The Potential of a Precision Agriculture (PA) Practice for In Situ Evaluation of Herbicide Efficacy and Selectivity in Durum Wheat (Triticum durum Desf.)

Abstract

Precision agriculture (PA) practices based on the use of sensors and vegetation indices have great potential for optimizing herbicide use and improving weed management in field crops. The objective of this research was to evaluate the efficacy of commercial herbicide products and their selectivity in durum wheat by measuring the Normalized Difference Vegetation Index (NDVI). Field trials were conducted in Velestino and Kozani, Greece (2020–2021 and 2021–2022) in four site-years with the following treatment list: untreated control (T1), 2,4-D at 300 and 600 g a.e. ha⁻¹ (T2 and T3, respectively), pyroxsulam + florasulam at 18.82 + 3.71 g a.i. ha⁻¹ + cloquintocet-mexyl at 18.82 g a.i. ha⁻¹ (T4), and mesosulfuron-methyl + iodosulfuron-methyl-sodium at 15 + 3 g a.i. ha⁻¹ + mefenpyr-diethyl at 45 g a.i. ha⁻¹ (T5). Site-years and treatments affected weed NDVI, weed biomass, crop NDVI, and grain yield (p ≤ 0.05). At Kozani, weed NDVI was lowest in T4 plots in 2020–2021 (0.31) and 2021–2022 (0.33). Treatments T4 and T5 resulted in lowest weed biomass in 2020-2021 (14–16 g m⁻²) and 2020-2021 (19–22 g m⁻²). At Velestino, T3 reduced weed biomass by 92 and 87% when compared to T5 in 2020–2021 and 2021–2022, respectively. Approximately, 67% and 73% of the variability in weed biomass in 2020–2021 and 2021–2022, respectively, at Kozani could be explained by weed NDVI. These parameters were strongly correlated in Velestino (R² ≥ 90%). Low crop NDVI at Kozani indicated herbicide injury in T3 plots, confirmed by yield losses. During 2020-2021, yield was 30, 38, and 40% higher in T4 plots than in T2, T1, and T3 plots, respectively. At Velestino, yield in T1 plots was 25, 27, 27, and 29% lower than in T2, T4, T5, and T3 plots, respectively, in 2020–2021. Similar results were obtained in 2021–2022. The current study indicates that NDVI can be used as a reliable, non-subjective indicator of herbicide efficacy and selectivity in winter cereals. The methodology used in this work should also be evaluated in other crops and under different soil and climatic conditions.

1. Introduction

Weed management is an essential agronomic practice to achieve high yields in durum wheat (Triticum durum Desf.) [1,2,3,4]. The primary method of controlling harmful weeds affecting crop yield and quality in this crop (and in winter cereals in general) is the application of herbicides. In particular, synthetic auxins (HRAC/WSSA Group 4) such as 2,4-D can be very effective in controlling broadleaf weeds; acetyl co-enzyme A carboxylase (ACCase)-inhibitors provide control of grass weeds, and acetolactate synthase (ALS)-inhibitors (HRAC/WSSA Group 2) belonging to sulfonylureas, triazolopyrimidines, and some other chemical families are commonly used to control both groups of weed species [5]. Optimizing the use and efficacy of herbicides currently available for use in this crop has become a priority. This is due to the lack of new herbicide modes of action, the phase-out of several active ingredients, and the occurrence of several weed biotypes that are developing resistance patterns to the above herbicide groups [6].

Precision agriculture (PA) practices have great potential to optimize herbicide use and contribute to the development of effective and sustainable weed management systems. Remote sensing techniques based on the use of smart sensors and vegetation indices can provide information on herbicide efficacy a short time after treatment [7]. For example, spectral weed indices for Kochia scoparia L., Ambrosia artemisiifolia L., and waterhemp Amaranthus tuberculatus (Moq.) J.D. Sauer have been used by combining a single wavelength and a normalized wavelength (450–920 nm) to evaluate herbicide efficacy and identify potentially herbicide-resistant populations of the above weed species [8]. Another recent example is the study by Xia et al. [9], who acquired multispectral and RGB images using Unmanned Aerial Vehicles (UAVs) and developed the Weed Spectral Resistance Index (WSRI) for herbicide-resistance screening in Echinochloa crus-galli (L.) P.Beauv. and Abutilon theophrasti Medik. The MERIS terrestrial chlorophyll index, the Ratio Vegetation Index (RVI), Red-Edge Ratio Vegetation Index (RERVI), the Plant Senescence Reflectance Index (PSRI), the Red-Edge Normalized Difference Vegetation Index (RENDVI), the Green Normalized Difference Vegetation Index (GNDVI), and the Normalized Difference Vegetation Index (NDVI) are among the various vegetation indices that have been used to estimate herbicide efficacy, and, in some cases, to distinguish between susceptible and resistant weed biotypes [10,11,12,13,14,15,16,17].

