bc | 32.5 a |
10 | 3.9 b | 1.881 a | 92.78 a | 20.6 b | 23.43 d | 4.613 d | –11.59 bc | 33.4 a |
20 | 4.61 c | 2.089 a | 107.09 a | 19.91 ab | 21.01 c | 3.877 c | –11.512 cd | 40.4 b |
80 | 6.48 d | 2.739 b | 138.68 b | 20.14 b | 16 b | 2.898 b | –11.652 b | 50.9 c |
160 | 7.59 e | 3.206 c | 149.56 b | 22.22 b | 13.54 a | 2.218 a | –11.869 a | 53.8 d |
Leaf N concentration varied significantly between genotypes and the effect of N-fertilization levels was highly significant (Table
All the analyzed leaf parameters (N, N/LA, SLA, δ15N, δ13C, C/N) were highly sensitive to variations in N fertilizer levels (Table
Additionally, the effect of changing light in outdoor conditions was evaluated in RGB indices obtained from canopy images (Table S1). For this purpose, 57 plots were photographed twice in nearly consecutive days, firstly in a sunny day and secondly in a partly cloudy day. All indices were strongly correlated between replicates (
All vegetation indices (either ground and aerial NDVI, RGBcanopy, or LCC) were strongly correlated with GY variation across the whole set of plots of the two trials. The best results were obtained by using the RGB-indices GA and GGA at the canopy level, which showed an exponential regression model and explained 70–72% of GY variability (Figure
Additionally, simple regression models from the P trial that explained GY across the different N fertilization levels were obtained by using the different VIs and validated for their accuracy in estimating the GY of the S trial (Table
GY est. | GA | GY = e ((GA−0.195)∕0.24) | 0.674 |
0.711 |
0.054 | <0.001 |
GGA | GY = e ((GGA−0.151)∕0.248) | 0.684 |
0.719 |
0.040 | <0.001 | |
NDVIaerial | GY = e ((NDVIaerial−0.192)∕0.094) | 0.452 |
0.543 |
0.002 | <0.001 | |
NDVIground | GY = e ((NDVIground−0.461)∕0.109) | 0.231 |
0.324 |
0.001 | <0.001 | |
GY exp. | GY 3-3 hybrids | – | – | 0.044 | <0.001 |
To further assess the accuracy of these indices, the determination coefficients for GY prediction within each N-input level across genotypic means were performed (Table
NDVIaerial | 0.019 ns | 0.415 |
0.092 ns | 0.074 ns | 0.621 |
NDVIground | 0.741 |
0.381 ns | 0.421 |
0.212 ns | 0.289 ns |
hue | 0.632 |
0.653 |
0.717 |
0.364 ns | 0.729 |
a |
0.706 |
0.634 |
0.627 |
0.524 |
0.464 |
b |
0.040 ns | 0.160 ns | 0.046 ns | 0.056 ns | 0.889 |
u |
0.709 |
0.608 |
0.666 |
0.491 |
0.569 |
v |
0.079 ns | 0.239 ns | 0.001 ns | 0.120 ns | 0.798 |
GA | 0.771 |
0.659 |
0.704 |
0.501 |
0.764 |
GGA | 0.872 |
0.664 |
0.774 |
0.555 |
0.748 |
LCC | 0.148 ns | 0.163 ns | 0.076 ns | 0.021 ns | 0.004 ns |
Leaf %N | 0.059 ns | 0.100 ns | 0.014 ns | 0.001 ns | 0.504 |
N/LA | 0.037 ns | 0.046 ns | 0.006 ns | 0.014 ns | 0.189 ns |
C/N | 0.085 ns | 0.069 ns | 0.006 ns | 0.009 ns | 0.365 ns |
δ13C | 0.002 ns | 0.353 ns | 0.019 ns | 0.073 ns | 0.001 ns |
δ15N | <0.001 ns | 0.007 ns | 0.005 ns | 0.025 ns | 0.065 ns |
SLA | 0.004 ns | 0.002 ns | 0.026 ns | 0.008 ns | 0.030 ns |
LCC was the best predictor of leaf N concentration across the entire trial, explaining more than 80% of N variability, moderately surpassing the fitting accuracy of the RGBleaf indices (Figure
A table depicting the determination coefficient between the RGBleaf indices, NDVIaerial, NDVIground, and LCC against leaf N across genotypic means within each of the N fertilization levels is presented (Table
NDVIaerial | 0.545 |
0.242 ns | 0.352 ns | 0.755 |
0.650 |
NDVIground | 0.113 ns | 0.223 ns | 0.285 ns | 0.403 |
0.057 ns |
hue | 0.081 ns | 0.155 ns | 0.188 ns | 0.094 ns | 0.012 ns |
a |
0.588 |
0.261 ns | 0.567 |
