Content uploaded by Piero Toscano
Author content
All content in this area was uploaded by Piero Toscano on Aug 14, 2015
Content may be subject to copyright.
Remote Sens. 2015, 7, 2971-2990; doi:10.3390/rs70302971
remote sensing
ISSN 2072-4292
www.mdpi.com/journal/remotesensing
Article
Intercomparison of UAV, Aircraft and Satellite Remote Sensing
Platforms for Precision Viticulture
Alessandro Matese
1,
*, Piero Toscano
1
, Salvatore Filippo Di Gennaro
1,2
, Lorenzo Genesio
1
,
Francesco Primo Vaccari
1
, Jacopo Primicerio
1,3
, Claudio Belli
4
, Alessandro Zaldei
1
,
Roberto Bianconi
4
and Beniamino Gioli
1
1
IBIMET CNR–Istituto di Biometeorologia, Consiglio Nazionale delle Ricerche, via G. Caproni 8,
50145 Firenze, Italy; E-Mails: p.toscano@ibimet.cnr.it (P.T.);
f.digennaro@ibimet.cnr.it (S.F.D.G.); l.genesio@ibimet.cnr.it (L.G.);
f.vaccari@ibimet.cnr.it (F.P.V.); j.primicerio@ibimet.cnr.it (J.P.);
a.zaldei@ibimet.cnr.it (A.Z.); b.gioli@ibimet.cnr.it (B.G.)
2
DSAA-Dipartimento di Scienze Agrarie, Alimentari e Ambientali, Università di Perugia, Borgo XX
Giugno 7, 06123 Perugia, Italy
3
Dipartimento di Scienze Agrarie, Forestali e Agroalimentari, Università di Torino,
Via Leonardo Da Vinci 44, 10095 Grugliasco, Italy
4
Terrasystem s.r.l., Via Pacinotti, 5, 01100 Viterbo, Italy;
E-Mails: c.belli@terrasystem.it (C.B.); r.bianconi@terrasystem.it (R.B.)
* Author to whom correspondence should be addressed; E-Mail: a.matese@ibimet.cnr.it;
Tel.: +39-055-303-3711; Fax: +39-055-308-910.
Academic Editors: Georg Bareth, Pablo J. Zarco-Tejada, Clement Atzberger and Prasad S. Thenkabail
Received: 14 November 2014 / Accepted: 17 February 2015 / Published: 13 March 2015
Abstract: Precision Viticulture is experiencing substantial growth thanks to the availability of
improved and cost-effective instruments and methodologies for data acquisition and
analysis, such as Unmanned Aerial Vehicles (UAV), that demonstrated to compete with
traditional acquisition platforms, such as satellite and aircraft, due to low operational costs,
high operational flexibility and high spatial resolution of imagery. In order to optimize the
use of these technologies for precision viticulture, their technical, scientific and economic
performances need to be assessed. The aim of this work is to compare NDVI surveys
performed with UAV, aircraft and satellite, to assess the capability of each platform to
represent the intra-vineyard vegetation spatial variability. NDVI images of two Italian
vineyards were acquired simultaneously from different multi-spectral sensors onboard the
OPEN ACCESS
Remote Sens. 2015, 7 2972
three platforms, and a spatial statistical framework was used to assess their degree of
similarity. Moreover, the pros and cons of each technique were also assessed performing a
cost analysis as a function of the scale of application. Results indicate that the different
platforms provide comparable results in vineyards characterized by coarse vegetation
gradients and large vegetation clusters. On the contrary, in more heterogeneous vineyards,
low-resolution images fail in representing part of the intra-vineyard variability. The cost
analysis showed that the adoption of UAV platform is advantageous for small areas and that
a break-even point exists above five hectares; above such threshold, airborne and then
satellite have lower imagery cost.
Keywords: precision agriculture; Unmanned Aerial Vehicle (UAV); remote sensing
1. Introduction
Precision Agriculture (PA) could be defined as the site specific management of crops heterogeneity
both at time- and spatial-scale [1] in order to enhance the efficiency of agricultural inputs to increase
yields, quality and sustainability of productions. Precision Viticulture (PV) falls in the area of PA and
aims at [2]: identifying within a degree of stability the inter-annual spatial variation of the grape yields
and quality; identifying which are the causes that determine such variability and if they are related to
some site specific management practices. For these reasons, PA and PV approaches take advantage of
those technologies that are able to detect with high accuracy the spatial heterogeneity of vineyards that
is driven by several intrinsic factors (soil, crop management, irrigation, vineyard nutritional state, pest
and disease control), and external variables (climate), and that determine the inter-annual and
intra-vineyard variability of yield and quality. Some new instruments have already demonstrated to be
suitable for PV. The Unmanned Aerial Vehicle (UAV) remote sensing platforms are among the
technologies that have been recently applied to remote sensing of vegetated areas [3–5] and applied to
PV [6–8], proving a high flexibility of use, low operational costs and very high spatial resolution [9],
down to 1 cm.
In parallel, traditional remote sensing technologies based on satellite and aircraft platform, are
continuously improving in terms of spatial and temporal resolution, thus enhancing their suitability for
PV applications. Each of these technologies has pros and cons that involve technological, operational
and economic factors. Satellite surveys can map large areas at the same time, but on the other hand still
have coarse resolution for PV, and may suffer from cloud cover and from constraints in relating imagery
timing to specific phenologic phases because of the fixed-timing acquisitions. Aircraft surveys can be
planned more flexibly, but can pose difficult and costly campaign organization efforts [10]. UAVs are
well suited for small scale and research applications, while their limited payload and short flight
endurance still remain areas of weakness for their wide scale implementation in PV. These factors pose
a “scale dilemma”, making the identification of the most effective technology strictly dependent on the
spatial scale and the purpose of the survey, and calls for an improved assessment of technical, scientific
and economical performances of the different remote sensing platforms to assess their optimal
operational context. In all operational-oriented studies, a cost comparison between different technological
Remote Sens. 2015, 7 2973
solutions is of vital importance to define for each of them the cost/effectiveness range of application and
their respective limits of convenience.