As mentioned above, NDVI is one of these vegetation indices that can be used to rapidly evaluate herbicide efficacy under real field conditions [17]. This index quantifies the normalized difference between near-infrared (NIR) and red (Red) reflectance and can be calculated using broad and narrowband reflectance data [18]. It has been proposed as a non-destructive tool for estimating various parameters of vegetation growth and plant health, since the process of absorption of red light and subsequent reflection of energy in the near-infrared (NIR) range is a characteristic of healthy vegetation [19]. Thus, considering the way NDVI is calculated [20], deterioration of vegetation health due to herbicide application can be detected by the reduced NDVI values. Herbicide application is a factor that causes stress in treated weeds. Inhibition of carotenoids, bleaching of leaves, and lower chlorophyll content in meristematic tissues are the typical symptoms that occur after treatment [21]. Although NDVI is not exclusively aimed at detecting herbicide injury symptoms, it has been used to detect the aforementioned changes in vegetation growth and health status caused by herbicide applications [7].

As recent field trials have shown, scanning the canopy of herbicide-treated weeds with a handheld spectrometer and measuring NDVI can be a reliable estimate of herbicide efficacy in situ, approximately two weeks after treatment [17]. In these cases, low NDVI values are measured after the application of effective herbicides, resulting in herbicide injury symptoms in the foliar area of treated weeds. In contrast, high NDVI values indicate poor herbicide performance as treated weeds continue to grow without showing injury symptoms, either due to herbicide application errors (incorrect application rate or pressure, weed growth stage, etc.) or due to the occurrence of herbicide-resistant populations [17]. In addition, it can be hypothesized that by measuring NDVI when the sensor is held above the crop canopy, herbicide injury symptoms to crop plants can be detected. This was previously reported by Kong et al. [22], who used NDVI to evaluate the selectivity of commercial sulfonylurea herbicide mixtures (mesosulfuron-methyl plus iodosulfuron-methyl-sodium) in bread wheat (Triticum aestivum L.) and highlighted the role of herbicide safeners in preventing crop injury after herbicide applications. In other studies, NDVI was used to evaluate herbicide symptoms in maize (Zea mays L.) and soybean (Glycine max (L.) Merr.) caused by herbicide drift into these crops [19].

The main objective of this study was to evaluate the efficacy of commercial herbicide products against weeds in durum wheat fields at two agricultural sites with different climatic conditions and their selectivity on the crop by using NDVI measurements as an indicator of both parameters in the field and shortly after treatment. To verify that NDVI is a reliable estimate of herbicide efficacy, weed density data were collected later in the season. The effects of herbicide use on durum wheat grain yield data were also evaluated to determine if potential herbicide injury, as determined by NDVI values, corresponded to lower crop yields at the end of the growing season.

2. Materials and Methods

2.1. Site Description

Field trials were conducted during the 2020–2021 and 2021–2022 growing seasons in two durum wheat fields. The first experimental field was located in the Kozani region in the village of Ptelea (40.327° N, 21.896° E), while the second was located near the town of Velestino in Thessaly in the village of Megalo Monastiri (39.442° N, 22.675° E). Regarding crop rotation, durum wheat was grown as a monoculture in all sites in recent years. In Ptelea, the soil texture (0–30 cm) was clay loam with the following characteristics: 48.2% clay, 24.1% silt, 27.7% sand with an organic matter content of 6.09% and a pH of 7.33. In Megalo Monastiri, the soil texture (0–30 cm) was sandy clay loam with the following characteristics: 26.9% clay, 35.3% silt, 37.8% sand with an organic matter content of 6.45% and a pH of 7.47. In both growing seasons, the average monthly air temperature in Kozani was lower than in Velestino. On some days, temperatures at Kozani were very low during the night and early morning hours, falling below −10°C in January and February 2021 and also in January 2022. In February 2022, the lowest temperature was recorded at −6.3°C, but severe freezing temperatures (up to −10.8°C) occurred in March 2022, when durum wheat was in the middle of tillering (Zadoks 25), as is common in this region (

Table 1. Climatic conditions in the two experimental fields during the 2020–2021 and 2021–2022 growing seasons.