0.500 |
0.149 ns |
b |
0.526 |
0.489 |
0.566 |
0.437 |
0.265 ns |
u |
0.607 |
0.171 ns | 0.534 |
0.507 |
0.073 ns |
v |
0.581 |
0.546 |
0.429 |
0.316 ns | 0.192 ns |
GA | 0.097 ns | 0.013 ns | 0.185 ns | 0.517 |
0.222 ns |
GGA | 0.261 ns | 0.026 ns | 0.229 ns | 0.359 ns | 0.026 ns |
LCC | 0.587 |
0.426 |
0.401 |
0.168 ns | 0.005 ns |
N/LA | 0.810 |
0.127 ns | 0.643 |
0.496 |
0.552 |
C/N | 0.973 |
0.947 |
0.918 |
0.980 |
0.939 |
δ13C | 0.318 ns | 0.083 ns | 0.069 ns | 0.158 ns | 0.073 ns |
δ15N | 0.006 ns | 0.004 ns | 0.167 ns | 0.183 ns | 0.295 ns |
SLA | 0.167 ns | 0.069 ns | 0.219 ns | 0.022 ns | 0.121 ns |
Leaf N was strongly negatively correlated across N levels with δ15N and the C/N ratio and to a lesser extent with δ13C and SLA (Table S2). Correlations of these traits with GY were also negative but weaker, except for SLA which did not correlate.
Most of the RGB indices (both at the leaf and canopy scales), the LCC and the NDVI correlated with N/LA across N regimes, but always more moderately than they correlated with leaf N concentration (Table S2). The association of δ15N with NDVI, LCC, and RGB indices (at the both scales) was highly significant and in some cases their correlation coefficients were higher than the respective coefficients between δ15N and leaf N. Similarly, δ13C was fairly well correlated with most of the RGB indices (especially at the leaf scale) and LCC. Regarding the C/N ratio, LCC was the best predictor but this correlation was smaller than with leaf N concentration. However, most of the RGBleaf indices (a*, b* u*, v*, GA), the RGBcanopy indices (hue, u*, GA, GGA) as well as NDVIground and NDVIaerial correlated more strongly with the leaf C/N ratio than they did with leaf N (Table S1). Finally, SLA correlated strongly with the RGBleaf indices GA and GGA, and slightly with both NDVIs.
The relationships between leaf N, N/LA, C/N, δ13C, δ15N, and SLA with GY across genotypes within N fertilization treatments were almost all non-significant except for leaf N in the 160N treatment (Table
As previously found in other studies in wheat grown under different stress conditions (Casadesus et al.,
For its part, the range of variability in the RGBcanopy index, GA, was much wider (only 63% of values were in the range of 0.5–0.8) and GA values in the low N treatments were somewhat smaller (average GA = 0.46 in the 0N treatment) than those of the NDVI, and in fact GY correlated much better with GA than with the NDVI. Even so, the RGBcanopy indices also seemed to saturate for high GY but to a lesser extent than the NDVI because they mainly depend on changes in pigment concentration and the canopy LAI is less affected in the visible region than in the NIR region (Casadesus et al.,
In the case of the airborne NDVI data, the correlation with GY was also much lower than with GA taken on individual plots with GY. In fact, the images from the ADC multispectral camera have around four-fold less resolution than current digital camera technology (3.2 vs. 12 MP in our study, respectively). Although many ADC images were employed to obtain mosaics of the entire field trials, the resolution obtained at the flight altitude generated pixels which were mixed between pure vegetation, shadows and soil components. Such effects were successfully separated in the imagery collected at the near-canopy level with the RGB camera due to the higher resolution obtained. Altogether, the NDVIaerial provides a much lower amount of information than the GA and other VIs derived from RGB images taken at the plot level.
