3.1. Plant Height Validation
UAS-based maize plant height was estimated according to the method introduced in Section 2.5
. The width of the centerline-developed polygons was set to 10 cm, and the 99th percentile in height was selected for validation. It is worth clarifying that the UAS platforms did not fly on exactly the same dates when the ground-measured height datasets were collected (Table 3
). To enable a height comparison as fair as possible, the flight experiments closest to the ground data collection dates were selected.
provides a comparison of the ground-measured maize height and UAS-measured maize height in the field trial. Rather than separating the results from the upper and lower portions of the field trial, the figure merged results from both portions and the height comparison of individual dates is differentiated by color, as explained in Table 3
. The left subfigure shows the overall height comparison across various dates for both upper and lower portions of the field trial. The 1:1 relationship is marked with a dashed black line. The right subfigures illustrate a height comparison for individual dates, and the red solid lines depict linear regressions between ground observed height and the UAS-estimated equivalent. In the right subfigures, the coefficient of determination (R2
) for most individual linear regressions ranged from 0.26 to 0.43, whereas a dramatically lower value (R2
= 0) was observed for the cyan dots (i.e., 6 May 2016). A possible reason for the failed correlation of cyan dots (i.e., 6 May 2016) could be caused by a relatively low height variation (<0.4 m) as compared to other dates, making SfM processing noise a dominant negative factor for obtaining a higher correlation. Moreover, the reduced correlation could also be affected by the inconsistency of height extraction norms between the UAS method and the ground survey. Specifically, within individual plots, the UAS method estimated height by leveraging signals at 40 mm resolution (CHM pixel resolution) in centerline polygons, while the ground survey extracted height by averaging measurements from three plant samples.
In addition, the green and black dots (i.e., UAS images taken on 17 May and 8 June 2016, respectively) reveal a slight height overestimate. This is because the flight experiments were conducted two to three days after the ground-measured plant height was collected. On the contrary, the underestimate from blue dots stemmed mainly from the flight occurring one day earlier than ground height collection at the rapid growth stage (i.e., around V10). Slight underestimate from magenta and cyan dots was primarily caused by the inclusion of the soil in centerline-based polygons due to the small size of the plants.
However, according to the left subfigure on a broader time scale, the scattered comparison demonstrates a good detection of the actual growth rates across the growing season (R2
= 0.88). This verified the validity of the UAS height estimation introduced in Section 2.5
3.3. Lodging Assessment
In the proposed lodging detection method, the values of the predefined parameters thrd90 and thrd99 are subject to change during the growing season depending on the actual plant height of that growth stage. In the present work, the DJI Phantom 4 flight image dataset collected on 30 June 2016 (i.e., 90 DAP), was selected for subsequent lodging detection assessment. Empirical tuning of the predefined parameters in order to ensure the optimal lodging detection performance suggested values for thrd90 and thrd99 equal to 0.15 m and 0.45 m, respectively. Assessing the lodging severity at an earlier growth stage may suggest a different set of thrd90 and thrd99 settings depending on the spatial patterns of canopy structure; however, this is beyond the scope of discussion in this study. In terms of the grid cell spacing in a row, i.e., Lg, its value was chosen to stay consistent with the seed spacing in a row (i.e., ), and this paper rounded it to Lg = 0.2 m. Larger cell spacing is prone to underestimating the lodging severity because the lodging plants in a cell may not be detected by using the conditional statement (6). Small cell spacing, on the contrary, is liable to produce overestimated lodging counts by summing up excessive grid cells determined as lodging through conditional statement (6).
provides the demonstration of the UAS-based lodging detection over eight consecutive rows on the lower portion of the field trial. The left subfigure shows the background geo-referenced orthomosaic image produced from DJI Phantom 4 flight images taken on 30 June 2016, while the grid cell-based lodging detection results are superimposed on the right subfigure. From the left subfigure in Figure 8
, the most severe lodging was observed at the third row. This is intuitively consistent with the detection results shown on the right subfigure. In addition, lodging plants at other rows are also successfully detected and geo-located.
A complete correlation matrix displaying the correlation coefficients among the ground-measured lodging rate (GLR), UAS-estimated lodging rate (ULR), canopy structural complexity metrics, and yield is presented in Table 4
. For calculating GLR, stand counts manually collected in the field were used as Np
; while, in the ULR approach, Np
was estimated via Equation (10) in an effort to minimize the usage of field data, and the ULR value was calculated via Equation (11). The left-most column shows how the height metrics, yield, and ULR correlate to the GLR. High negative correlations (0.55 <
< 0.60) were found between the GLR and Hmean
, as well as H50
. Reduced correlation values were obtained when the percentile in height increased. This is because the lodging rate depends little on local maxima or minima of 3D canopy structure, but on the ratio of numbers of height measurements with high and low values. Slightly better correlations (
≥ 0.60) were observed when using Herr
as these two metrics further reveal canopy structural heterogeneities based on CHMs. Herr
was negatively correlated to the lodging rate because, as defined in Equation (3), Hmean
in grid cells usually decreases when lodging rate increases. This also explains why Hcv
was positively correlated to lodging rate, as defined in Equation (2), while a lodging grid cell is also inclined to produce a higher Hstd
than a non-lodging one.
By using the lodging detection method introduced in this work, the most significantly high correlation was seen between GLR and ULR (r
= 0.71). Intuitively, the direct comparison between ULR and GLR is shown in Figure 9
. The results substantiate that the proposed UAS-based lodging detection method has a great potential to accurately reflect lodging severity and could potentially replace manual measurements in the open field environment.
