Grassland covers roughly 40% of the total world land area and this extension corresponds to approximately 52.5 million km2
]. Moreover, grassland ecosystems have extremely variable features, since they are strongly influenced by environmental conditions, such as topographic location, and by anthropogenic effects, such as (in)organic fertilizer application [2
]. Therefore, efficient management plans of grasslands require insight into the health status and spatial variation in order to improve conservation and to achieve optimal growth. The planning of grassland management is even more important if investigated at the field scale, since the farmers need a guide for identifying the optimal time for fertilizer application and for predicting the ideal time of harvest [3
]. Up until now, the schedule of grassland management was mainly organized according to qualitative information deducted from farmers’ experiences. The integration of quantitative and spatial data into the already existing management plans could significantly improve the health status of grasslands and the feeding quality of the final harvest.
Currently, traditional field-based methods are usually preferred, even if their application is strongly restricted because they are time-consuming, destructive, and cannot be repeated often enough for high spatial resolution investigations over large areas [3
]. Over the years, alternative techniques have been developed, and, nowadays, remote sensing is recognized as a suitable technique thanks to its ability to characterize the land surface in a fast and relatively cheap way. For these reasons, it has been widely used for estimating various biophysical and biochemical vegetation variables in grasslands, such as the Leaf Area Index (LAI) or chlorophyll [5
]. The subsequent introduction of hyperspectral sensors has allowed researchers to improve the retrieval of grassland traits considerably [3
]. Based on its narrow contiguous wavebands, hyperspectral sensors are more sensitive to vegetation variables since they provide a continuous reflectance spectrum of the vegetation target [4
]. Nevertheless, the hyperspectral images include hundreds of spectral bands, many of which are strongly correlated. Consequently, it has been necessary to select a subset of data in order to decrease the dimensionality of the dataset and reduce redundant information. Therefore, the selection of the subsets can be carried out by taking into account the sensitivity of the vegetation variables to the spectral bands [17
Further improvements could be achieved by adapting the spatial resolution to the size of the object investigated. In this way, the image captures the spatial detail needed to resolve important patterns in the field. Given the fine spatial scale of grassland biophysical traits, Unmanned Aerial Vehicles (UAV) can provide the necessary fine spatial scales because they can be flown at a low altitude. In order to analyze biophysical traits of grassland plants, it is necessary to have the smallest possible resolution. Therefore, the application of Unmanned Aerial Vehicles (UAV) could provide a useful support, as they permit us to drastically reduce altitude in order to increase the spatial resolution and determine it based on the object size under examination and the aim of the research [18
]. Furthermore, UAVs can provide improved access to more remote targets and can be deployed flexibly, eliminating problems due to clouds. For these reasons, the utility of hyperspectral cameras mounted on UAVs has been explored in a few papers [21
]. Indeed, although the application of hyperspectral imaging and of UAVs is significantly widespread for quantitatively the Earth system, their combination is still rather limited.
Two different statistical approaches are commonly used to analyze the relationship between spectral measurements and vegetation traits: univariate and multivariate regression models. The first approach is mainly based on a regression model between a so-called spectral vegetation index, a limited set of combined spectral bands, and the vegetation trait of interest, while the multivariate approach is based on the application of regression modeling between all observed spectra variables and a vegetation trait [22
Depending on the spectral resolution of the sensor applied, the calculation of vegetation indices can be based on both narrow or broad spectral bands, obtaining different values for the same index. Nevertheless, when comparing the results of the two different approaches for computing the same index, some studies have shown that the use of narrow-band indices results in improved statistical relationships between biochemical and biophysical traits and vegetation indices [15
]. Moreover, earlier studies have shown that narrow-band information of hyperspectral images is important for providing information for the estimation of biochemical and biophysical vegetation variables by applying the vegetation indices [14
]. However, in this way, it is not possible to fully exploit the hyperspectral data potential of the large number of spectral bands available. For this reason, numerous researchers have focused their attention on Stepwise Multiple Linear Regression (SMLR), which takes advantage of almost all of the hyperspectral information available [7
]. However, SMLR suffers from multi-collinearity problems and the extensive spectral overlap of individual biochemical properties [35
]. Partial least squares regression (PLSR) has been recognized as an alternative technique to SMLR, and it has been used earlier in the spectral data analysis for wheat [8
] and heterogeneous grasslands [36
Several studies have dealt with validity and accuracy in the analysis of grasslands using the two statistical methods described. However, no studies have been published on the evaluation of these techniques for estimating grassland variables relevant for forage quality assessment, such as crude ash or metabolic energy. In addition, grassland traits such as sodium are also relevant to assess the taste for feeding cattle. Thus, the utility of the two statistical methods for estimating biochemical traits in grasslands from spectral datasets and the potential influence of the type and amount of fertilization on these grasslands needs investigation.
