This study investigates whether a water deficit index (WDI) based on imagery from unmanned aerial vehicles (UAVs) can provide accurate crop water stress maps at different growth stages of barley and in differing weather situations. Data from both the early and late growing season are included to investigate whether the WDI has the unique potential to be applicable both when the land surface is partly composed of bare soil and when crops on the land surface are senescing. The WDI differs from the more commonly applied crop water stress index (CWSI) in that it uses both a spectral vegetation index (VI), to determine the degree of surface greenness, and the composite land surface temperature (LST) (not solely canopy temperature).
Lightweight thermal and RGB (red–green–blue) cameras were mounted on a UAV on three occasions during the growing season 2014, and provided composite LST and color images, respectively. From the LST, maps of surface-air temperature differences were computed. From the color images, the normalized green–red difference index (NGRDI), constituting the indicator of surface greenness, was computed. Advantages of the WDI as an irrigation map, as compared with simpler maps of the surface-air temperature difference, are discussed, and the suitability of the NGRDI is assessed. Final WDI maps had a spatial resolution of 0.25 m.
It was found that the UAV-based WDI is in agreement with measured stress values from an eddy covariance system. Further, the WDI is especially valuable in the late growing season because at this stage the remote sensing data represent crop water availability to a greater extent than they do in the early growing season, and because the WDI accounts for areas of ripe crops that no longer have the same need for irrigation. WDI maps can potentially serve as water stress maps, showing the farmer where irrigation is needed to ensure healthy growing plants, during entire growing season.
Recent developments in unmanned aerial vehicles (UAVs) have extended the
practice of remote sensing in precision agriculture research
(Berni et al., 2009b; Gonzalez-Dugo et al., 2012; Hoffmann et al., 2016; Turner et
al., 2011; Vergara-Díaz et al., 2016; Zarco-Tejada et al., 2009,
2013b). UAV platforms now enable the collection of remotely sensed
temperatures with higher spatial and temporal resolution than those
collected by satellites and manned aircraft. Temperatures from canopies are
closely related to air and soil water content, actual transpiration and crop
water stress (Idso et al., 1986; Jackson et al., 1987). If a crop has insufficient water supply,
stomata will close in order to limit water loss through transpiration. This
leads to energy being stored and thus higher canopy temperatures relative
to those seen in crops with ample water supplies
(Guilioni et al., 2008). Canopy temperature is
therefore positively correlated with crop water stress and negatively
correlated with soil moisture and transpiration. Quantifying crop water
stress allows farmers to assess the need for irrigation.
Idso et al. (1981) and Jackson et al. (1981) developed the well-documented and commonly used crop water stress index (CWSI) (Barbosa
da Silva and Ramana Rao, 2005; Feldhake et al., 1997; Sepaskhah and
Kashefipour, 1994; Xiang and Tian, 2011). They related canopy temperatures
to evaporation and presented the stress index as
In this study, a UAV-based WDI was applied to barley fields three times during the spring and summer of 2014 in order to assess whether it can detect intra-field variations of crop water status at different crop growth stages. Both the early and the late growing seasons were investigated to assess whether the WDI possesses the unique potential to give accurate results both when the surface is partially composed of bare soil and when the crops are senescing. The latter required the WDI to be based on a VI that can determine the greenness of crops, i.e., whether they are ripe or not ripe. We classified yellow barley as ripe or prematurely ripe and green barley as not ripe. Thus, in this study, the VI served two purposes: detection of the canopy cover in the early growing season, as originally intended, and detection of the developmental stage of crops in the late growing season. The normalized green–red difference index (NGRDI) was used as the VI, and its suitability to be incorporated in the WDI will be discussed below. Further, we extended the original WDI with an extra set of thermal observations in order to accommodate any offset between surface and air temperature measurements caused by, for example, atmospheric effects (Anderson et al., 1997). The composite LST and the VIs (on which the WDI was based) were collected in both overcast and moderately sunny weather conditions.
The objective was thus to investigate whether a WDI based on UAV imagery can provide accurate crop water stress maps during different growth stages of barley and in different weather situations.