Month	Kozani
	Average Temperature (°C)		Maximum Temperature (°C)		Minimum Temperature (°C)		Monthly Precipitation (mm)
	2020–2021	2021–2022	2020–2021	2021–2022	2020–2021	2021–2022	2020–2021	2021–2022
November	7.9	8.8	18.2	22.7	−1.5	−0.6	5.8	59.2
December	6.2	4.0	14.3	13.4	−0.6	−7.6	58.1	48.6
January	4.0	2.0	16.9	16.2	−11.4	−10.4	130.4	37.4
February	5.5	4.5	18.6	17.6	−11.3	−6.3	14.0	26.2
March	5.6	3.0	17.7	18.7	−3.8	−10.8	35.8	42.4
April	10.1	11.1	27.4	25.6	−2.7	0.4	32.2	35.0
May	17.8	17.3	31.2	31.9	7.1	5.8	15.4	13.2
June	20.9	21.9	36.3	33.7	7.9	12.7	66.2	79.6
Month	Velestino
	Average Temperature (°C)		Maximum Temperature (°C)		Minimum Temperature (°C)		Monthly Precipitation (mm)
	2020–2021	2021–2022	2020–2021	2021–2022	2020–2021	2021–2022	2020–2021	2021–2022
November	13.9	14.6	21.9	23.8	6.1	7.4	32.8	50.2
December	13.1	10.7	20.1	20.0	7.2	1.4	74.6	64.8
January	11.0	8.5	22.6	19.3	−2.2	−1.9	58.6	147.0
February	11.1	10.6	21.8	19.8	−1.9	2.4	22.0	71.0
March	11.7	9.5	20.6	23.7	2.3	0.0	50.8	62.4
April	15.4	16.5	26.8	28.1	5.1	7.4	8.0	35.8
May	22.3	21.3	34.6	32.6	12.6	12.0	9.2	11.8
June	25.4	26.8	41.1	38.3	13.4	18.8	17.0	28.8

).

In contrast, the weather was much warmer in Velestino, where air temperatures were higher in all months of the two growing seasons. The lower air temperatures observed in the first experimental field could be due to the higher altitude of the first experimental field in Kozani (695 m) as compared to Velestino (119 m). In Kozani, precipitation was highest in January 2021 (130.4 mm), but much lower in the following months of the 2020–2021 growing season. In the second growing season (2021–2022), precipitation and snow events were similar to those in 2020–2021, except in January 2022, when precipitation was more than 90 mm lower than in January 2021. In Velestino, precipitation was higher in 2021–2022 than in 2020–2021, peaking in January 2022, when 147.0 mm of precipitation was deposited on the ground. In the months between January and April, precipitation was 176.8 mm higher in the second growing season than in the first.

2.2. Experimental Setup and Design

All experimental runs were conducted on already established durum wheat fields. Before explaining the experimental design, general crop management practices are presented. Specifically, according to landowners, the soil of each field was plowed to a depth of about 30 cm in late September and then cultivated twice with a cultivator to a depth of 20 cm (1 week before sowing) to prepare the seedbed. At this time, a complete fertilizer 21-17-0 (YaraMila^® STAR PLUS^®, Yara Hellas S.A., Athens, Greece) was incorporated into the soil to provide the crop with 63 kg N ha⁻¹ and 51 kg P₂O₅ ha⁻¹. Durum wheat was sown on 10 November 2020 and 11 November 2021 in Velestino and on 16 November 2020 and 13 November 2021 in Kozani. The durum wheat cultivar ‘Simeto’ (Greenco Ltd., Thessaloniki, Greece) was grown in both fields; row spacing was 15 cm and the seeding rate was 230 kg ha⁻¹. At the beginning of the tillering stage (Zadoks 21–23), an inorganic nitrogen fertilizer (40-0-0) was applied at a rate of 130 kg ha⁻¹ as top dressing (YaraVera^® AMIDAS^®, Yara Hellas S.A., Athens, Greece) to supply 52 kg N ha⁻¹ to the crop, and another time at the end of tillering (Zadoks 29) at a rate of 70 kg ha⁻¹, corresponding to 28 kg N ha⁻¹. The crop was grown under rainfed conditions. These crop management procedures were identical in both experimental fields.