The comparison of data with different native resolution involves the application of spatial statistics,
and requires tackling the problem of spatial autocorrelation. All maps display spatial autocorrelation,
needing dedicated statistics that take this into account by adopting a spatial lag, in analogy to the time
lag in time series analysis. Furthermore, although methods are becoming available to compare maps
accounting for the spatial structures present in the data, the most practiced procedures still rely on
cell-by-cell evaluations.
In this paper we deployed simultaneous UAV and aircraft NDVI surveys and quasi-simultaneous
RapidEye NDVI satellite images, acquired over two vineyards in Italy, to assess the capability of each
system to represent the intra-vineyard vegetation patterns, to evaluate the similarities of images taken at
different spatial resolutions and to perform a pros and cons evaluation that combines operational and
economic factors. The final outcome of this assessment is the development of a logical framework with
the aim of providing guidelines for the choice of the appropriate detection platform as a function of the
scale of analysis in PV.
2. Materials and Methods
2.1. Experimental Site
Two vineyards, hereafter referred as V1 (45°31′02′′N, 12°31′01′′E) and V2 (45°43′05′′N, 12°32′10′′E)
were chosen as test sites in the Veneto Region alluvial plain (Italy). The two vineyards have similar
extension (2.5 ha) and the same agronomic characteristics. Cabernet Sauvignon (Vitis Vinifera L.) vines,
grafted on 420A rootstock, are trained to free cordon with a single horizontal wire 1.5 m high and
downward shoots. Vines spacing is 2.5 × 1.3 m between rows and plants, respectively, while the row
orientation is North-South with flat topography. Climatic characterization for the period 1996–2013
made use of data collected by a nearby agrometeorological station (45°43′05′′N, 12°28′46′′E). The study
was performed in summer 2012, one of the warmest of the long-term period and second only to 2003,
with mean temperatures 1.5 °C higher than the historical average (June–August), and a lower cumulated
rainfall (90 mm compared to 230 mm average).
2.2. Remote Sensing Platforms
Three different remote sensing platforms were employed to map the NDVI vegetation index at the
two sites (Table 1).
2.2.1. UAV Images
A flight campaign was made on 18 September 2012 using a UAV platform, based on a modified
multi-rotor Mikrokopter OktoXL (HiSystems GmbH, Moomerland, Germany) able to fly by remote
control or autonomously with the aid of its Global Position System (GPS) receiver and its waypoint
navigation system. The sensor utilized to acquire UAV multispectral images was a Tetracam ADC Lite
(Tetracam Inc., Chatsworth, CA, USA), described in detail in Table 1. All images were taken between
12:00 and 13:00 in clear sky condition, and a white reference image to compute reflectance was taken
Remote Sens. 2015, 7 2974
by framing a Teflon calibration panel just before the flight. The flight altitude has been fixed at 150 m
(AGL), with a UAV flight speed of 4 m/s. Those settings allowed a 72% image forward overlap, while
a waypoints route planned ad hoc ensured a 40% image side overlap, high enough to guarantee an
optimal photogrammetric processing.
Table 1. Remote sensing platforms.
UAV AIRCRAFT SATELLITE
Platform Mikrokopter OktoXL Sky Arrow 650 TC/P68 RapidEye
Camera
Tetracam ADC Lite
ASPIS
REIS
Number of channels 3 12 5
Spectral wavebands
520–600 nm
630–690 nm
760–900 nm
415–425 nm
526–536 nm
545–555 nm
565–575 nm
695–705 nm
710–720 nm
745–755 nm
490–510 nm
670–690 nm
770–790 nm
790–810 nm
890–910 nm
440–510 nm
520–590 nm
630–685 nm
690–730 nm
760–850 nm
Dimension 114 × 77 × 22 mm 270 × 250 × 200 mm 656 × 361 × 824 mm
Weight 0.2 kg 10 kg 62 kg
Resolution 2048 × 1536 pixel 2048 × 2048 pixel
12000 pixel linear
CCD per band
Pixel size 3.2 μm 7.4 μm 6.5 µm
Focal length 8.5 mm 12 mm 633 mm
FOV 42.5° × 32.5° 12.5° × 12.5° 15.7° × 10.5°
Output data 10 bit RAW 8 bit RAW 16 bit NITF
Image size 6 MB 4 MB
462 MB/25 km along track for
5 bands.
Flight quote AGL 150 m 2300 m 630 km
Flight speed 4 m/s 90 knot -
Ground resolution 0.05 m/pixel 0.5 m/pixel 5 m/pixel
Ground image
dimension
116.5 × 87.5 m 1024 × 1024 m 77 × 45 km
Total frames 100 2 1
Remote Sens. 2015, 7 2975
PixelWrench2 software (Tetracam Inc., Chatsworth, CA, USA) was used to manage and process ADC
images, providing a batch file conversion from RAW to TIF. Ortho-rectification of the images was
performed by means of a 5 m resolution digital elevation model (DEM). Afterwards, the captured images
were assembled into a mosaic by Autopano Pro 3.6 Software (Kolor SARL, Challes-les-Eaux, France).