This study was carried out in two adjacent spring barley fields constituting
a 400
The two barley fields located in western Denmark (orange square:
56.037644
The fields were routinely irrigated according to normal practice on sandy soils. In 2014, irrigation was applied approximately on 23 May, 29 May, 5 June, 15 June and 2 July. The irrigation of both fields takes approximately 2 days, and 25 mm of water was applied on each occasion using a traveling irrigation gun. Barley was sown on 14 March 2014 and harvested on 22 August 2014. Data were collected on three dates (22 April, 18 June and 2 July), and the growth stages of the barley were, according to Tottman (1987), 13, 61 and 69, respectively. On 22 April, the barley had approximately three unfolded leaves and a height of 0.08 m. On 18 June, the barley was in the early stages of anthesis and 0.95 m tall, while on 2 July, the anthesis was complete and the crops were 1.1 m tall.
The WDI is defined in Moran et al. (1994) as
Illustration of a standard shape of the trapezoid used for the VIT
index calculations formed by four hypothetical vertices representing the
extreme water and canopy conditions: (1), (2), (3) and (4). “C” represents
corresponding values of measured
Corresponding
The upper and lower VI values (
Data used to compute
The NGRDI is computed according to Hunt Jr. et al. (2005)
as
Computations of
Maximum NGRDI values of 0.25 were detected on 18 June and minimum values of
RGB and thermal data were collected with a fixed-wing UAV on three occasions
during 2014: 22 April, 18 June and 2 July. The lithium-battery-driven UAV
was in the air for approximately 20 min on each occasion. The thermal camera was
an Optris PI 450 camera that detects infrared energy in the 7.5–13
The thermal camera was triggered by its internal auto-trigger and each image was assigned GPS coordinates, as described in Hoffmann et al. (2016), by comparing the timestamp on images with the timestamp and GPS position extracted from the autopilot log file (created by autopilot software, SkyCircuit Ltd SC2). Thermal images were converted into unsigned 16-bit data to enable stitching of images in Agisoft PhotoScan software (professional edition version 1.0.4).
The RGB data were collected with a digital Panasonic DMC-LX5 Lumix camera in
which exposure times are adjusted automatically with a built-in light
intensity meter. The RGB camera was triggered using predefined settings in
the SkyCircuit software, and georeferencing was conducted directly in the
autopilot log file, as each triggered image was marked next to its
corresponding GPS position. No conversion of RGB images was needed for the
PhotoScan framework. A flying height of 90 m above the ground allowed
resolutions of 0.1 m pixel
Half-hourly averages of air temperature measurements (
NGRDI maps were compared with UAV-RGB orthophotos created in PhotoScan. The
NGRDI map from 18 June was also validated against measurements of leaf area
index (LAI) and against near-field remote sensing measurements of the
normalized difference vegetation index (NDVI). LAI
measurements were obtained on 18 June with a plant canopy analyzer (LAI2000)
at six locations selected according to empirically confirmed different crop
yields from previous years. LAI measurements were repeated four times at
each location and a mean represented the final LAI value for each specific
location. NDVI data were collected on 22 June with a MobilLas mounted on a
tractor with three sensors: a near-infrared laser range finder (AccuRange
4000, Acuity Research Inc., CA, USA), a two-band radiometer (Crop Circle
ACS-210, Holland Scientific Inc., NE, USA) and a global positioning system
(GPS 16, Garmin International Inc., KS, USA). The two-band radiometer was an
active canopy sensor providing its own illumination, and measurements were
made independently of solar radiation. NDVI was calculated as in
Carlson and Ripley (1997) with reflectance measurements made
at near-infrared (780 nm) and red (670 nm) frequencies. Measurements were
made at a 50
The surface-air temperature difference maps were compared with measurements from time domain reflectometry (TDR) probes providing information on volumetric soil water content. Six TDR probes were placed at the same locations as those for the LAI measurements: three in the field north of the road and three south of the road. Measurements were collected on 9 days during the growing season 2014: 27 March, 24 April, 8 May, 13 May, 27 May, 4 June, 13 June, 26 June and 17 July. The TDR100 instrument (Campbell Scientific, Logan, UT, USA) was deployed as a probe system with a central hub connecting the six individual probes. The central hub ensures that measurements can be taken without disturbing the vegetation and soil where probes are installed (Thomsen, 2006). The probe rods measured soil water content representing the first 0.2 and 0.5 m of the soil profile. Three repeat measurements were made at each location, and volumetric soil water contents were calculated using the apparent dielectric constant, as in Topp et al. (1980).