The overall objective of the present study is to investigate the utility of hyperspectral images acquired using an unmanned airborne platform for predicting vegetation traits in grasslands. Specific sub-objectives can be summarized as follows:
Compare two regression methods based on vegetation indices and the PLSR approach in order to choose the best strategy for predicting bio-physical and bio-chemical plant traits of grasslands;
Investigate the influence of the amount and the type of fertilization on grassland traits and the resulting spectral curve;
Evaluate the influence of the phenology of grasslands and the timing of spectral data collection on the established regression relations.
2. State of the Art of UAVs Applications
In the last few years, UAV applications in the civil field have become increasingly popular. Their use is becoming widespread in many scientific disciplines, such as earth observation and precision agriculture [38
], in which high temporal frequency and detailed spatial resolution are necessary in order to improve the health status of soil and crops [40
This has resulted in the development of suitable instrumentation for such purposes. With the recent technical electronic and optical improvements for aerial platforms and the devices mounted on them, the field has developed fast. Indeed, thanks to the evolution of the integrated circuits and radio-controlled systems, it is possible to adopt the use of UAVs [43
]. They can fly at lower flight altitudes than traditional airborne platforms, allowing an increase in spatial resolution, reaching difficult-to-access areas and acquiring images in flexible data, eliminating cloud cover problems. Moreover, their large market penetration and continuous development have led to a drastic reduction in their cost [19
Furthermore, the size and weight of electronic devices, used for capturing images, have been modified, and strongly decreased overall, so that they can be mounted on the UAVs [44
]. Also, the introduction of small digital multispectral sensors has further allowed us to expand the areas investigated and to improve the results obtained [45
], as shown in several previous studies for grassland monitoring.
After comparing and assessing the potential of several multispectral sensors and small RGB digital cameras, the authors of [46
] has shown the limitations of RGB cameras and the superiority of multispectral sensors in grassland traits estimation. In the same way, the authors of [47
] has shown that there is a strong relationship between vegetation indices, calculated from multispectral images, and grassland bio-physical parameters, useful for temporal changes and, consequently, future crop growth monitoring. Subsequently, the introduction and the miniaturization of hyperspectral sensors has allowed us to obtain further improvements in grassland analysis [21
For these reasons, various studies have applied the combination of UAVs and hyperspectral sensors in grassland mapping and monitoring and have documented its great potential for detecting water stress [48
] and for estimating grassland biophysical traits [14
In this paper, the best strategy for estimating structural and biochemical traits of grasslands from UAV-acquired hyperspectral images and the study of the type and the amount of the fertilization influence on their prediction capacity was explored. The application of nitrogen (N) to grassland is one of the major management tools which farmers have to influence yield and quality. Currently, there are knowledge gaps in how organic fertilizer application affects N balances in grasslands. Especially weather conditions can influence nitrogen mineralization of the organic N of slurry and thus influence the long-term efficiency of slurry application [69
]. Therefore, to study this long-term effect of slurry, a grassland experiment was planned by applying different amounts of slurry for the years of 2012, 2013, and 2014. The level of fertilizer applied was chosen based on the customs of German farmers. Moreover, this data combination seemed especially eligible for studying the best statistical strategy for the estimation of structural and biochemical traits from UAV-acquired hyperspectral images as the different levels of N fertilization provided a wide, well-defined range of both biomass yield and N concentration at every flying date.
The hyperspectral images of an experimental grassland field at the farm Haus Riswick, near Kleve in Germany (Figure 1
), were taken by applying an octocopter UAV, equipped with the Wageningen UR Hyperspectral Mapping System (HYMSY) (Figure 3
]. Two different flight campaigns were carried out, one on 15 May 2014 and the other on 14 October 2014. Following from the flight campaigns, the structural and biochemical grassland traits analysis was also carried out.