Coordinates of the 50 PVC white panels (0.25 × 0.25 m) randomly located inside each vineyard were
measured with a high-resolution (0.02 m) differential GPS Leica GS09 GNSS (Leica Geosystems A.G.,
Corporate Legal Services, Heerbrugg, Switzerland) to georeference the images. The QGIS software
(Quantum GIS Development Team 2014, Quantum GIS Geographic Information System, Open Source
Geospatial Foundation Project, http://qgis.osgeo.org) was used to carry out this task, utilizing ground
referenced panels and a set of ortho-photos with a ground resolution of 0.5 m. A FieldSpec Pro
spectroradiometer (ASD Inc., Boulder, CO, USA) was utilized to perform a radiometric calibration in
field as described by Primicerio et al. [6], so each pixel DN (digital number) was converted first into
spectral radiance and then into reflectance as described in Goward et al. [11].
2.2.2. Aircraft Images
The aerial data of the two vineyards were acquired on the same day as UAV, in a single swipe with a
Sky Arrow ERA platform [12] at a flight altitude of 2300 m above ground level, corresponding to
a 0.5 m spatial resolution. The aircraft was equipped with the ASPIS (Advanced SPectroscopic Imaging
System) remote sensing system [13] (Table 1), coupled with a Systron Donner C MIGITS III INS/GPS
unit (Systron Donner Inertial, Concord, MA, USA) and a Riegl LD90 series laser altimeter (RIEGL
Laser Measurement Systems GmbH, Horn, Austria). Spectral bands of red and near infrared were
processed in order to calculate NDVI vegetation index. Radiometric correction to the sensor was applied
by means of the proprietary software of Terrasystem srl. An atmospheric correction was carried out using
the ENVI FLAASH module (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes, ITT
Visual Information Solutions, USA), an algorithm developed by Spectral Sciences, Inc. (Burlington,
MA, USA). Geometric correction, which is necessary to eliminate the internal optical distortions of the
sensor and those caused by the altitude, was performed using the software PCI Geomatica (PCI
Geomatics Corporate, ON, Canada), through a methodology that envisage the acquisition of ground
control points on georeferenced high-resolution images. An aerial model was used as orthorectification
algorithm for the aerial data. The aerial orthoimage (a mosaic of two frames) has been returned at 0.50
m ground resolution, and georeferenced in the WGS 84-UTM 32 North reference system.
2.2.3. Satellite Images
A multispectral image acquired on 15 September 2012 (at 11.03 am) was provided by BlackBridge
from the RapidEye archive. RapidEye is a constellation of five satellites that acquire multispectral data
at a spatial resolution of 6.5 m, resampled to 5 m pixel size, in the range of the visible and near infrared.
Images are provided in NITF 16-bit format, while temporal resolution is five to six days for nadir data
and one day for the off-nadir. Features of the RapidEye’s spectral bands and technical specification are
presented in Table 1. RapidEye Level 2 product embeds radiometric correction natively. As for aerial
data, the atmospheric correction was carried out using the ENVI FLAASH module, and the geometric
correction with PCI Geomatica software. We used the rational function orthorectification algorithm,
Remote Sens. 2015, 7 2976
while the orthoimage (single frame) has been returned at 5 m ground resolution and then georeferenced
in the WGS 84-UTM 32 North reference system.
2.3. Statistical Framework
Multispectral images acquired by UAV, aircraft and satellite platforms, were elaborated to
calculate NDVI (Normalized Difference Vegetation Index) [14], which is a structural vegetation index
utilized for the production of vigor maps according to the methodology extensively described by
Matese et al. [8]. A set of statistical tools were applied to analyze the images with different purposes:
Basic statistics and histograms for native resolution images were performed using Matlab software.
Aircraft and UAV images have been resampled to match satellite resolution of 5 m by means of a
block-averaging function.
Quadrant decomposition was applied, allowing the decomposition of an image in sub-bocks based on
their internal homogeneity: the bigger the sub-blocks, the higher the inherent homogeneity of the image.
This method, similarly to a spectrum for a signal, enables the identification of the information that is
contained within spatial scales.
The heterogeneity of a map cannot be simply described in terms of descriptive but should account for
the spatial structure, or patchiness, of the described variable statistics [15]. CV (coefficient of variation)
is a measure of relative variance and was calculated as the ratio of standard deviation to mean value
expressed as percentage. For each field, NDVI values were also used to compute geo-statistical
information, such as the variogram and the trend (gradient), that were used to assess the within field
variability. Native resolution image data were processed using Matlab code in order to calculate the
variograms for both the vineyards. Experimental (Semi-) Variogram function [16] calculates the
experimental variogram and Variogramfit function [17] performs a least squares fit of various theoretical
variograms to an experimental, isotropic variogram. Nugget (N) is the height of the jump of the
variogram at the discontinuity at the origin, Sill (S) represent the limit of the variogram tending to infinity
lag distances and Range (R) is the distance in which the difference of the variogram from the sill becomes
negligible. The variogram was computed using the maximum distance, dmax = 125 m. Trend was
calculated using gradient (F) Matlab function, where F is the image matrix and returns the x and y
components of the two-dimensional numerical gradient.
The degree of similarity between images can be described by means of similarity indexes able to
capture the degree of correlation between spatial structures. In particular two similarity indices, Lee and
Pearson, were applied in this work using the map comparison statistic software developed by the
Research Institute for Knowledge Systems [18]. Lee’s index [19] offers an approach to calculate
bivariate spatial association reconciliating Pearson’s r statistic as a spatial measure of bivariate
association and Moran’s I [20] as a univariate measure of spatial association. Basically, the correlation
found between the mean fields is corrected for the degree to which X and Y are spatially autocorrelated.
Lee index measures the extent to which both map 1 and map 2 are spatially autocorrelated and their
neighborhood mean fields are correlated as well. The Pearson correlation (R) was calculated on the basis
of a cell-by cell evaluation.
