3.1. Crop Height Determination with UAV Point Cloud
The GSD was resolved for the orthomosaic and digital surface elevation, with an average of 15.1 points/cm2
produced in the point cloud. The geolocation error of the generated data (reflected in the RMSE) was calculated using the initial and reconstructed GCP locations and varied between 1.5 and 4 cm over time, representing an acceptable level of accuracy given the ground control point error (<5 cm) and the high resolution of the DSM (~2.5 cm). Figure 4
(top row) presents the derived DSM outputs from Photoscan (gained from the merged and filtered point cloud), together with the crop height level estimates (middle row), retrieved as discussed in Section 2.4
, that is, by subtracting the baseline terrain map collected immediately after the crop sowing date (on 4 April), from each subsequent DSM survey. Finally, the pattern of the crop height anomaly around the mean was also evaluated (bottom row).
As can be seen in Figure 4
, an examination of the retrieved DSM and crop height from the UAV provides considerable insights into the small-scale variability of the crop systems. The UAV-based SfM methodology was able to discriminate areas with both abundant and sparse vegetation (the heterogeneity in the maps highlights this aspect), allowing for the detection of intra-field variability at specific times throughout the crop growing season. Several characteristics of the pivot development can be depicted from the DSM and crop height retrievals. Areas with lower vegetation can be seen throughout the field. Although some areas (stretching to the center) seem to recover by the end of the season (i.e., around the access road on the right of the center pivot), many areas identified as under-performing relative to the field average seem to be maintained in the 25 June scene, including around the perimeter of the field. Certainly, it is likely that vegetation along the periphery of the pivot receives less water from the sprinkler irrigation system due to wind effects and associated spray-losses. Such a vegetation buffer response is noted across many of the pivot systems, with vegetation on the interior of the field faring better than that on the exterior. Additional factors explaining the observed intra-field variability may include: (i) heterogeneity in the soil texture and composition; (ii) chemical properties (i.e., salinity) influencing the growth and development of the crop; and (iii) uneven distribution of irrigation and fertilizer application rates, which are delivered through a combined fertigation system (with possible lower efficiency at the terminal end of the pipe). In this context, regular monitoring of crop dynamics throughout the season represents an important aspect of precision agricultural application and agricultural management decisions (e.g., for irrigation scheduling, fertilizer application and harvesting) [72
]. Although some areas still remain under-performing until harvesting date, it is evident from the crop height graphs that no areas within the field get worse (in terms of crop height) during the season, demonstrating a good management practice and the absence of any adverse conditions that may affect specific areas of the pivot (i.e., meteorological events). While evaluation of the temporal intra-field variability of crop status at different growth stages has been reported in previous studies [45
], these analyses were either limited to smaller fields, or used a lower number of flights to characterize the variability.
More detailed information can be inferred from the texture of the DSMs (Figure 4
, top row). A wide range in elevation is present within the field, as emphasized by the SW-NE gradient. Indeed, at its maximum, there is a vertical difference of almost 4 m from one side of the field to the other (also verified by the GCP range in Table 1
) with a spatial pattern that remains largely unchanged until 28 April. The presence of low (or no) vegetation within the field through the first three UAV campaigns drives this consistency in the spatial pattern of surface elevation. In line with this, the crop height derived on 18 April is characterized by the presence of dark areas in the upper part of the field where the actual height presents some negative values. This is probably caused by the bare soil map (4 April) that raises above the true terrain level, leading to negative height when subtracted from the DSM of 18 April. In practice, although negative values are often removed (as a crop height below zero has no physical meaning), we decided to leave them to highlight the inconsistent retrieval that can arise from typical UAV sensing systems. A better understanding of this behavior can be obtained by examining Figure 4