However, in Kozani, a strobilurin fungicide, namely pyraclostrobin (Comet^® 20 EC, BASF Hellas S.A., Athens, Greece) was applied at a rate of 250 g a.i. ha⁻¹ at the end of the booting growth stage (Zadoks 49) in both growing seasons. This is a common agricultural practice in the wider region to prevent fungal infections and improve crop physiology [23]. Fungicide applications were made using a Gloria^® 405 T pressure sprayer (Gloria Haus & Gartengeraete GMBH, Witten, Germany) with a 2.4 m wide boom equipped with six flat fan nozzles (Teejet^® 8002). The sprayer was calibrated to deliver 400 L ha⁻¹ of water at a constant pressure of 280 kPa. No fungicides were applied in Velestino. No infestation with insect pests was detected in any experimental run.

Regarding the experimental setup, in the middle of the tillering stage of durum wheat (Zadoks 25), we walked each field on the two diagonals to identify the predominant weed species [17]. In Velestino, the field was dominated by the annual broadleaf weed Erodium cicutarium (L.) L’Hér. ex Aiton, Sinapis arvensis L., and Papaver rhoeas L. In Kozani, the dominant weeds were the annual broadleaves Bifora radians Bieb., Galium aparine L., Adonis aestivalis L., and the perennial broadleaf Cardaria draba (L.) Desv., while we also identified several large patches of the annual grass Avena sterilis L. subsp. ludoviciana which is the most widespread and one of the most troublesome weed species infesting winter cereals and legumes in Greek agriculture [24,25,26]. In both fields, positions with representative weed flora were selected in each growing season to conduct the experiments. All experimental runs were conducted in a Randomized Complete Block Design (RCBD) with five treatments repeated four times, resulting in a total of twenty experimental plots. Plots were 1.5 m long and 4 m wide and had an area of 6 m² resulting in a total experimental area of 120 m². Weed-free borders of 0.2 m were maintained between adjacent plots.

In order to repeat the experimental runs at each site, we established new plots in both Kozani and Velestino during the 2021–2022 growing season. This is because if we used the same plots at each site as we did in 2020–2021, the study would not be truly replicated in time, as the second year’s data would show the effects of two consecutive years of treatments on the parameters studied. Therefore, we established new plots at each site in 2021–2022 and repeated the field trial in a total of four site-years in which herbicide treatments were the same. In particular, the treatment list included an untreated control, 2,4-D at the minimum and maximum recommended application rates, a prepackaged triazolopyrimidine herbicide mixture, namely pyroxsulam + florasulam, and also a prepackaged sulfonylurea herbicide mixture, namely mesosulfuron-methyl + iodosulfuron-methyl-sodium (

Table 2. Application rates, commercial products, and manufacturer names of herbicides included in the treatment list in all four site-years.

Treatment	Herbicide	Rate (g a.i./a.e. ha−1)	Commercial Product	Manufacturer
T1 1	-	-	-	-
T2	2,4-D 2	300	CrossbowTM 600 EC	Corteva TM Agriscience Hellas S.A. (Athens, Greece)
T3	2,4-D	600	CrossbowTM 600 EC	Corteva TM Agriscience Hellas S.A. (Athens, Greece)
T4	Pyroxsulam + florasulam	18.82 + 3.71	BroadwayTM 85 WG 3	Corteva TM Agriscience Hellas S.A. (Athens, Greece)
T5	Mesosulfuron-methyl + iodosulfuron-methyl-sodium	15 + 3	Atlantis® WG 4	Bayer Hellas A.G. (Athens, Greece)

¹ Untreated control. ² 2,4-D rate is expressed as g a.e. ha⁻¹. All other herbicide rates are expressed as g a.i. ha⁻¹. ³ The product contained cloquintocet-mexyl as a safener, applied at a rate of 18.82 g a.i. ha⁻¹. ⁴ The product contained mefenpyr-diethyl as a safener, applied at a rate of 45 g a.i. ha⁻¹.

