Assessment of sunflower water stress using infrared thermometry and computer vision analysis


 The objectives of the current study were to implement affordable and non-invasive measurements of infrared thermometry, leaf relative water content (RWC), crop water stress index (CWSI), leaf area index (LAI) from computer vision analysis and seed yield of sunflowers. The experiment was designed as split-plot based on randomized complete blocks with three replications. Treatments were four different levels of deficit irrigation as the main plots and three fertilization treatments were applied as sub-plots. Results showed a significant effect (P ≤ 0.01) of water stress and fertilizer on CWSI during different stages of sunflower growth. Changes in fertilizer amount and type resulted in a change in lower (dTLL) and upper (dTUL) limits of canopy-air temperature difference. A combination of chemical fertilizer with biofertilizer could help to decrease CWSI. From computer vision analysis, the normalized difference red blue index (NDRBI) had a strong linear relationship with RWC and CWSI for sunflowers (R2 of 0.87 and 0.93, respectively) and the normalized difference green blue index (NDGBI) had a linear relationship with seed yield (R2 = 0.79). Therefore, analysis of digital RGB images and CWSI were efficient, non-destructive and low-cost methods to assess crop water status for sunflowers under different irrigation and fertilizer treatments.


INTRODUCTION
Water resources are under pressure due to the rapid increase in world population, natural resources pollution and climate change. As water scarcity increases and more agricultural water is diverted to other sectors, advanced technology, methods and equipment must be used for irrigation systems and scheduling so that the limited water resources would meet the needs of farmland irrigation (Wenting et al. ). These methods could be classified as destructive (e.g., stem water potential using a pressure bomb, non-destructive (e.g., infrared thermography (Cohen et al. )/ thermometry (Jackson et al. 1982)), expensive requiring high technology and know-how (e.g., using a large number of sensors for precision irrigation or cameras for thermal infrared/multispectral/hyperspectral as payload of unmanned aerial/unmanned aerial and terrestrial vehicles [UAV, UTV respectively] (Gago et al. ), and low-cost (e.g., digital camera (Zakaluk & Sri Ranjan )). Among these methods, infrared thermometry techniques measure canopy temperature non-invasively by measuring reflected long wave radiation from plant canopy surfaces and has high temporal resolution to monitor plant water stress (Cohen et al. ). To transform canopy temperature data into information on plant water status, many indices have been proposed. One of the most common and popular indices is CWSI (Cohen et (Carroll et al. ). It should be emphasized that these measurements (canopy measurements (T c ) and leaf RWC) were time-consuming and required a large number of manual sampling to describe a whole field ( Jackson ). Therefore, these techniques are considered less effective, frequently challenging, and time and resource consuming (Lee & Lee ).
Color digital image processing within the visible (400-700 nm) region through computer vision algorithms has been used with some success in the diagnosis of crop growth and water status as a low-cost and efficient tool (Lee & Lee ). Many studies have used spectral veg- Overall, farmers can monitor crop growth parameters (such as LAI and yield), crop water status (such as leaf RWC and CWSI) using inexpensive equipment (digital camera and infrared thermometer) and computer vision techniques at the canopy level for a sunflower crop under different management. However, more information is required related to the application of the mentioned approach to detect crop growth and water status of sunflower at the canopy level under different management practices and specifically the effect of type and amount of fertilizer on CWSI. The main objectives of this study were: (1) determine upper and lower baselines for calculating CWSI under fertilizer treatments in the semi-arid conditions; (2) evaluate the ability of reflectance-based indices for detecting water stress in comparison to plant measurements; (3) determine the relations between image feature parameters and LAI and seed yield; (4) determine the effect of water stress during the growing season on RWC, LAI and sunflower yield.