Mean values from WDI maps were validated with measured stress values
computed as
Maps of NGRDI,
Imagery and indices from 22 April:
Imagery and indices from 18 June:
Imagery and indices from 2 July:
Visual comparison of NGRDI maps and corresponding orthophotos (Figs. 4–6, panels a and c) confirms that the NGRDI can be used to quantify greenness from the RGB data. This is observed relative to soil (Fig. 4) and relative to yellowing barley (Figs. 5 and 6). The green stripes along the edges of the fields in the orthophoto and a northeast–southwest-orientated stripe of green vegetation in the field north of the gravel road in Fig. 4c are translated into brighter areas in Fig. 4a (purple lines), indicating more green vegetation and a higher canopy cover. Comparing Fig. 5a with c and Fig. 6a with c shows that the NGRDI is also able to quantify the degree of crop ripeness. Areas with ripe crops attain lower (darker) NGRDI values (areas highlighted with blue). The same NGRDI response is observed in areas with bare soil (Fig. 4a), and therefore the NGRDI responses to ripe crops and bare soil are indeed similar.
The NGRDI map from 18 June was compared with approximately 1300 spatially
distributed measurements of NDVI from 22 June, and a correlation coefficient
(
Correlation coefficients (
High LAI values were measured in areas with high NGRDI values and likewise with low LAI and NGRDI values. A correlation coefficient of 0.85 was achieved between these two parameters. On 18 June, low LAI measurements were located in areas with ripe crops (see Appendix C). Yellow crops on this date were presumably prematurely ripe and therefore smaller. Correlations between the NGRDI map and LAI and NDVI measurements bode well for the quality of collected RGB data and for the NGRDI as a greenness index. A correlation has further been conducted between LAI and NDVI to test whether NGRDI or NDVI best represented the LAI measurements (see Table 2). The correlation coefficients are equally high. According to Knipling (1970), crop reflectivity response in the visible spectrum depends on leaf chlorophyll content and information from the visible bands thus enables a distinction between green and yellow crops. However, vegetation indices based on spectrally narrow bands collected with advanced sensors in the visible and near-infrared spectra are most commonly used when crop conditions are assessed (Baluja et al., 2012; Garcia-Ruiz et al., 2013; Lelong et al., 2008; Primicerio et al., 2012; Sugiura et al., 2005; Zarco-Tejada et al., 2012). Rasmussen et al. (2016) concludes that there is no difference in the ability to detect spectral crop response between consumer-grade RGB cameras (broad bands) and multispectral sensors providing narrow-band information in visible and near-infrared spectra.
There was clear agreement between
Appendix D shows the location of the six TDR probes on the
Correlation coefficients (
The shape of the elongated warmer areas in
Comparison of Figs. 5b, c and 7c reveals that thermal UAV data contain more information than it is possible to see simply by looking at the RGB images, i.e., more than what can be seen with the naked eye. White areas highlighted by red circles (Fig. 5b) indicate higher temperatures and higher soil water deficits that cannot be detected in the orthophoto (Fig. 5c). Crops in the areas highlighted with red circles in the 18 June map become yellow by 2 July (Fig. 6c). The temperature data from 18 June thus predict where crops on 2 July will become prematurely ripe as a consequence of being water stressed and receiving insufficient irrigation.
Figures 4–6 (panels a, b and d) show that variations in both
Mean values and standard deviation (in parentheses) of WDI maps
along with the measured stress index
Mean values and standard deviations of the WDI maps are shown in Table 4,
along with computed CWSI indices and measured stress values
The WDI is highest in April and lowest in June and July. On 22 April, the barley was in an early growth stage, so the root networks in the fields were not yet large nor deep. Further, the low LAI values on this date indicated that small fractions of the field of view were occupied by canopy. The LST collected on 22 April is thus an expression of soil water in the uppermost soil layer, which may dry rapidly. This led to the high WDI of 0.58 in 22 April. WDI results from fields with low LAI and shallow root networks are correct according to measured stress values and thus also according to actual evapotranspiration. However, they only represent soil water availability in the top few centimeters of the soil and are therefore less representative of the deeper soil water content. The relatively low stress values of 0.40 and 0.41 obtained in June and July agree well with the repetitive irrigation of the fields initiated on 23 May. The large variation in the WDI map from 18 June is due to the large amount of available energy detected on this day. Crops with insufficient soil water availability will heat up in accordance with the large amount of available energy. Variations in soil water will result in large variations in crop temperature.