The results of the grassland traits analysis in both campaigns showed different values for each of the parameters under investigation, but similar statistical relationships (Table 2
, Table 3
and Table 4
). The Pearson correlation among the structural traits (height, fresh and dry matter yield) is significantly high (>0.89), while among the biochemical traits, the correlation is substantially different depending on the feature considered (Table 2
): for example, metabolic energy and crude fiber are negatively correlated (−0.98), while metabolic energy and crude protein show a very low correlation (0.04). Indeed, the metabolic energy has not been directly measured but it has been calculated by applying Equations 1 and 2, and surely this influences the correlation. On the contrary, the correlation between crude protein and crude fiber is quite low (0.10). Normally, when the analysis is carried out on a growing plant at subsequent dates, they are negatively correlated [70
]. In this case, however, the grassland was only analyzed at one harvest date and, therefore, it is not possible to find a good correlation among them. This is also confirmed by the analysis of the N fertilization influence. Indeed, N fertilization increases crude protein concentration while normally it has little effect on the crude fiber concentration since crude fiber depends mainly on temperature, day length, and light intensity (Figure 4
The use of a UAV, applied to adjust the spatial resolution to the size of the grassland field investigated, obtaining a GSD of 200 mm for the hyperspectral image, due to the reduced altitude (70 m) and speed (5 m/s) compared with the traditional quotas and speed of the airplane [18
]. Moreover, it also permits us to acquire images in flexible dates, exploiting the same sunny illumination conditions and eliminating problems due to clouds, in such a way that the imagery of two different periods can be compared. In addition, the possibility to mount a hyperspectral camera on it has allowed us also to fully take advantage of all the potentials of remote sensing, processing hundreds of narrow contiguous spectral bands to which the vegetation is sensitive [4
]. Indeed, in general, quite homogeneous grassland vegetation cover has relatively small variations in reflectance properties of the canopy, which could be visible and analyzed only from a continuous reflectance spectrum. For these reasons, the field under investigation was analyzed by applying hyperspectral data. From this detailed spectral data, the variation in plant traits of grassland, caused by the type and the amount of fertilization (Figure 5
), correspond to a relative small variability in reflectance values (Figure 6
). Thus, in order to identify this variation, the narrow-band vegetation indices were calculated instead of broad vegetation indices, as shown in previous studies [14
]. Narrow-band vegetation indices improve the results in processing the hyperspectral images since they allow us to exploit all of the information contained in the spectrum [14
The selected narrow-band vegetation indices (Table 1
) were computed separately on the three sub-datasets in which the original dataset was split in order to understand the influence of the type and the amount of fertilizer and the weight of the grassland growth status. Results showed that, although narrow vegetation indices detect the effects of the type and the amount of fertilization, their prediction ability is moderate. Indeed, the results of the two integrated sub-datasets, including both May and October data, but one characterized only by inorganic fertilization and the other by both organic and inorganic fertilization, are comparable. On the contrary, grassland health status influences the estimation ability of this technique (Table 6
). This is perfectly in line with the results of Table 5
. Indeed, the best results (higher correlation values) were obtained using a linear regression model on the data related to the more productive part of the season in May. This is justified by the distribution of the two datasets. Indeed, the distribution of data acquired during the two campaigns does not follow a specific linear relation, but it forms two separate clusters. For that reason, it is not possible to fix the best narrow vegetation index, since it depends on the grassland health status and the trait under analysis. In particular, the best estimator for grassland structural traits is MCARI/OSAVI [57
], which shows a value > 0.5, while for the most biochemical features, it is CIred-edge
] (Figure 8
). Moreover, this approach shows promising results for estimating grassland structural traits, but not biochemical features, with the exception of metabolic energy. The UAV-based set-up allows the delivery of spatial maps showing the spatial distribution of grassland traits following the application of this statistical approach (Figure 9
and Figure 10
Like the narrow vegetation indices, the PLSR method was implemented on the same three sub-datasets, in order to analyze the consequence of the type and the amount of the fertilization and the influence of the grassland growing season. Also, in this case, results showed that the estimation ability of this approach is mainly influenced by the grassland growing season and not by the different type and the amount of fertilization. Indeed, the results related to the two different populations including both May and October, one with only the inorganic fertilization and the other with both organic and inorganic fertilization, are comparable. Instead, comparing one of them with the sub-dataset related to the combined data of the inorganic and organic fertilization data of the May survey, it is possible to see that the method performs better with the integrated sub-dataset of May and October, with the exception of crude ash, Na, and K. Indeed, in the integrated sub-dataset of May and October, including both inorganic and organic fertilization, the R2
values are good (>0.7) for all characteristics, except for crude ash (0.4), Na (0.2), and K (0.3) (Table 7
). Also, the value of the RPD is good (Table 7
) while, in the sub-dataset of May including both inorganic and organic fertilization, the R2
values are quite high for all traits, and regarding the RPD, only K has a value higher than 3 (Table 7
). Probably, the higher coefficient of determination of the May-October group is related to a higher number of LVs, which could be a sign for overfitting. In addition, the leave-one-out cross-validation approach as adopted in this study cannot be treated as a completely independent assessment of prediction quality. In order to generate a more robust validation and representative error estimates, a larger dataset is necessary so that a training and independent test dataset can be used to test the PLSR approach, as suggested in recent studies [67
]. Therefore, for future application of the described approach, an additional evaluation using an independent test set would be required to assess the use of PLSR in practice. Moreover, other future applications could contemplate the PLSR use for integrating different vegetation indices over the whole spectrum in order to evaluate their performance and to compare them with the performance of the vegetation indices computed on pure spectral reflectance values.
From the above results, the best strategy for detecting structural and biochemical features is hyperspectral remote sensing in combination with PLSR. The outcome from this study is in agreement with the results of previous research studies [3
]. For example, [3
] compared a vegetation index approach with PLSR for estimating fresh grassland biomass. In their study, PLSR gave a better retrieval of fresh biomass with a RMSECV of 2.10, slightly better compared to the findings from this study (Table 7
). However, for biochemical grassland traits, this finding gives new opportunities to map grassland quality before harvest, which is one of the requirements for the future adoption of precision farming practices in grassland management [3
In order to confirm the results of this paper and achieve a schedule of precision grassland management, the approach presented in this study should be tested on other fields of different size and geographic position. In addition, more additional points need to be investigated. For example, a possible future development could consider more observations over the growing season, since this study has taken into account only the beginning and the end of the growth.