Remote Sens. 2015, 7 2977
2.4. Cost Analysis
The direct comparison of specific costs for the three platform is nonetheless not feasible due to the
different aggregation of their cost estimates [21], as the satellite images are a commercial product where
all the operational and development costs are included in the price per image figure and the
aircraft campaign utilizes an external vector for aircraft missions. In this study we chose a top-down
approach to account for all the expenses associated to data acquisition and processing, grouped into three
broad categories:
Acquisition costs (C) cover all the expenses to get the raw images. For the satellite this is the
purchase price of the commercial image, for the aircraft it includes the cost of the flying vector
and the expenses for the deployment of the sensor platform and payload, while for the UAV it
includes also all the costs for organizing and conducting the acquisition campaign.
Georeferencing and orthorectification (P1) includes the man-hour costs to obtain a georeferred and
orthorectified image. The price for a single man-hour was considered at 50 Euros.
Image processing (P2) covers all the correction and elaboration, priced in man-hour, needed to
get the final results. The process is similar for each platform, the only difference is the different
resolution (and hence computing time) of the three starting images, and the fact that satellite and
usually aircraft images do not require soil filtering as their resolution do not permit distinguishing
between vines and inter-row. Also for that phase 50 Euros per single man-hour was considered.
3. Results and Discussion
3.1. Histograms and Basic Statistics
The basic statistics performed on the images at native resolution highlighted differences between the
three platforms in their range of values. For both V1 and V2 the histogram of UAV values is broader
and shows NDVI values between 0.2 and 0.9; values range for aircraft images is between 0.3 and 0.7
while satellite images show a narrower NDVI interval between 0.5 and 0.65 (Figure 1a,b).
This different behavior between the three platforms is also confirmed by the descriptive
statistics (Table 2), where UAV images show a higher standard deviation compared to aircraft and
satellite images.
Table 2. Basic statistics at native resolution.
Platform N. Values Average Standard Deviation Skewness CV (%) Trend (NDVI)
V1–UAV 4,956,789 0.589 0.08 −0.38 14.61 0.00005
V1–AIRCRAFT 103,199 0.601 0.06 −0.77 9.98 0.00054
V1–SATELLITE 1012 0.624 0.02 −0.28 3.68 0.0014
V2–UAV 8,233,791 0.536 0.09 −0.11 17.16 0.00037
V2–AIRCRAFT 96,588 0.477 0.07 −0.77 15.93 0.00094
V2-SATELLITE 959 0.567 0.03 0.09 5.29 0.0044
Within each platform, the two vineyards did not show substantial differences in the range of values
with V1 showing a higher average NDVI and a lower standard deviation, compared to V2 that also has
a more Gaussian distribution (Figure 1).
Remote Sens. 2015, 7 2978
CV was greater in V2 than V1 and more variation going from low (satellite) to high resolution (UAV)
was detected. Trend was greater in V2 than V1 showing a regular horizontal drift.
(a)
(b)
Figure 1. (a) V1 NDVI histogram as percentage of total values; and (b) V2 NDVI histogram.
The larger range of values (Figure 1) and the higher coefficient of variability (CV) detected by the
UAV platform is explained by its higher resolution, which in a highly heterogeneous crop, such as
vineyards, enables the identification of the alternation of canopies (higher NDVI values) and
inter-rows (lower NDVI values related to grass cover or bare soil). On the contrary, the typical vineyard
discontinuity was not detected by the satellite resolution that averages canopy and inter-row reflectance
values, therefore providing a narrower distribution. Aircraft data fall in between UAV and satellite in
terms of NDVI histograms (Figure 1) and variability (Table 1), confirming that spatial resolution is the
key parameter controlling the amount of spatial information that is effectively sampled by each instrument.
3.2. Coarse Resolution Inter-Comparison
The comparison of images from the three platforms after the re-scaling at the satellite resolution
(5 m) highlights similar behaviors of vegetation patterns and their spatial structure showing also some
differences among platforms (Figure 2a,b).
Remote Sens. 2015, 7 2979
From a visual inspection of the V1 images, a low vigor zone can be observed in the UAV image in
the central zone of V1 that is progressively less pronounced in the aircraft and then in satellite image
(Figure 2a). Similarly, vineyard 2 shows a gradient in the West-East direction that is more evident in
satellite images while it is smoother in the UAV images and the aircraft pattern shows a more abrupt
distribution in two distinct macro zones along the same direction (Figure 2b).
(a)
(b)
Figure 2. (a) V1 images rescaled at 5-m resolution for the three platforms; and (b) V2 images
rescaled at 5-m resolution for the three platforms.
The similarity analysis enabled a further insight of these analogies and discrepancies: higher
correlations were observed between UAV and aircraft images for both vineyards (R = 0.635 for V1
and R = 0.881 for V2) (Table 3), while the correlation between UAV and satellite was high for V2
(R = 0.779) and low for V1 (R = 0.286). Similarly, aircraft vs. satellite correlation was high in V2 (0.78)
and low in V1 (0.42). Lee’s similarity index was high for V2 between all platforms and only between
Remote Sens. 2015, 7 2980
UAV and aircraft for V1. Overall, all the cross-platform combinations of Pearson correlation and Lee
index had consistently larger values in V2 than V1, suggesting that the presence of gradients and spatial
patterns, like those observed in V2 (Figure 2), tends to increase spatial correlation parameters. On the
other hand, more homogeneous patterns like in V1 are intrinsically less correlated. In terms of statistical
comparison, the lower resolutions seems more informative in presence of vineyards showing variability
according to a spatial gradient (trend), with respect to more homogeneous vineyards or vineyards
showing a more irregular vegetation distribution (lower trend).
Table 3. Similarity analysis indices.