Study area and experimental treatments
The field experiment was carried out during the summer

Soil water content (SWC) measurements
Before each irrigation event, soil samples at two depths 0-60 cm (0-30; 30-60 cm) were obtained from the centre of the plots to measure the SWC by the gravimetric method (Black ), and the amount of irrigation water was calculated based on the soil moisture deficit (to fill the FC). Also, the irrigation interval was estimated based on the cumulative evaporation from the pan. Equation (1) was applied to where I is the amount of irrigation water (m 3 ), W FC is soil water content at field capacity (%), W i is pre-irrigation soil moisture content (%), γ is soil bulk density (g cm À3 ), D is soil depth (0-0.6 m), and A is surface area of the plot (m 2 ).  (2):

Sample collection and plant measurements
where dT m , dT LL and dT UL are measured, lower limit and upper limit of dT, respectively. LAI, RWC and seed yield were measured under deficit irrigation and fertilization treatments.

Image acquisition
The imaging procedure was carried out using a 16-Megapixel resolution digital camera with 140-280 mm focal length. All the images were taken weekly at Nadir at a fixed distance of 1 m over the sunflower plants. To minimise the variation and shadows, all images were taken at the same time of the day under a similar light condition.

Image analysis
After image acquisition, the pre-processing steps included size reduction, background removal, normalization, noise removal and image binarization to improve image quality and remove unnecessary objects from images.
The first stage was dividing the main image into its three components (Red, Green and Blue) then Equation (3)  was normalized from 0 to 1.
The filtering of sunflower canopy and its background was carried out in the second step. In the visible range of the light spectrum, green vegetation has a reflectance peak in the green wavelength (495-570 nm), whereas there is not any apparent change in soil albedo. Therefore, in this research, the difference between sunflower canopy and the non-canopy area can be enhanced by the greenness index (Liu & Pattey ), as defined in Equation (4).
where r, g, and b represent the intensity of levels by normalized red, green and blue components in a digital camera.
Once a threshold was set, a binary mask image was created.
The pixel value ¼ 1 specifies a pixel with a greenness value higher than the threshold indicates a sunflower canopy. A pixel value ¼ 0 represents the background and includes soil and plant residues. Then all the individual masks were applied to the relevant image, so the background was removed from all images ( Figure 3). Also, the color of the sunflower seed head section was sharply different from that of the canopy section and it made noise, so to separate sunflower seed head the by threshold method was used.
The digital images were further processed to calculate plant stress indices which are represented in Table 2. They have been used to estimate the RWC, CWSI, LAI and sunflower yield. Figure 4 shows the basic procedure of the proposed method and image analysis algorithm.
Analysis of variance (ANOVA) was conducted on all observed crop and reflectance data by using SAS 9.2 software (SAS Institute, Cary, NC, USA). In all cases, the coefficient of determination (R 2 ) and root mean square error (RMSE) were used to assess the most appropriate regression equation.

Baseline equations and CWSI
NWSB is shown in Figure 5 for three growth stages, vegetative, reproductive, and maturation stages as well as the whole of sunflower season, under different levels of fertilizer. To calculate NWSB and CWSI, regression curves were fitted for mentioned periods to get a and b coefficients (Table 3). As shown in Table 3, the range of R 2 for the NWSBs at different stages of growth was observed between  NDGBI normalized difference green blue index NDGRI normalized difference green red index GRS green red slope transformation  water conditions. Also, these differences confirm that this equation needs to be validated for region and crop type.
As seen from  (Table 4). As seen from    and sunflower (Gholamhoseini et al. ).
Based on the analysis of variance results, the differences between the irrigation and fertilizer treatments for both seed yield were statistically significant (P< 0.01) (  (Table 4). This is mainly due to the fact that head diameter and seed number per head reduction cause to decline sunflower seed yield under water stress (Gholamhoseini et al. ).
Among the fertilizer treatments, the lowest and highest values of seed yield were reported as 2.26 t ha À1 and 2.79 t ha À1 under F 100 and F 50 treatment, respectively (Table 4).
It seems that an appropriate equilibrium between available soil nitrogen and plant nitrogen requirements causes seed yield to increase in F 50 and F 75 treatments. This is mainly due to the fact that bio fertilizer and chemical fertilizer application improve soil biological activity and nutrient mobilization from different sources (chemical and organic sources) and adjust organic matter decomposition dynamics and the plant nutrient availabilities (Reddy et al. ). Also, the increase of microbial biomass is directly related to soil health and thus enhances the balance of nutrient elements, and nutrient availability in the root zone that promotes growth and ultimately affects a higher yield (Biari et al. ).