(bottom row), which plots the anomaly of crop height around the mean value within the field, from which the percentage of crop height localized between a certain range (−0.6:0.6 m) for each UAV campaign was estimated. It should be noted though, that the negative values at 18 April represent only 0.7% of the entire distribution and they are mostly localized in the left and right areas at the top of the field, close to the perimeter (displayed as dark blue areas). It is clear from the localization of these negative regions that this inconsistency is most probably arising from errors due to lower overlap in the UAV imagery at the borders of the flight path compared to its center. This condition affects the number of key points that can be detected between images, producing a point cloud with lower density and therefore, with lower accuracy. Also, as stated in Section 2.2
, the flight performed during the second UAV campaign (18 April) was affected by a camera-bay blockage, which required a cropping procedure to remove the dark area over each image. The reduced size (and overlap) of the pictures definitely affected the imagery post-processing, as the vertical size of the image was reduced by almost one quarter of its total length (Table 1
On 18 April, there are two banded regions in the upper- and lower-middle areas of the field that show positive anomalies above the mean (see bottom row of Figure 4
), that do not seem to be maintained in the subsequent campaigns. It should be noted that the crop height retrieved on this date is likely affected by a higher level of uncertainty compared to those in later UAV flights, as the thin structure of the vegetation at the early stage (emergence) is not well captured by the generated point cloud. As reported by Grenzdörffer et al. [46
], errors are strongly related to the stage of crop growth. That is, crop height is better determined by the UAV-SfM system if the canopy surface is dense and homogeneous. In early development stages, crop height is more challenging to retrieve as the vegetation cover is lower and the individual stalks are small and do not normally form a closed canopy.
The pattern of the crop height anomaly around the mean (Figure 4
, bottom row) is also able to identify areas of higher and lower performances around the field, bringing attention to specific areas that are seen to enhance the growth and development of the maize throughout the season. In particular, the “positive” anomaly captured on 28 April in the S-E part of the field is followed by a similar structure in the next two UAV flight dates, confirming the capacity of that particular area to provide conditions more favorable to crop development (e.g., perhaps driven by better soil nutrient and water content availability). Similar responses are shown for the vertical stripe on the left side of the field, which is also repeated on 9 May and 25 June. Such improved crop height response in these particular zones may be a consequence of a different soil texture, type or salinity properties, which led the plants to be higher than the crop height average: by up to 75 cm on 9 May and 62 cm on 25 June. A closer examination of the last two campaigns indicates that the anomaly around the mean is more emphasized on 9 May rather than on 25 June, demonstrating a stabilization in terms of average height within the field during the last month of the growing season. Based on the growth trend of the crop, the penultimate field campaign matches the flowering period, after which the canopy structure is subject to an increased density (with expanding leaves and corn cob) rather than any significant rise in height. This period between image collections could have given the plants time to stabilize in terms of height, producing a smaller anomaly compared to the previous campaign.
To highlight the ability of the SfM to retrieve fine structural changes in crop height, three sub-areas of 30 m × 30 m were extracted from within the field and presented in Figure 5
. A healthy crop (Area 1), a static no-crop region (Area 2) and a problematic area (Area 3), were retrieved throughout the growing season, with changes in crop coverage and plant height depicted over the sampling dates. Perspectives from Areas 1 and 3 illustrate that the crop canopies were adequately reproduced by the point clouds, with crop growth well captured across the different stages. Visual interpretation of the images in Figure 5
reflects the spatial agreement between the derived crop height and the underlying orthomosaic, confirming the efficiency and accuracy of the UAV-SfM approach. As can be seen, Area 1 depicts a relatively stable canopy coverage for all dates, especially when compared with Area 3, which shows the impacts of soil or irrigation-related problems on the canopy structure. Further examination indicates that Area 1 and 3 follow the same growth trend up until 28 April, after which the crop development and subsequent canopy response diverge. For instance, by 9 May, Area 3 shows a loss in canopy structure, with half of the area of the 3D model and underlying orthomosaic reflecting the low canopy structure (although this is partly recovered by the time of the last retrieval on 25 June). Area 1 on the other hand, manifests a more linearly increasing crop height trend and is absent of any significant canopy alterations throughout the season. As expected, Area 2 remains consistent throughout the UAV collection period, reproducing the static bare soil section in the center of the pivot.
The dynamic crop height across the entire field was also visualized by considering the histograms of all pixel-based elevations for each of the individual UAV campaigns. The histograms were constructed by placing each z-coordinate of the crop surface models in a bin-size of 1 cm, providing an accurate representation of the distribution and frequency of the crop height estimates. As can be seen from Figure 6
, the time-evolution of plant height is apparent across the growing season. The early stage distribution from 18 and 28 April, presents a sharp peak and narrow Gaussian distribution, reflecting the low crop height (25–50 cm) characteristic of the maize field at the beginning of the growth cycle. In the case of these low canopy retrievals, it can be observed that the distribution depicts some negative values in the field for the April 18 campaign, suggesting some inconsistency in the results. As stated before, negative values of crop height were obtained because ground values in the first campaign had higher elevation compared to those created on 18 April, thereby leading to negative height in the derived crop surface model. For this particular date, the crop coverage within the field is still low and the captured bare soil is almost the same as the previous campaign (4 April). However, any error in the UAV GPS coordinates, even though these have been corrected through the introduction of GCPs, still presents as a source of uncertainty, ultimately producing lower crop height estimates relative to the first survey (4 April). The plant height corresponding to the last two UAV surveys (9 May and 25 June) shows the highest intra-field variability, reflected by lower peaks and a wider range of the distribution: from 0 to 1.6 m and from 0 cm to 2.35 m, for the respective campaigns. Despite the similarity in the distribution of the last two UAV campaigns, 9 May shows a bimodal distribution that is not present on 25 June. This is explained by the presence of areas with low vegetation and with heights less than 50 cm, which represent 11.3% of the values in the distribution. On the other hand, 77% of the field has a height above 0.992 cm, which represents the average crop height. This intra-field variability, clearly depicted in the histogram, is also emphasized in calculations of the 25th and 75th percentile, whose values are 0.914 cm and 1.30 cm, respectively.
Interestingly, it can be noted that the crop growth occurring in the two-weeks between 28 April and 9 May (i.e., about 1 m in height) represented a more rapid height development relative to the following month and a half between 9 May and 25 June (about 60 cm). The fast crop development after 28 April, which may explain (in part) the high spatial variability captured by the fixed-wing UAV on 9 May, is further confirmed by the differences in the median values recorded from the last three campaigns: 0.25 cm, 1.20 cm and 1.86 cm.
The typical length of the maize season in the study region usually ranges between 70 and 90 days, depending on the sowing date as well as climatic and environmental variables [73
]. Indeed, the timing of crop phenology from one growth stage to another (development periods) is directly affected by temperature changes [75
] and it is strongly correlated with the cumulative daily temperature [78
] (i.e., degree days). Generally, cool temperatures tend to slow down growth, while warm temperatures hasten maturity [74
]. The rapid development of crop height that occurs between 28 April and 9 May reflects the crop reaching the end of its vegetative stage, which starts approximately two weeks before flowering (around 9 May in our case) [80
]. During this rapid growth phase, the stalk follows a substantial development, leading plant height to increase dramatically [82
Being able to track crop growth at the intra field scale provides an important metric with which to understand and assess the multiple developments that can occur in diverse areas within a field of such size. From this, specific management decisions can be implemented to improve the response of problematic crop areas, hence reducing the risks of potential yield losses. Timely and accurate prediction of crop height during the growing season is important in farm management, as it can be used by farmers and operators for improved decision-making [83
] and by government and researchers agencies for informing food security policies [84
]. Further assessment of the accuracy of these retrievals is provided below.