V1
V2
Pearson (R) Lee (L) Pearson (R) Lee (L)
SATELLITE vs. UAV 0.286 0.246 0.799 0.701
SATELLITE vs. AIRCRAFT 0.426 0.346 0.776 0.689
AIRCRAFT vs. UAV 0.635 0.547 0.881 0.747
3.3. Image Decomposition
The analysis of image structure performed on aircraft images with the quadrant decomposition
method highlights a substantial difference in the distribution of dimensional classes between the two
vineyards. V2 showed a higher degree of fragmentation, in particular in the east part, while V1 is
represented by larger blocks and has therefore more homogeneous zones (Figure 3a,b).
Figure 3. (a) V1 quadrant decomposition results and (b) V2 quadrant decomposition results.
The distribution of dimensional classes differs between the two vineyards and is consistent between
the two platforms (UAV and aircraft).
The decomposition of V1 showed a higher presence of larger classes, especially in the aircraft image
decomposition, while V2 was more heterogeneous and well represented by smaller classes. The smaller
Remote Sens. 2015, 7 2981
classes represented are those of the native resolution of each acquisition platform (0.05 and 0.5 m for
UAV and Aircraft, respectively).
The analysis of the curves in Figure 4 enabled the quantification of the level of information that is
not resolved moving from UAV and Aircraft to satellite resolution. Considering that satellite resolution
is 5 m (represented by the dashed line in Figure 5), this spatial decomposition highlights how a relatively
large fraction of information is represented by classes actually smaller than the satellite resolution. But
it is worth noting that such fraction is higher in V2 with respect to V1, this means that satellite resolution
is able to provide an appropriate representation of vineyards in those cases where the spatial structure of
vegetation is more homogeneous.
Figure 4. (a) Distribution of dimensional classes for Aircraft resolution and (b) distribution
of dimensional classes for UAV resolution.
3.4. Variogram Analysis
The trend of variogram (Figure 5) variations with the spatial resolution describes the effect of spatial
heterogeneity, providing an assessment of NDVI spatial structures within the image domain [22]. In
general, all the variograms computed on all the images reached a sill well before dmax (maximum
distance). The sill is an indicator of the spatial variability of the data.
All experimental variograms computed on the images are linear at the origin, without any nugget
effect, then increase promptly and reach almost the whole image variance at a very short range in V1 for
UAV and AIRCRAFT (R
aircraft
= 3.88 m; R
uav
= 0.29 m) with respect to V2 (R
aircraft
= 10.34 m;
R
uav
= 4.72 m), confirming that V1 is poorly structured with respect to V2 (Figure 5a–c). The degree of
image spatial variability was attributable to the sill and for all the platforms it was higher in V2 than in
V1 and was higher for UAV than aircraft and satellite.
Remote Sens. 2015, 7 2982
(a)
(b)
(c)
Figure 5. NDVI variograms of three platforms at native resolutions. The lines represent the
fitted variograms models. The parameter of the variogram model are reported in the legend:
C = sill, N = nugget, and R = range. (a) Satellite, (b) Aircraft, and (c) UAV
The analysis of images fragmentation, performed with variograms and quadrant decomposition,
provided a further insight highlighting the importance of the typology of vegetation structure fragmentation:
Remote Sens. 2015, 7 2983
V1 was characterized by larger clusters of vegetation while V2 is characterized by higher heterogeneity
and smaller vegetation sub-classes. With this approach, V1 distribution of pixels clustering, results are
consistent between UAV and aircraft acquisition, confirming a distribution toward larger classes. For
this vineyard (V1), all remote sensing platforms appear to be highly informative independently from
their native resolution; in fact most of the variability falls in the bigger classes (>2 m), thus closer to the
satellite native resolution. On the contrary, V2 vineyard was characterized by a higher variability and
the classes of information that cluster together are placed at substantially smaller resolution and, in this
case, the satellite platform succeeds only in part to represent the variability of the vineyard that is for
over 50% described by classes smaller than the satellite resolution.
The images comparative analysis acquired by the three different platforms involved the use of
different methodologies in order to understand if when scaled to the spatial resolution of the satellite,
these were comparable. The results show that the loss of information at lower resolution (satellite)
relative to the aircraft and UAV are variable between vineyards, and cannot be simply described by
pixel-based statistical indexes, such as Pearson correlation. In fact, V2 showed either a higher
fragmentation into small spatial scales that are not resolved by the satellite platform (Figure 4b and 5b),
and higher values of Pearson correlation and Lee indexes at the 5 m satellite resolution, with respect V1
(Table 1). These results indicate that high Pearson correlation and Lee values at coarse resolution do not
provide any insight of the actual amount of spatial variability that is contained in smaller classes, which
can only be quantified with structural and variogram analysis.
Other authors, as reported by Garrigues et al. [22], have studied many methods to quantify spatial
heterogeneity from empirical (i.e., local variance), probabilistic (i.e., Variogram and fractal) to
mathematical (i.e., Fourier or wavelet transform, but mainly applied to low-resolution images and
comparing different satellite sensors). D’Oleire-Oltmanns et al. [23] compared UAV and satellite images
for soil erosion assessment, proving an identification of gullies on different scales. Hall et al. [24]
presented a review of remote sensing platforms, satellite and aircraft, demonstrating the high potential
application of such technologies in precision viticulture.
3.5. Inter-Row Separation
Remote sensing representation of vineyards presents specific peculiarities, because of the alternation
of vertical vine canopies with a horizontal surface that can be bare soil or covered by grass. This
characteristic implies that the remotely sensed images contain information other than the vine canopy,
i.e., the inter-row soil and the shading produced by canopies. In this sense, while satellite resolution
necessarily implicates the averaging of row and inter-row information, smaller resolution of the same
order of magnitude of canopy projection, enables performing a filtering of the image with the purpose
of excluding the information coming from the inter-row. The possibility of removing the spectral
response of the inter-row is of particular interest in the case of grass-covered inter-rows that can result
in a biased representation of vine canopy status. This procedure can be easily performed starting from
UAV images by mean of establishing a region of interest (ROI) in the center of each vine row with a
canopy buffer width [25]. A comparison of images from the three resolutions with an inter-row filtering
applied to UAV images is provided in Figure 6.
Remote Sens. 2015, 7 2984
Figure 6. (a) Vineyard portion of Satellite image; (b) Vineyard portion of Aircraft image;
(c) Vineyard portion of UAV image; and (d) Vineyard portion of UAV image with
inter-row filtering.
3.6. Acquisition and Processing Cost Analysis
The three different platform analyzed in this study provide data products that illustrate the capability
of remote sensing technologies to monitor and map vineyards with different levels of accuracy,
emphasizing the impact of spatial resolution on vineyard variability assessment and analysis.
The cost analysis was applied at two different spatial scales related to our study: 5 ha, which was the
area actually mapped by UAV in this study, and 50 ha, which was the area actually mapped by the
aircraft survey. Table 4 summarizes the operational costs associated with this case study. The number of
images required to map a certain area (N) scales exponentially from satellite to aircraft and UAV, also
resulting an increasing cost for the image processing chain (P1 and P2). The acquisition cost was fixed
for satellite at both scales, since the same image can cover both areas, while it increases from
5 to 50 ha by a factor of 1.36 and 2.66 for aircraft and UAV, respectively.
The cost/benefit analysis was calculated on the basis of a service that is offered by a third party, thus
not including the investment cost of purchasing an aerial platform or a UAV, instrumentation,
maintenance, etc. Overall, on small fields (5 ha) the use of UAV appears to be the most cost effective
solution due to the low cost for the data acquisition (Figure 7). On the contrary, when the plots reach a
larger dimension (50 ha analyzed here), the UAV solution appears to be the least economic. The satellite
solution does not imply any significant difference, while the aircraft solution is placed in the middle,
showing only a marginal additional cost in the data acquisition related to the slightly higher flight time,
and some higher image processing costs, since the typical aircraft acquired image was in the order of
8–10 ha, requiring the processing of multiple images in the 50 ha case, and only one image in the 5 ha
Remote Sens. 2015, 7 2985
case. The break-even point that can be derived from Figure 7, i.e., the point at which two or more lines
intercept, is placed slightly above 5 ha for all three platforms, meaning that at such scale size, the three
technologies have approximately the same acquisition and processing cost.
Table 4. Category costs (Euro) for satellite, aircraft and UAV mapping. N is the number of
images that compose the mosaick, C the acquisition costs, P1 the georeferencing and
orthorectifing costs and P2 the image processing costs.
5 ha
50 ha
N C P1 P2 N C P1 P2
Satellite
1 2500 50 100 1 2500 50 100
Aircraft 1 2200 100 150 10 3000 500 300
UAV
100 1500 500 200 1000 4000 1000 300
Figure 7. Plot of category costs (Euro) for satellite, aircraft and UAV platform, considering
a 5 ha and 50 ha mapping area.
The presented techniques are promising tools for farmers to monitor their crops, but each of them, if
analyzed individually, can often be incomplete. In fact, if on one hand the applications of UAV and
aircraft may be optimal for a fine characterization of the fields in terms of resolution and to identify the
intra-vineyard variability, on the other hand satellite remote sensing is capable of mapping field
variability with a higher temporal continuity that is consistent across seasons and multiple years,
allowing monitoring of different vegetation stages during the growing season and to derive an historic
analysis on past seasons.
However, a parameter that can better target farmers to choose one or the other platform, or towards a
multiplatform approach, is represented by the real structure of the vineyard that can be assessed only by
UAV or aircraft application. It is hard, with satellite-only images, to assess factors, such as the actual
degree of heterogeneity of the field, the status of the inter-row area, and therefore to assess the
uncertainties associated with the satellite representation, lacking a fine resolution truth. In the presence
of pronounced intra-vineyard variability associated with a lack of well-defined structure and gradients,
Remote Sens. 2015, 7 2986
drones and aircraft can provide valuable information to tailor the use of pesticides, herbicides, fertilizer
and other applications based on how much is needed at a specific point in a field, saving the grower
money from unnecessarily overusing resources, while at the same time reducing the amount of runoff
that could flow into nearby rivers and streams.
Table 5 is an attempt to integrate all factors considered in this study, summarizing the strengths and
weaknesses of the three platforms used as experienced in the actual acquisition campaigns. The mission
attributes deal with the planning and execution of the surveys, the ability to reach the site (Range), to
deal with weather condition and scheduled practices of the farm (cloud cover and flexibility), the need
of multiple flights to obtain the whole scene (Endurance), and the overall reliability of the platform
installment. With respect to aircraft and satellite, UAV can operate closer to the target with more
flexibility on scheduling, and its acquisition are non-dependent on cloud cover conditions, but has a
much shorter range and endurance and an overall lower reliability, being still in the prototyping phase.
Satellite images on the contrary cover much larger areas, but are subject to fixed scheduling and strongly
depend on cloud cover. The aircraft platform sits in between these two with more flexibility than satellite
and better endurance than UAV.
Table 5. Comparative platform characteristics for different remote sensing platforms.
(++ optimal, + good, o average, - poor).