In addition, looking horizontally at the bottom line, Table 4
displays how the ULR correlated to other structural complexity metrics. Compared to the GLR, the ULR correlated more tightly to all of the structural complexity metrics as the proposed UAS-based lodging detection method mathematically assessed the lodging severity based essentially on the canopy structural complexity. Compared with other metrics, the significantly better responses of the Hmean
, and Hcv
to the ULR (
≥ 0.75) demonstrate that the proposed method possessed some similarity with the metrics of Hmean
, and Hcv
Accuracy assessment of the number of ground-measured lodging plants against UAS-estimated lodging plants was also conducted within individual rows (Figure 10
). Each blue dot represents an individual row in the lower portion of the field trial. In general, the estimated numbers complied with a linear relationship (red solid line) with ground measurements that were close to the 1:1 line (black dashed line) with a R2
However, the linear trend in Figure 10
was visibly contaminated by noisy blue dots. The noise can be divided into two categories, i.e., overestimates and underestimates. Overestimates of the number of lodged plants is illustrated by the blue dots below the 1:1 line (highlighted as the brown oval in Figure 10
). The reasons of overestimate are twofold. (1) Overestimates were susceptible to ground observational errors. More specifically, some plants that bent over 60° were mathematically estimated as lodging in the method but were not included by the ground data collector (Figure 11
). (2) Late in the season, plant structures were liable to be failed in forming a closed canopy in the point cloud, leading to a suppressed height estimation [17
] and overestimated lodging detection. The blue dots above the 1:1 line (highlighted as the pink oval in Figure 10
) refer to the underestimates of the number of lodged plants within individual rows, which also stemmed from two primary causes. (1) At times, maize leaves from non-lodging grid cells extended to lodging grid cells, which, to some extent, deceived the lodging detection method proposed and resulted in lodging grid cells being miscategorized as non-lodging ones. (2) In the proposed method, the plant stand count within an individual row was estimated as per Equation (10) (i.e., multiplying the average seeding rate by the length of the row centerline) in order to maximize the automation in the workflow instead of manned data collection in the field. However, after comparing it with the manual plant stand count in the field, it was found that the estimated stand count was prone to an underestimate given the seeding rate Rs
in the field trial (Figure 12
). Therefore, the underestimated stand count may in turn cause a slight decrease in estimating the number of lodged plants. On the other hand, accurate stand count based on early season imagery is one of the primary measurements that studies have been working to automate due to its importance as a metric for both farmers and researchers. Furthermore, it is believed to be a relatively straightforward metric to estimate using image processing algorithms [40
]. Once these algorithms are readily used and implemented with a UAS surveying approach, estimation errors introduced by using seed count are expected to be eliminated. Although not overly strong lodging correlations were observed in Figure 9
and Figure 10
, the UAS-based method provides potential and feasibility to identify lodged maize plants in the field.
The impact of weeds in this work was minimized by the centerline height extraction technique because: (1) height was estimated only around row centerlines with 10 cm width. By doing this, most of the weeds were filtered out by excluding any height signals outside of the centerline polygons; (2) weeds in the field were very low in height when compared to the maize plants. It was observed in the field that lodged maize plants were even higher than those remaining weeds growing along the rows.
It needs to be highlighted that the UAS image dataset used in this section was obtained on 30 June 2016, while the ground-measured lodging rate was collected on 25 July 2016. This is technically sound for comparison. As was discussed previously, the image datasets collected in July 2016, were not used for analysis due to poor canopy coverage. In addition, maize plants at this stage became increasingly less likely to further straighten up or to fall down [39
]. The overall UAS-based lodging detection results on the lower portion of the field trial, within individual rows, are superimposed on an orthomosaic image produced from a DJI Phantom 4 flight on 30 June 2016 (Figure 13
). A green polygon indicates a maize row less likely to be affected by lodging (low lodging rate), while a red polygon indicates a maize row more likely to be affected by lodging (high lodging rate).
A challenge facing practical implementation of the proposed method lies with its potential automation. In this work, most manual operations were associated with marking GCTs after a sparse 3D point cloud has been generated and when selecting centerline endpoints for each of the individual rows. Fortunately, the emergence of state-of-the-art real time kinetic (RTK)-equipped UAS platforms helped eliminate the need for tedious GCT-based georeferencing in SfM processing [42
]. Moreover, algorithmic attempts on automatic crop row detection have recently been made, which can facilitate the automated crop centerline identification [43
]. In other words, by using a RTK-enabled UAS platform and a more elaborate row centerline extraction technique, a complete automated lodging detection workflow can be achieved, regardless of the size of the field trial under study.
Although rare, it may happen that the actual canopy height of non-lodged maize plants is lower than the thresholds introduced in the proposed method due to a variety of factors, such as differences in plant type, management treatment, soil content, disease or insect infestation, and other human intervention. For large-scale production trials, as opposed to a small-scale field trial, like shown in this work, it is speculated that this issue may cause some false-lodging detection. Therefore, further examination on large-scale field trials is needed to better evaluate and improve the performance of the method.
To better understand the variability that may arise in linking the maize lodging severity and yield pattern, yield in response to the lodging rate is provided (Figure 14
). It intuitively confirms the fact that the plots that were not severely lodged usually produced higher yield in the field. Furthermore, the linear trend of yield vs. lodging rate obtained by using the UAS method was found to be closely consistent with that obtained by using ground measurements. In the Figure, results from some maize plots were observed to be off of the linear trend when the lodging rate was relatively low, particularly in the highlighted green oval. The lower accuracy achieved in the oval in Figure 14
is thought to be caused by a reduced germination rate in some plots, even if the lodging rate detected was low. In regards to this concern, improvement can be expected from excluding any plots with a relatively low germination rate (e.g., below 80%), provided that the germination rate information is available. Other possible causes are phenotypic differences in terms of the genetic suitability (yield potential), seed origin, or management treatments over plots in the field trial.