3.2. Evaluation of UAV-Based Retrievals with LiDAR Scans
How accurately the digital surface model and baseline terrain map are able to be determined by the UAV-SfM technique is critical to the accurate determination of crop height. Therefore, an evaluation procedure is needed for assessing the UAV-SfM reconstructions. In the following, a comparison between UAV and LiDAR scan data is presented, providing a mechanism to assess the accuracy of both bare soil and digital surface model retrievals (from which the crop height is extracted). Despite the higher spatial coverage of the fixed-wing UAV system, which allows for the survey of an entire 50 ha field in a single flight, the ground-based LiDAR system provides a higher point density and accuracy but over a much smaller area. A pixel-by-pixel comparison between the two datasets is performed, with varying numbers of measurement (i.e., pixel-height) analyzed based on the sampled resolution (i.e., 1,269,822; 317,455; 79,364; and 19,892 pixels were considered in dataset comparisons at 2.5, 5, 10 and 20 cm, respectively). Figure 7
presents a summary of the UAV and LiDAR retrievals for the bare soil surface acquired on 4 April.
Overall, the UAV-SfM results represented in the density scatter plots show very good correlation and low RMSE (few centimeters) at the four different resolutions. In all cases, the data are consistently retrieved and distributed around the 1:1 line. The 5 cm resolution result presents the lowest bias compared with the other three resolution retrievals, each maintaining a fairly consistent positive bias of approximately 5 cm. However, in all cases, the SfM elevation was slightly higher than the corresponding LiDAR values, which is most likely explained by errors in the SfM point cloud reconstruction and in the ground measurements [47
]. As expected, the highest correlation was obtained for the data at 2.5 cm and 5 cm (r2
= 0.99), followed by the 10 cm (r2
= 0.98). The worse correlation was determined for the 20 cm (r2
= 0.77), which also presents the highest value of RMSE, at 3.1 cm. RMSE values for the 2.5 and 5 cm retrievals show a well contained error, with a range that varies between 1.6 and 1.8 cm, while 10 cm resolution was only slightly higher (2.4 cm). The mean absolute error (MAE) and the relative root-mean-square error (rRMSE) further reinforce this response, with an observed range amongst the pixel-by-pixel comparisons between 0.0126 and 0.0246 cm for the MAE and 0.45% and 0.85% for the rRMSE.
Overall, these results clearly illustrate the ability of the UAV-SfM approach to accurately model the terrain surface elevation, which is critical for a reliable extraction of the crop height values. Employing a different point cloud density, hence producing different resolution scales (i.e., obtained from resolution-specific processing at each scale), did not particularly affect the results, whose accuracy is maintained even at a lower level of detail (i.e., 10 cm). It should be noted that while the results are only marginally different (in terms of correlation) using 2.5, 5 and 10 cm resolution, employing coarser resolution (i.e., 20 cm) data reduces the reliability of the dataset, having an r2
of 0.77. The reduced point cloud density obtained at 20 cm most likely explains the difference between the correlations obtained at the other resolutions. The correlation between the baseline terrain map retrieved with the UAV-based SfM approach and the LiDAR point cloud was also assessed by Malambo et al. [32
], who confirmed the importance of an accurate terrain map to extract reliable crop height. Also in that study, high values of the coefficient of correlation were achieved (0.88–0.97), although fewer measurements were utilized for the statistical comparison (n = 380), as a result of a raster grid interpolation.
During the course of crop development, vegetation growth is not a continuous or linear phenomenon but follows a series of generally well-defined crop stages [81
]. In the early development phase (emergence stage), the height of the crop is challenging to retrieve, as the vegetation density is low and lacking a closed canopy structure. Hence, UAV technologies will generally struggle to identify small stems and leaves, especially if the flight altitude is high (such as in this case, at more than 100 m). From an optical sensing perspective, the resolvable resolution in these conditions can result in a poor discrimination of vegetation from the underlying ground surface, resulting in a relatively homogeneous response signal. Furthermore, if the crop structure is either sparse or particularly thin, the UAV will deliver a lower crop height compared to the LiDAR. This “height dampening” is a consequence of the filtering step embedded in the SfM algorithms [58
], which smooths out the solution, hence leading to a poorer identification of single plants in the emergence stage. As the crop density and structure change during the development stages, crop height should be resolved with an increasing level of accuracy. The challenging retrieval of crop height in early development stages has been highlighted by Grenzdörffer et al. [46
], who considers the UAV-SfM limitation increasing at coarser resolution because within a single pixel, portions of one or multiple plants (and their shadows) merge into one single signal (canopy level). In Shi et al. [47
], the weak correlation between UAV estimates and ground truth was found to be a result of the inadequate image resolution, which was not able to distinguish the small tassels on top of the plants, that were measured on the ground. Furthermore, Bendig et al. [42
] and Malambo et al. [32
] pointed out the lower fidelity of the UAV-SfM system as a consequence of the restricted viewing perspective (nadir angle), which does not allow for a full 3D reconstruction of the canopy.
Following these considerations and the similarity in the statistical results achieved for the terrain base map at the four different GSDs, we carry out a subsequent evaluation of retrieved crop height using the same resolution scales, with the aim of testing their accuracy in providing the necessary intra-field variability: a key requirement for any precision agriculture based approach. For this purpose, the same portion of the field evaluated in the bare soil analysis, is now considered for crop height estimation on the penultimate UAV campaign (9 May), where the crop height variability is much more noticeable. It should be noted that due to the “blocking effect” caused by the first crop (see Section 2.4
for more details), some small areas at further distances could not be fully represented by the LiDAR scanned point cloud, hence generating “No Value” at those particular locations. For consistency in the UAV/LiDAR comparison of the results, the pixels of the UAV crop height map located at the same coordinates of the LiDAR “No Value,” have been removed. Figure 8
shows a pixel-by-pixel comparison of the UAV and LiDAR systems, which are reported in density scatter plots, while the comparison of their crop height maps is presented in Figure 9
. From our analysis, UAV-based crop height retrievals are shown to be quite reliable in detecting the variability within the considered area at the edge of the field. The pattern of the crop, depicted in the underlying maps (Figure 9
), is accurately predicted by the UAV system, which is also able to distinguish the adjacent crop rows that are separated by about 50 cm from each other. From these analyses, the heterogeneity in the vegetation along the periphery of the pivot is clearly impacted relative to the pivot average, with losses in canopy structure that are most likely explained by reduced irrigation efficiencies at the terminal end of the sprinkler system (see Section 3.1
for more details). It should be noted that scatter plots and crop height maps clearly represent two distinct distributions: the majority being bare soil, which is reflected by a high density area in the scatter plots (in yellow, where the height is nearly close to zero) and a smaller proportion being crop, which presents a linear trend with similar values for both UAV and LiDAR.
In terms of error, the high values of the rRMSE (with a range between 37% and 50%), are therefore the consequence of these different (and multiple) distributions in the scatter plots. As expected, the significant differences in crop structure within the area are better delineated using higher resolutions (i.e., GSD of 2.5 and 5 cm), which also produce the lowest RMSE (0.21 cm–0.22 cm). The solution at 10 cm seems to produce a less similar pattern to that generated by the LiDAR, which further worsens when using a 20 cm GSD. However, good similarity in crop height is reflected in the correlation indices of the density scatter plots (Figure 8
), with values of 0.57 at 20 cm resolution, 0.60 at 10 cm and 0.62 and 0.65 using 2.5 cm and 5 cm GSD, respectively.
In general, all the resolution scales generate lower crop height estimates compared to the LiDAR, reflecting the “height dampening” typical of the SfM filtering process [58
]. Although a negative bias is clearly visible at every resolution scale, the crop height above 0.5 m seems to be accurately predicted by the UAV systems, with values that remain close to the 1:1 line in all cases, albeit with some evident underestimation by the UAV system across all resolutions. It can be noted that a gradual reduction of the maximum crop height detectable by the UAV is present from 2.5 to 20 cm. In particular, the highest values of crop height (1.4 m) are retrieved using a 2.5 cm GSD, whilst the solution at 20 cm only managed to detect values up to 1 m height. This gradual reduction can be explained by the heavier height dampening factor produced by averaging the point cloud over a larger area (i.e., 20 cm).
Interquartile statistics of the two datasets plotted in Figure 8
, show that the 75th percentile of 2.5, 5, 10 and 20 cm GSD are 2.7, 9.3, 21.2 and 30 cm lower than the LiDAR, while the 25th percentile are 9.2, 11.3, 21.7 and 23.1 cm higher, testifying to a reduced range in the UAV crop height estimation. While the lower 75th percentiles may be explained by the smoothing effect embedded in the SfM algorithms, the higher 25th percentiles are likely explained by a reduced capacity of the UAV system in retrieving the smallest and thinnest plants, which present the lowest crop heights [46
]. However, although the UAV lacks the capacity to effectively retrieve the outliers (i.e., the min and max crop height), it provides a fairly strong correlation, even at courser resolution (10 cm GSD).
A further discussion can be made from Figure 8
, where a vertical feature is present on the left side of all the scatter plots. The UAV-SfM approach is generating crop height values up to and over 1 m, while the LiDAR system is not. One main source of discrepancy can be represented by unstable data acquisition conditions. As previous studies have explained [46
], the vegetation surface should be stationary during the aerial survey to ensure a successful image matching in post-processing and a highly accurate positioning determination. Nevertheless, in most agricultural environment, wind effects are a key factor influencing retrieved imagery [85
]. Crop movement due to high winds impacts the quality of the imagery by introducing positional error, since the same feature of the crop can be recorded in different positions. In addition, errors in the vertical crop height are most common during the later growth stages when the plants are taller and may be slanted during wind condition [31
]. Records from the farm weather station, reported that the wind speed at the time of the UAV flight on 9 May (10:30 a.m.), was about 10 km/h, which can partially explain the errors in the UAV retrievals. As shown in Figure 9
, the UAV at the highest resolutions (i.e., 2.5 and 5 cm) was able to distinguish the crop rows and the adjoining soil. The wind effect during the time of the flight may have bent some of the plants, which consequently covered the adjacent soil pixels. As a consequence, the UAV can retrieve higher values, as shown in the scatter plots. Another source of uncertainty is represented by small geolocation errors that could also have affected the reliability of the results, especially at the higher resolution (i.e., 2.5 cm). It should be noted that a UAV versus LiDAR pixel-by-pixel comparison at such a high resolution has not previously been reported in the literature. Similar evaluations of UAV derived crop height against LiDAR point cloud have relied on either an interpolation of the LiDAR data onto a 2 m grid [32
], or averaging the crop height onto a 0.3 m [57
] or a 0.5 × 0.6 m grid [56
], resulting in comparatively few measurements for the evaluation, relative to the present study. The correlation obtained between LiDAR and UAV plant height is consistent with previous findings in maize [32
] and barley [57
]. In line with these studies, a better correlation has been obtained between crop height measurements at the flowering stage, confirming that imagery captured at the end of the stalk elongation (flowering stage) is better correlated with ground truth data [86
As Figure 9
shows, crop height can vary considerably within the same field, even though irrigation and fertilizer application are applied with a consistent management practice across the pivot. However, other conditions can also influence the observed intra-field variability. These conditions can be quite variable and involve natural fluctuations in biological and plant physiological processes, soils and climate, all of which influence production levels and ultimately, potential profit at the farm gate. However, they are also impacted by ineffective irrigation practices, fertilizer variability, salinity and other soil property issues, as well as the decision-making skills of the farmer. Being able to “scout” for these intra-field issues offers insight not only for yield prediction but also in taking remedial action to address changes as they appear. When combined with GPS technology commonly deployed on farm equipment, such information can help guide the delivery of agricultural inputs to increase yield and the profitability of crops. While UAVs may provide the capacity to deliver such guidance, they also come with some caveats, some of which are discussed in the following section.
3.3. Application and Limitations
The dynamic estimation of the maize crop height was monitored quite accurately by the fixed-wing UAV throughout the growth cycle. While some previous studies have highlighted the limitation of poor point cloud generation by the UAV-SfM technique in landscapes where vegetation was dense and complex (i.e., dead or dry bushes with many overlapping branches in coastal areas) [87
] as well as in maize crop monitoring [89
], the work presented here provides additional insights on the structural variations within deep canopies. Results suggest that the UAV-SfM technique performs reasonably well in terms of reproducing the canopy cover at later stages of the growing season (flowering and maturity) and when complex crop structure and density increase. The comparison of the retrieved DSM, from which the crop height maps were generated, revealed some of the issues with the UAV-based retrievals: especially in reproducing the vegetation at early stages of development. In these cases, the UAV-based SfM approach generates a point cloud that struggles to detect the small-scale plant structure but which is captured by the LiDAR system. While one solution to this would be to reduce the flight altitude of the UAV (to obtain a higher resolution), current technologies still demand a compromise between areal coverage and flying height, due to power and related flight time constraints. In an ongoing (but unrelated) study, a lower flight height (20 m) and higher resolution (0.5 cm) digital surface model was able to distinguish leaves and fruit within a single plant, providing considerable insight on the health and condition of the vegetation. However, the covered area was limited to 0.5 ha due to the low flying speed of the rotary UAV, which was required to generate sufficient overlap between the images at such altitude and due to power supply constraints that kept flights to less than 30 min. The limited area coverage that characterizes rotary UAV versus fixed-wings systems represents an issue for precision agricultural application at larger scales (i.e., commercial scale monitoring). Because of the lower cruise speeds typical of rotary UAVs, an increased flight altitude (relative to that used in this study) would be necessary to cover the same 50 ha agricultural field in a single flight. However, this would also result in a coarser resolution retrieval (i.e., larger GSD) unsuitable for precision agricultural purposes. While LiDAR onboard a UAV could undoubtedly increase the level of detail with which a crop can be retrieved [90
], obtaining crop height (and hence crop status and intra field variability information) using relatively cheap instrumentation with a good level of accuracy remains a priority to encourage broad user-uptake. In this context, to improve the accuracy of the UAV systems retrieval, Harwin et al. [87
] suggested to perform the RGB data collection from multiple points of view. Although this allows a more reliable 3D reconstruction of the dense cloud, it would dramatically increase the time required for the data collection and subsequent processing.
A more generic limitation to the UAV based retrievals is the computational cost involved in image processing. In this study, the time needed to generate a final crop height map can be considerable: up to 2.5 days if an ultra-high resolution of 2.5 cm is required. These times decrease to 24 h, 9 h and 6 h for 5 cm, 10 cm and 20 cm, respectively. It should be noted that the processing undertaken here was achieved using a high-performance server (16 core, 2.4 GHz and 128 Gb RAM server workstation), which delivers solutions at a much faster rate compared to a standard desktop system. For a rapid assessment of the field variability and condition of the crop, an efficient but still accurate methodology is required. Certainly, it is not feasible for a farmer or farm manager to reproduce the type of processing performed herein, especially if they are to manage the multiple fields of a commercial-scale concern. While tuning the resolvable resolution provides some computational relief (by a factor of 7; see Section 2.3
), it remains impractical to suggest this as an operational methodology in a real world scenario.
In providing some practical guidance, the employed SfM approach is able to deliver insights into intra-field crop variability with relatively good accuracy and timing at a 10 cm GSD. While higher resolutions result in improved insights, these come at the cost of an unacceptable processing time. It should be noted that the SfM workflow can deliver a 10 cm resolution product by lowering the accuracy in the point cloud generation (hence reducing the processing time). In practice, the same resolution could be achieved with a UAV flight performed at 500 m altitude but which would be challenging for both safety and operational reasons due to increased risk of wind effects and line of sight restrictions. This represents an important aspect of the fixed-wing UAV-based SfM approach described herein, which can provide precision agricultural solutions as needed and in a timely manner.