UAV Aircraft Satellite
Mission
Range - + ++
Flexibility ++ + -
Endurance - ++ ++
Cloud cover dependency ++ + -
Reliability o + ++
Processing
Payload o + ++
Resolution ++ + o
Precision ++ + o
Mosaicking and geocoding effort - o ++
Processing time o + +
The image processing attributes deal with the computational chain deployed from the raw images to
the final products. It includes the precision and resolution attainable on the maps and the effort and
computing time to mosaic, orthorectify and produce the outputs. The strengths of UAV acquisition are
of course in the higher resolution and precision, but at the cost of a greater effort for mosaicking and
geocoding. Given the low number of images in the aircraft survey, an almost automatic processing
code was implemented, reducing time and costs of elaboration. Satellite images on the contrary
require no mosaicking and geocoding, at the price of a much lower resolution. The lower payload of the
UAV platform while requiring dedicated and miniaturized sensors, was not a limiting factor in our
acquisition campaigns
3.7. Operational Discussions
In the context of PV, vegetation mapping serve as a base to perform variable rate applications (VRAs).
In this sense, from an operational perspective, the different platforms might be used with different
Remote Sens. 2015, 7 2987
purposes and provide input for different variable rate applications. One of the most interesting VRA
applications is the Variable Rate Spraying that consist in dosing the quantity of pesticides in function of
the canopy volume, a technique that proved to enable an overall saving of up to 58% of application
volume [26] with consequent reduction of pollution and of operation costs. In this specific case,
high-resolution images are likely to represent the optimal solution for an efficient dosing of treatment,
while low-resolution satellite images risk underestimating the canopy volume and therefore drive the
application of insufficient treatment coverage.
On the contrary, in the case of other VRAs, such as selective harvesting, a technique that enables the
machine selection and harvesting of grapes of different qualitative classes and with different product
destinations [27], the use of low-resolution images would provide a sufficiently accurate representation
of grape quality macro-classes.
A further aspect to be considered is that the low spatial resolution cannot account for inter-row
management practice and, as a final result, it outputs an averaged spectral reflectance of the canopy and
inter-row, independently from the agronomic management adopted.
4. Conclusions
The understanding of the intra-vineyard variability is a keystone to implement effective PV practices,
especially in Mediterranean environment where the land-use patterns are highly fragmented and
vineyards present high heterogeneity because of soil, morphology and microclimate variability. Our
study, based on the comparison of different remote sensing platforms, highlighted that different
resolutions provide similar results in the case of vineyards characterized by pronounced vegetation
gradients and large vegetation clusters. On the contrary, in vineyards characterized by small vegetation
gradients and high vegetation patchiness, low resolution images fail in representing intra-vineyard
variability and its patterns. Furthermore, considering the peculiarity of vineyards crop structure, our
work points out the impossibility of distinguishing canopy and inter-rows in the case of low-resolution
images, something that limits the applicability of this platforms in the case of variable rate spraying.
The cost analysis shows that, beyond technical aspects, an economic break-even between UAV and
the other platforms exists between 5 and 50 ha of area coverage, and also that aircraft remote sensing
remains competitive with satellite above such threshold.
Acknowledgments
This work was supported by a grant from the Italian Ministry of Economy and Finance for the project
“Innovazione e Sviluppo del Mezzogiorno-Conoscenze Integrate per Sostenibilità ed Innovazione del
Made in Italy Agroalimentare-Legge n. 191/2009” and by the Italian MIUR (Progetto Premiale AQUA
to CNR).
We gratefully acknowledge BlackBridge (www.BlackBridge.com) for RapidEye satellite image and
Iptsat srl, the Italian local official distributor of RapidEye (www.iptsat.com) for data analysis support.
The authors are grateful to Netherlands Environmental Assessment Agency (MNP) as the owner and
RIKS BV as the developer of the MCK software.
Remote Sens. 2015, 7 2988
Author Contributions
Alessandro Matese conceived and designed the experiments, supported the statistical analysis, and
wrote the manuscript. Piero Toscano performed the statistical data analysis and wrote the manuscript.
Salvatore Filippo Di Gennaro proposed the field trial’s design and supported the statistical analysis.
Lorenzo Genesio analyzed the data and wrote the manuscript. Francesco Primo Vaccari supported the
field experiment and helped with editorial contributions. Jacopo Primicerio performed the UAV flights
and worked on the costs and comparisons between the platforms. Claudio Belli performed the aircraft
flights. Alessandro Zaldei supported in development and measurements. Roberto Bianconi analyzed
aircraft and satellite images. Beniamino Gioli processed the imagery and helped with editorial contributions.
Conflicts of Interest
The authors declare no conflict of interest.
References
1. Whelan, B.M.; McBratney, A.B. The “null hypothesis” of precision agriculture management.
Precis. Agric. 2000, 2, 265–279.
2. Bramley, R.G.V.; Hamilton, R.P. Understanding variability in winegrape production systems. 1.
Within vineyard variation in yield over several vintages. Aust. J. Grape Wine Res. 2004, 10, 32–45.
3. Salamí, E.; Barrado, C.; Pastor, E. UAV flight experiments applied to the remote sensing of
vegetated areas. Remote Sens. 2014, 6, 11051–11081.
4. Kelcey, J.; Lucieer, A. Sensor correction of a 6-band multispectral imaging sensor for UAV. Remote
Sens. 2012, 4, 1462–1493.
5. Zarco-Tejada, P.J.; González-Dugo, V.; Berni, J.A.J. Fluorescence, temperature and narrow-band
indices acquired from a UAV platform for water stress detection using a micro-hyperspectral imager
and a thermal camera. Remote Sens. Environ. 2012, 117, 322–337.
6. Primicerio, J.; Di Gennaro, S.F.; Fiorillo, E.; Genesio, L.; Lugato, E.; Matese, A.; Vaccari, F.P.
A flexible unmanned aerial vehicle for precision agriculture. Precis. Agric. 2012, 13, 517–523.