In the case of the LCC, it correlated strongly and linearly with grain yield across fertilization levels (Figure
Concerning the applications in breeding, the determination coefficients within N levels across genotypic means (Table
The importance of leaf N concentration for N management and breeding lies not only in its potential contribution to grain N (Gallais and Coque,
Our study highlights the potential of RGB indices for precise crop N management and for phenotyping genotypic performance under a wide-range of N conditions. As widely reported, LCC proved to be a very good indicator of leaf N concentration across nitrogen fertilization levels, therefore enabling monitoring of N application (Hirel et al.,
By contrast, in the highest N-fertilization level (160 kg ha−1) the RGBleaf indices and LCC were probably saturated because they did not correlate with variations in leaf N concentration. For its part, the NDVIaerial had an irregular trend as it was significantly correlated to changes in leaf N concentration at three of the five N fertilization levels (0, 80, and 160 kg ha−1) and these correlations were especially strong in the high N levels. As discussed above, besides of some plot variability and soil exposure, the poorer performance of the NDVIaerial may be mainly explained by the relatively poor spectral resolution at the single plot level of the multispectral aerial images. Even so, according to our results this approach seems efficient for its implementation in aerial platforms.
Besides the leaf N concentration discussed above, other leaf N parameters like the N concentration on an area basis (N/LA) and the C/N ratio were strongly associated with GY across N fertilization levels. In the case of the leaf δ15N, its value gradually decreased as the N application rate increased. This trend is due to the absorption of N from chemical fertilizers that are highly depleted in 15N, whereas in the low N treatments plants absorb the N available in the soil, which is usually 15N-enriched (Bateman et al.,
In agreement with previous studies (Dercon et al.,
Previous studies noted the relevance of SLA for the compositional and ecophysiological characterization of plants (Reich et al.,
Regarding the relationship between VIs and the C/N ratio, most of the RGB indices (at the canopy and leaf levels) and both NDVI approaches were demonstrated as being even better correlated to the leaf C/N ratio than to leaf N concentration (Table S2). This finding may have considerable economic implications as the C/N ratio informs not only about the crop N status but also about the aerial biomass quality, including digestibility and nutritional quality (Van der Wal et al.,
The tested vegetation indices based on RGB images and to a lesser extent the NDVI demonstrated a high-throughput for the accurate prediction of several traits that are highly valuable for maize breeders and agronomists such as grain yield, leaf N concentration and the ratio of carbon to nitrogen under a wide range of N fertilization levels. Proper N fertilization management may be assisted considerably by using these parameters as decision criteria controlling the expected production and the uptake of N by the above-ground biomass. Beyond this, maize breeding programs may benefit from these findings through their application during the characterization of genotypic performance within N fertilization levels. In this way the selection of the most efficient genotypes in terms of grain production and/or N uptake may respond to the needs of low N stress tolerant maize varieties.
Vegetation indices derived from RGB images proved to be broad-use because they were previously employed satisfactorily in other crops under biotic and water stress conditions (Casadesus et al.,
Although the performance of the RGB indices (obtained from JPEG images) worked well in this study, future research may address the possibility of further improve their accuracy by using input images saved in a lossless compression format as TIFF or PNG. Despite of storage inconvenient, their larger capability (16 bit per pixel instead of 8 bit) may maintain higher quality detail from the visible spectrum. Another important consideration is the effect of changing light conditions when making these outdoor measurements. Despite of the good strength and repeatability of the results (Table S1) fluctuating ambient lighting should be considered as a possible source of error. Further research should also be targeted toward implementation and evaluation of similar RGB phenotyping methods in remotely piloted aerial platforms (Elazab et al.,
BP, JC, MZ, and BM managed and directed the maize programme in the Southern Africa regional office of CIMMYT in Harare, Zimbabwe. MZ, PZ, and AH carried out the UAV flights for the obtainment of aerial measurements. On the other hand, JA, BM, JC, MZ, and OV conducted the field measurements and the collection of samples. AH and PZ processed the aerial images. OV analyzed the samples and other data and wrote the paper under the supervision of JA and with contributions from all the other authors.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
This article was supported by grants from the MAIZE CGIAR Research Program and the Project AGL2013-44147-R from the Ministerio de Economy Competitividad of the Spanish Government. OV is a recipient of a research grant (APIF) sponsored by the University of Barcelona. We thank the personnel from the CIMMYT Southern Africa Regional Office at Harare for their support during the field measurements and sampling. The trials were planted under the Bill and Melinda Gates funded project Improved Maize for Africa Soils. Finally we thank Dr. Jaume Casadesús for providing the BreedPix software.
The Supplementary Material for this article can be found online at: