UAV Photogrammetric Surveys for Tree Height Estimation
[ad_1]
1. Introduction
Precision agriculture (PA) is defined as a particular agricultural management strategy based on the observations, measurements, and responses of a set of quantitative and qualitative variables affecting agricultural production [1]. The first step in the PA approach is the acquisition and collection of data from different sensors, such as optical multispectral and geophysical ones, to obtain a deep characterization of plants and monitor a crop [2]. Therefore, PA commonly includes a wide range of techniques and methodologies that need to be properly combined to develop increasingly sustainable agricultural management [3,4]. In this regard, the combination of data from different sources is bound to the proper management of the reference system, commonly handled through geographic information systems (GISs), which are highly suitable tools for multidisciplinary studies [5,6]. Among all, geophysics and geomatics techniques represent fundamental support tools for PA applications, giving essential information for cultivation protection and monitoring [7].
In particular, geophysical analyses involve the study of the chemical characteristics of plants and the shallow part of the subsoil, using ad hoc instruments, such as georadar systems [8,9]. On the other hand, geomatic techniques allow obtaining quantitative parameters describing the plants, in terms of height, crown extension, and volume. The use of global navigation satellite systems (GNSSs) in the agricultural field dates back to the 90s, with the development of farm machines equipped with GNSS receivers. However, the use of GNSS instruments to acquire field measurements cannot ensure high-density datasets without resulting in very time-consuming and expensive surveys [10]. Therefore, in this context, they are commonly employed as a support for other techniques. In the early 2000s, drones became the new major players in precision agriculture, being a very useful platform to transport different types of sensors (i.e., LiDAR, multi-spectral, and optical cameras). Indeed, the main advantages of using unmanned aerial vehicles (UAVs) are related to the fast performance and lower costs compared to other geomatic techniques [11,12].
Several studies addressed the use of UAV remotely sensed data for forestry or agricultural applications, mainly relying on LiDAR (light detection and ranging) datasets or multi-spectral imagery [9,13,14,15,16,17,18,19,20]. Moreover, most of the applications are framed in forest environments, where tall and dense trees are present, and the main focus is crown identification and classification [14,15,17,21,22,23,24]. On the other hand, recent studies have been conducted over areas with lower vegetation, using photogrammetric or LiDAR data sources [16,25,26,27,28,29,30]. In particular, tree height extraction exploiting UAV-acquired imagery has been addressed by several researchers, with different specific characteristics of the observed plants, providing promising results [14,21,23,31,32]. Compared with LiDAR point clouds, photogrammetric data are known to be generally noisier and less accurate since they do not provide information about the terrain surface but only the upper layer [14]. Conversely, information representing the ground level can be retrieved from LiDAR data by exploiting multi-return echoes penetrating the canopy [33,34]. This fact commonly ensures a better accuracy of the ground representation, which is crucial for a fair reconstruction of the 3D structure of vegetation [14]. Nevertheless, this benefit is obtained with more expensive instruments, requiring also more advanced carrying platforms.
Whatever the employed technique, analysis in the agricultural field generally relies on raster-based methods, through canopy height models (CHM), or others combining point clouds and raster data [35,36,37]. A CHM represents tree crowns and their heights from the ground; hence, it is commonly obtained as the difference between a digital surface model (DSM) and a related digital terrain model (DTM). The reference DTMs can be generated by exploiting automatic filters to identify the points representing the ground surface. Originally, the algorithms for vegetation filtering were designed for LiDAR datasets, although they can be employed also for photogrammetric applications [24]. Indeed, different software implements automatic filters for the vegetation that yet lead to better performances in forests rather than in areas with low trees [14,25,26,27]. However, fruit farmland or vineyards are characterized by the presence of sparse trees and relatively low heights, ranging approximately from 1 m to 4 m [38]. A similar situation concerns soils contaminated by pollutants or heavy metals undergoing phytoremediation or reclamation by planting trees or shrubs that can improve the chemical and physical characteristics of the soils and reduce the pollution level [39,40]. These latter contexts are commonly characterized by very low vegetation, in the order of half a metre, and height extraction applications in these contexts are poorly available in the literature, except for LiDAR. Some possible limitations of the SfM algorithm for the extraction of the height of low trees are addressed in the study by Matsuura et al. (2023) [41], developing a stereo-matching-based methodology applied for relatively low trees (50–60 cm).
This paper presents the first results of research concerning the use of aerial photogrammetry acquired by low-cost UAVs to retrieve dimensional parameters of the trees, in particular their height. This study belongs to a wider project aimed at finding algorithms and procedures for analysing and monitoring soils contaminated with heavy metals and undergoing phytoremediation processes, based on the definition of dimensional, geophysical, and chemical parameters. The final purpose is to define specific algorithms for different soils, plant species, and microbial populations, allowing the identification of the optimal conditions for the contaminated site’s restoration. To this aim, two study areas, named “Area 1” and “Area 2”, with different characteristics in terms of mean tree height (Area 1: 5 m; Area 2: 0.70 m) are chosen to carry out the geomatic surveys to test the applicability of the procedure even in a challenging context of shorter trees seedlings. Three low-cost UAV campaigns are performed in Area 1, with different flying altitudes to evaluate the impact of the images’ resolution (ground sampling distance—GSD), while only one UAV survey is performed in Area 2. Starting from the photogrammetric datasets, the implemented workflow involves the elaboration of the reference digital terrain model (DTM), the dense point clouds, and the digital surface models (DSMs) and a GIS-based procedure for the extraction of tree heights based on a local maxima approach. Finally, the accuracy of the tree heights is assessed through a statistical analysis exploiting the availability of field measurements.