7. Zhang, C.; Kovacs, J.M. The application of small unmanned aerial systems for precision agriculture:
A review. Precis. Agric. 2012, 13, 693–712.
8. Matese, A.; Capraro, F.; Primicerio, J.; Gualato, G.; Di Gennaro S.F.; Agati, G. Mapping of vine
vigor by UAV and anthocyanin content by a non-destructive fluorescence technique. In Proceedings
of the 9th European Conference on Precision Agriculture (ECPA), Lleida, Spain, 7–11 July 2013.
9. Hunt, E.R., Jr.; Hively, W.D.; Fujikawa, S.; Linden, D.; Daughtry, C.S.; McCarty, G. Acquisition
of NIR-Green-Blue digital photographs from unmanned aircraft for crop monitoring. Remote Sens.
2010, 2, 290–305.
10. Berni, J.; Zarco-Tejada, P.; Suarez, L.; Fereres, E. Thermal and narrowband multispectral remote
sensing for vegetation monitoring from an unmanned aerial vehicle. IEEE Trans. Geosci. Remote
Sens. 2009, 47, 722–738.
Remote Sens. 2015, 7 2989
11. Goward, S.N.; Markham, B.; Dye, D.G.; Dulaney, W.; Yang, J. Normalized difference vegetation
index measurements from the advanced very high resolution radiometer. Remote Sens. Environ.
1991, 35, 257–277.
12. Gioli, B.; Miglietta, F.; Vaccari, F.P.; Zaldei, A.; de Martino, B. The sky arrow ERA, an innovative
airborne platform to monitor mass, momentum and energy exchange of ecosystems. Ann. Geophys.
2006, 49, 109–116.
13. Papale, D.; Belli, C.; Gioli, B.; Miglietta, F.; Ronchi, C.; Vaccari, F.P.; Valentini, R. ASPIS, A
flexible multispectral system for airborne remote sensing environmental applications. Sensors 2008,
8, 3240–3256.
14. Rouse, J.W., Jr.; Haas, R.H.; Schell, J.A.; Deering, D.W. Monitoring vegetation systems in the
Great Plains with ERTS. In Proceedings of the Third ERTS Symposium, NASA SP-351,
Washington, DC, USA, 10–14 December 1973.
15. Pringle, M.J.; McBratney, A.B.; Whelan, B.M.; Taylor, J.A. A preliminary approach to assessing
the opportunity for site-specific crop management in a field, using yield monitor data. Agric. Syst.
2003, 76, 273–292.
16. Matlab Central (Experimental (Semi-) Variogram function by Wolfgang Schwanghart).
Available online: http://www.mathworks.com/matlabcentral/fileexchange/20355-experimental-
semi-variogram (accessed on 1 September 2014).
17. Matlab Central (variogramfit function by Wolfgang Schwanghart). Available online:
Http://www.mathworks.com/matlabcentral/fileexchange/25948-variogramfit (accessed on 1
September 2014).
18. Visser, H.; de Nijs, T. The map comparison kit. Environ. Modell. Softw. 2006, 21, 346–358.
19. Lee, S.I. Developing a bivariate spatial association measure: An integration of Pearson’s r and
Moran’s I. J. Geogr. Syst. 2001, 3, 369–385.
20. Moran, P.A.P. The interpretation of statistical maps. J. R. Statist. Soc. 1948, 37, 243–251.
21. Valerdi, R.; Merrill, J.; Maloney, P. Cost metrics for unmanned aerial vehicles. In Proceedings of
the AIAA 16th Lighter-Than-Air Systems Technology Conference and Balloon Systems Conference,
Arlington, VA, USA, 26–28 September 2005.
22. Garrigues, S.; Allard, D.; Baret, F.; Weiss, M. Quantifying spatial heterogeneity at the landscape
scale using variogram models. Remote Sens. Environ. 2006, 103, 81−96.
23. D’Oleire-Oltmanns, S.; Marzolff, I.; Peter, K.D.; Ries, J.B.; Aït Hssaïne, A. Monitoring soil erosion
in the Souss Basin, Morocco, with a multiscale object-based remote sensing approach using UAV
and satellite data. In Proceedings of the 1st World Sustainability Forum, Sciforum Electronic
Conference Series, 1–30 November 2011. Available online: http://www.sciforum.net/conference/
wsf/paper/562 (accessed on 9 March 2015).
24. Hall, A.; Lamb, D.W.; Holzapfel, B.; Louis, J. Optical remote sensing applications in viticulture–A
review. Aust. J. Grape Wine Res. 2002, 8, 36–47.
25. Zarco-Tejada, P.J.; González-Dugo, V.; Williams, L.E.; Suárez, L.; Berni, J.A.J.; Goldhamer, D.;
Fereres, E. A PRI-based water stress index combining structural and chlorophyll effects:
Assessment using diurnal narrow-band airborne imagery and the CWSI thermal index. Remote Sens.
Environ. 2013, 138, 38–50.
Remote Sens. 2015, 7 2990
26. Llorens, J.; Gil, E.; Llop, J.; Escolà, A. Variable rate dosing in precision viticulture: Use of
electronic devices to improve application efficiency. Crop Prot. 2010, 29, 239–248.
27. Bramley, R.G.V.; le Moigne, M.; Evain, S.; Ouzman, J.; Florin, L.; Fadaili, E.M. On-the-go sensing
of grape berry anthocyanins during commercial harvest: Development and prospects. Aust. J. Grape
Wine Res. 2011, 17, 316–326.
© 2015 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article
distributed under the terms and conditions of the Creative Commons Attribution license
(http://creativecommons.org/licenses/by/4.0/).