Relationships between image feature parameters and sunflower water status, LAI and seed yield
To determine the plant water status, the correlation between the indices extracted from color images (NDGBI, NDRBI, GRS, GBS, RBS, NDGRI, VARI, SAVI Green , CC and GMR) and the indicators of sunflower water stress (RWC and CWSI) were investigated and the best relationship between them was determined (Table 5 and Figure 6(a)-6(d)). Based on the results of this The other advantages of the RWC index would be its relation with the CWSI (Soleymani et al. ), CWSI variations are explained due to water transpiration through the stomata so CWSI prediction could be acceptable based on the blue and red components. In Figure 6(a)-6(d), the relationship between NDRBI index and water stress indices (RWC and CWSI) has been illustrated under water and fertilizer treatments. As seen in Figure 6(c) and 6(d), an increase in the NDRBI index was found as CWSI increases, or in other words, water stress increases.
Therefore, it can be concluded that the reduction of Based on Figure 6(a)-6(d), the strong linear relationship between sunflower moisture indicators (RWC, CWSI) and extracted index from the image, NDRBI, indicates that this index can be used to detect sunflower water stress. The linear relationship is presented as in Equation (5): During the growth period only, the LAI were analyzed.
The results of investigating LAI and image parameters showed that LAI had a significant correlation with CC (canopy cover) (P < 0.01) and with other vegetation indices   (6).
where y is LAI, k is the initial value of the curve function, and d is the curve shaping parameter, and CC is canopy cover.
According to The relationship between the NDGBI index extracted from the images and seed yield are shown in Figure 6(g) and 6(h). As observed in Figure 6(g) and 6(h), the lowest index of NDGBI was observed under non-water stress treatment. By increasing the water stress, the value of this index increases, while grain yield decrease. The combination of chemical and bio fertilizers (F 50 ) showed also a better correlation between the NDGBI index and grain yield.
The NDGBI index has linear relationships with seed yield, so that the linear equation with the highest coefficient of determination (0.79), in Equation (7), was known to be the best regression for determining seed yield.
in which y represents the seed yield and x represents NDGBI. Both parameters a and b were obtained by the least squares method.
In a study by Wang et al. (), the CC index was reported to be the best model for estimating biomass with a coefficient of 90% for rice. Elazab et al. () showed that NDGRI index is the best model for the wheat yield estimation under water stress condition. The close relationship between NDGBI with seed yield is an important tool for evaluating seed yield in sunflower (Figure 6(g) and 6(h)).
Although field studies are the most common method to determine yield, field surveys can be time-consuming, difficult and costly. So, image indicators could present an appropriate estimate of the sunflower seed yield under different management conditions.
Error analysis results can only obtain the overall performance of the indices extracted. Mean absolute error in predicted CWSI, RWC, LAI and seed yield were 0.03, 1.65%, 0.55 m 2 m À2 and 0.21 t ha À1 , respectively. The distribution of percentage error in predicted CWSI, LAI, RWC and seed yield is shown in Figure 7. Reasons for the different errors may be due to differences in input surface reflectances, and the effect of water stress and fertilizer on the stomatal response to visible light.

CONCLUSION
The current study evaluated the ability of infrared thermometer and digital camera (RGB images) to monitor sunflower crop water status (CWSI and RWC) and crop growth indices (LAI and seed yield) in a sunflower field, The strong correlation between these indices and crop growth parameters indicated their potential suitability to develop strategies and make a decision by farmers and managers to track crop growth and water stress using digital cameras and image processing methods as a reliable, fast, less expensive approach. However, future studies should aim at verifying and/or modifying relationships between CWSI and image indices presented in this paper for other crops and under different climatic conditions.

DATA AVAILABILITY STATEMENT
All relevant data are included in the paper or its Supplementary Information.