3. Results
The main statistics of the residual values between extracted and measured heights for each flight are reported in Table 5.
Concerning Area 1, the results are within the expected method’s accuracy obtained for higher trees or in forestry contexts since the mean absolute values of the residuals range between 14 cm and 34 cm. As expected, the lower the flying altitude, i.e., the lower the GSD, the better coherence we found with the field measurements, from Flights 1 to 3. In particular, Flight 3 (30 m of flight altitude) shows very promising results, with 90% of the residuals below 50 cm. This result was also unexpected due to the high values of the RMSE related to the georeferencing process (Table 3). Nevertheless, a difference between the three flights arises since the former tend to underestimate the tree heights, while the third slightly overestimates their values (considering the field measurements as a reference in the comparison, as in Equation (2)). This fact is also evident from the frequency distribution histograms of the residuals in Figure 6a and should be further investigated to understand possible sources of this behaviour. Although Flight 3 involved a slightly smaller area compared with the previous ones, with a total of 41 identified trees, we can observe an almost normal distribution of the differences close to zero-centred. Finally, the relation between the residuals and the corresponding “real” height value (from the field measurement) has been analysed through the parameter. Very similar values are found for Flights 1 and 2, equal to 8.7% and 8.8%, respectively, whereas for Flight 3, is equal to 5.5%. These outcomes confirm the good results for the first site, meaning that the errors, i.e., the residuals in the height determination, are low compared to the absolute tree heights. This is further proved by the RMSE values, resulting in 0.47 m, 0.51 m, and 0.37 m, for Flights 1, 2, and 3, respectively.
The situation is different in Area 2, where Table 5 and Figure 6b show that values are entirely shifted toward the negatives, with a mean value of about minus 60 cm, i.e., nearly 30 cm more than the highest flight on Area 1 (Flight 1). This fact was expected from a first graphical inspection of the dense point cloud where plants were barely visible, mainly due to their very low heights, which affected the whole procedure. As a first aspect, independently of the following elaborations, it was difficult to obtain a good representation of shorter tree seedlings from the photogrammetric dense point cloud, considering the limited extension of the canopies and the foliage. Secondly, low heights negatively impact vegetation filtering, which cannot exploit automatic tools but rather requires a manual procedure. As a final point, comparing DSMs and DTMs, both of which derive from interpolation processes entraining sources of uncertainty, requires proper accuracies to enable small variations to be detected. Reasonably, these aspects mutually influence each other as well as the plant heights that are both the origin of the problem and the reason why the method is unusable in this context. Indeed, even if the field measurements are known to have accuracies lower than the cm-level, the obtained differences are comparable or sometimes higher than the actual tree heights in the area. Moreover, the same outcome is confirmed by the RMSE value, equal to 0.67 m.
Figure 7 shows the box and whisker plots of the estimated and measured tree heights for the three flights in Area 1. The minimum, maximum, and median values of the two datasets are highlighted, showing the higher variability in the estimated values compared to the measured ones as well as the underestimation trend for the first two flights, which is inverted in the case of Flight 3. Note that since the statistics related to the residuals already revealed the behaviour in Area 2, we decided not to show box charts for the second site.
The relationship between the estimated and measured tree heights for the three flights performed in Area 1 is presented in Figure 8. Looking at the linear regression lines, we can say that overall, the models, i.e., the estimated height values, are able to replicate the observed samples. Flight 1 and Flight 3 exhibit similar coefficients of determination (R2 = 0.67; R2 = 0.65), whereas the value is slightly lower for Flight 2 (R2 = 0.50). Similar considerations of the previous charts are made for Figure 8 relating to Area 2.
4. Discussion and Conclusions
To date, the monitoring and management of agricultural lands is a topic of high interest in the context of environmental protection and precision agriculture (PA), also due to the European Common Agricultural Policy (CAP) [57,58]. Indeed, there is a growing development in related research where geomatic surveys play a fundamental role.
This study focuses on the use of photogrammetric data to retrieve dimensional parameters of trees. The preliminary results of the tree height estimation from imagery acquired with low-cost UAV platforms are presented. The proposed methodology is based on photogrammetric processing exploiting the SfM technique, coupled with a GIS-based analysis. This analysis is straightforward and mainly based on open-source software, such as CloudCompare and Qgis, making again the whole procedure very flexible. Two study areas were chosen, named “Area 1” and “Area 2”, and a total of four UAV flights were performed. Starting from the UAV dense point clouds, the reference DTMs were produced by applying automatic filtering procedures or implementing a manual classification of ground and vegetation. The canopy height models obtained as differences between each DSM and the related DTM were analysed in the GIS environment using a local maxima approach choosing ad hoc regions as individual investigation areas around each tree. Thus, the candidate tree tops, corresponding to the maximum CHM pixel values, were identified and the related heights were considered in the analysis. The availability of in-field measurements concurrent with the UAV surveys supported the validation process and the statistical analysis, allowing for the method’s accuracy assessment.
In Area 1, the comparison with the field measurements provided promising results, with mean absolute residuals ranging between 14 and 34 cm. The best values are related to the campaign with 30 m of flying altitude, proving that the image resolution (GSD) is a fundamental parameter to obtain higher quality in the photogrammetric products. This fact is also supported by other studies, such as [22,29]. In particular, Pourreza et al. [29] used a DJI Phantom 4 flying at different altitudes (25 m, 50 m, and 100 m) over an area with tree heights comparable to ours, confirming that accuracy is negatively dependent on the flight altitudes. Moreover, all the flights in Area 1 exhibit 50% of the residuals lower than 38 cm, and the frequency distribution histograms of the residuals follow almost normal distributions. Results in the first site are also coherent with the outcomes of a similar study by Zarco-Tejada et al. [28], where they obtained an RMSE of 35 cm for tree heights ranging between 1.16 m and 4.38 m. Concerning the coefficients of determination related to the tree height estimation, their higher value (R2 = 0.83 versus our R2 = 0.67) could be related to the difference in the field measurements used for the validation. Indeed, we are aware that the use of a metric rod inevitably suffers from the weather conditions, in particular the wind, the tree characteristics in terms of foliage and the crown’s extension and the subjective nature affecting the observations. Birdal et al. [26] found an RMSE equal to 28 cm for trees ranging between 1.2 m and 7.1 m, which is also consistent with our findings, whereas their R2 parameter is significantly higher (0.95). Similar results can be found in [14,21,23,31], whose findings outline RMSE and R2 values comparable to ours in contexts of higher trees. However, results of the different flights in the first site exhibit different behaviours, since Flights 1 and 2 generally underestimate the trees’ heights, while Flight 3 tends to overestimate. Since results of other studies generally confirm the underestimation trend of UAV extracted heights [21,29], this fact requires further investigation, which will be developed in future studies.
As expected, the context in Area 2 entailed different conclusions since the mean value of the residuals is about −60 cm, i.e., sometimes higher than the actual tree heights in the area. Even if the flight altitude was comparable with the one of Flight 1, the mean value of the residuals is 30 cm higher. This fact is reasonably not only due to the altitude rather than to the “absolute” height of the individual trees that made the whole procedure very challenging. Indeed, we want again to stress the complexity of the context chosen in the case of Area 2, where tree heights are always below 1.5 m and the foliage is poorly grown. We identified three main aspects impacting this outcome: (i) goodness of the photogrammetric dense cloud in representing shorter tree seedlings; (ii) manual vegetation filtering; and (iii) impact of the interpolation in the DSM and DTM generation and differencing. Although the obtained results make the methodology not usable for this range of heights, bearing in mind the presence of different plant species in Area 2, we tried to relate the results with this parameter. However, no evidence has arisen, unless a slightly better behaviour of the lentisk compared with the other two species (poplar, oleander), but probably reliable results would require higher plants to allow their differentiation by foliage and crown extensions.
Overall, the obtained results confirm the suitability of UAV photogrammetric data acquired from low-cost instruments for tree height extraction. The implementation of the same procedure to multitemporal datasets acquired at low-flight altitudes could allow even the determination of the rate of growth of trees over medium–long temporal scales. These become key data in the analysis of phytoremediation processes when assessing the plant’s health and the effectiveness of the chosen restoration method. Moreover, since the usage of low-cost UAV equipment is easily accessible even for non-specialist users, the possibility of easily adopting the same workflow represents a benefit for all those involved in the agricultural field. Nevertheless, the application of the same methodology in the context of shorter tree seedlings requires further investigation, and possible future development of this research may involve UAV campaigns at lower flying altitudes and an intermediate case study of Area 2 after the plants’ growth.
Finally, it should be highlighted that this research is framed in the context of a wider project, “Tecnologie di CARatterizzazione Monitoraggio e Analisi per il ripristino e la bonifica (CARMA)”, involving both geomatics and geophysical data [59]. Thus, the presented study is the first step to reaching the project’s goals of integrating geometric, biochemical, and geophysical data in a single workflow to assess tree health and rates of growth. Considering this multidisciplinary context, future developments may involve the employment of more advanced technologies, such as LiDAR, to improve the accuracy, especially in areas with shorter tree seedlings. In addition, future analysis will exploit both nadiral and oblique images, as suggested in [26,29]. Indeed, the use of oblique images leads to improvements in the final accuracy since they allow the point cloud to better capture the ground points under and around trees, resulting in enhanced classification processes.
[ad_2]