The WDI provides very accurate estimates of crop water status, and
therefore the WDI maps provide precise irrigation maps. WDI maps are most
valuable in the late growing season. At this stage, the remotely sensed data
represent plants' available water more sensitively than they do in the early
growing season, where the majority of the remotely sensed data represent
water availability in the top few centimeters of the soil profile. The WDI
accommodates absolute stress values. However, there is a degree of empirical
assessment associated with the determination of upper and lower VI values in
constructing the trapezoid. The
The added utility of WDI response in late growing seasons is dependent on the color change in ripening crops, and crops with color change that is similar to that in barley (green to yellow) will probably be suitable for beneficial monitoring with WDI applications. We therefore expect that the added utility of the WDI will be valid for many types of cereals. However, further studies will need to be conducted to generalize WDI advantages.
CWSI thresholds for stressed and not stressed crops are species dependent (Feldhake et al., 1997), and so the same will be true for WDI thresholds. WDI thresholds, which determine the amount of irrigation needed, are beyond the scope of this paper; they will need to be analyzed in future studies.
In this study, the UAV-based WDI was applied to barley fields in April, June and July (2014), to investigate whether crop water deficits at subfield scale could be determined at different crop growth stages. Data from both the early and the late growing season were investigated to assess whether the WDI has the unique potential to be applicable both when the land surface is partly composed of bare soil and when crops on the land surface are senescing.
We found that the WDI maps determined accurate absolute water stress values and variations within the barley fields, in agreement with measured stress values from the eddy covariance tower, at different growth stages. This implies that the WDI accounts for areas with ripe and prematurely ripe crops that no longer need large volumes of irrigation. The robustness of the WDI during different growth stages emphasizes its added utility compared with the more commonly used CWSI. Further, the WDI has the potential to become an efficient and powerful irrigation tool in areas where overcast weather is common. We also found that the surface-air temperature difference maps alone can predict where crops will become prematurely ripe with insufficient irrigation. The study also demonstrated that a lightweight UAV system, a consumer-grade camera and an uncalibrated and uncooled thermal camera can be combined to produce accurate maps of crop water stress. In this way, reliance on camera calibration and costly multispectral cameras can be reduced. The WDI response in the late growing season was crop-color dependent, and studies applying the setup we have presented to other crop types are needed in order to confirm the general added utility of the WDI.
The data used in this study are available upon request from the corresponding author (helene.hoffmann@ign.ku.dk).
Computed trapezoids for the 3 data collection days. Red line is
(
NGRDI from 18 June against NDVI from 22 June. NDVI seems to saturate at high values of NGRDI.
Location and size of LAI measurements on NGRDI map from 18 June. Size of green stars indicates the magnitude of LAI, according to the legend. Darker areas on the background map represent low NGRDI values and areas with ripe/yellow crops. Brighter areas represent high NGRDI values with high green canopy cover. A correlation coefficient of 0.86 is obtained between NGRDI and LAI.
The location of the six TDR probes on the
The soil water content measured at the six TDR probes during the growing season. Solid lines represent measurements at 0.2 m depth and semi-dashed lines represent 0.5 m depth. Black dashed lines indicate days of UAV data collection: 22 April, 18 June and 2 July.
TDR probe 1 and 5 (red and orange) are located in areas with higher LST. Figure D2 shows that probes 1 and 5 measure lower soil water contents than the other four probes, especially on 18 June (black dashed lines in Fig. D2 show dates for UAV data collection). Comparison of the seasonal trend in Fig. D2 with temperature maps from 22 April and 2 July (Figs. 4b and 6b) shows agreement between soil water content and temperature maps: the smaller temperature variance in the map from 22 April corresponds well with the small variability in soil water content among TDR probes in the same period. Similarly, the location of warmer areas in the June and July maps corresponds well with the lower soil water content seen at probes 1 and 5 during June and July.
Helene Hoffmann performed most of the data analysis, collected the UAV data and wrote the manuscript; Rasmus Jensen also collected UAV data and contributed to the analyses; Thomas Friborg and Hector Nieto contributed to the design of the study; Anton Thomsen collected validation data; Jesper Rasmussen contributed to the interpretation of data. All authors contributed to the editing of the manuscript.
This work was conducted within the Danish Hydrological Observatory, HOBE, and was funded by the Villum Foundation. Edited by: P. Stoy Reviewed by: two anonymous referees