).

It should also be noted that the commercial product containing pyroxsulam + florasulam (Broadway^TM 85 WG) as herbicidal active ingredients also contains cloquintocet-mexyl as a safener at a concentration of 7.1% (v/v). Considering the recommendation on the product label to apply 265 g ha⁻¹, this resulted in the application of cloquintocet-mexyl at a rate of 18.82 g a.i. ha⁻¹. The product containing mesosulfuron-methyl + iodosulfuron-methyl-sodium as active ingredients (Atlantis^® WG) contained mefenpyr-diethyl as a safener at a concentration of 9% (v/v). Since 500 g ha⁻¹ of the product were applied, mefenpyr-diethyl was applied at 45 g a.i. ha⁻¹.

The four herbicide treatments were applied with four separate Elettra Venus^TM 5 pre-pressure sprayers (Viopsec Kalimeris SMPC, Athens, Greece) that were calibrated to deliver 300 L ha⁻¹ of spray solution through a conical brass nozzle (in each sprayer) at a constant pressure of 180 kPa for the 2,4-D treatments (T2, T3) and 300 kPa for the pyroxsulam + florasulam (T4) and mesosulfuron-methyl + iodosulfuron-methyl-sodium (T5) treatments. At the time of herbicide application, the durum wheat plants were in the mid-tillering stage (Zadoks 25). The weeds were between the 4- and 6-leaf growth stages (BBCH 14–16). Applications were made on 9 March 2021 and 13 March 2022 at Velestino and 31 March 2021 and 30 March 2022 at Kozani.

2.3. Data Collection

Approximately two weeks after treatment in each site-year, NDVI was measured using a Trimble^® GreenSeeker^® portable sensor (Trimble Agriculture Division, Westminster, CO, USA). The sensor unit has self-contained illumination in both the red and near-infrared (NIR) regions and measures reflectance in both the red (visible; 660 nm) and NIR (near-infrared; 770 nm) regions of the electromagnetic spectrum according to the following equation [20]:

$$NDVI=\frac{NIR-Red}{NIR+Red}.$$

(1)

Prior to measurements, four 0.25 m² metal quadrats were placed in each plot in areas with uniform weed flora and away from plot edges; each quadrat contained two rows of crop and weeds in the inter-row spaces between them. To measure weed NDVI in each quadrat, the sensor was oriented parallel to the direction of the crop rows. The sensor was held at a height of 40 cm and moved slowly to scan the weed canopy (within the boundaries of the quadrat) in the areas delineated by the plant rows for 5 s to obtain an estimate of the herbicide effect on the leaf area of the treated weeds [17]. To avoid confounding weed and crop NDVI and to ensure that weed NDVI was recorded, crop plants were manually removed from the quadrat immediately prior to measurements. To measure crop NDVI, each quadrat was moved to adjacent areas of the plot and the sensor was held at a height of 40 cm above durum wheat rows and moved slowly to scan the crop canopy (within the boundaries of the quadrat) for 5 s to detect potential injury symptoms due to herbicide application [22]. To avoid confounding crop and weed NDVI and to ensure that crop NDVI was recorded, weeds were manually removed from the quadrat immediately prior to measurements. At Velestino, the exact dates of weed and crop NDVI measurements were 25 March 2021 and 28 March 2022, while at Kozani, weed and crop NDVI’s were measured on 14 April 2021 and 13 April 2022. All NDVI measurements were taken around midday (between 10:00 am and 2:00 pm), as in previous relative studies conducted in winter cereals [17,22].

Weed biomass was also sampled at Velestino on 2 May 2022 and 5 May 2022 and at Kozani on 18 May 2022 and 20 May 2022, in the four 0.25 m² metal quadrats near the center of each plot. The aboveground weed biomass was harvested by cutting the vegetation with scissors at a height of 5 cm [27]. The weed samples were carefully stored in large numbered plastic bags, taken to the Agronomy Laboratory of the Agricultural University of Athens and dried in an oven (DHG-9025, Knowledge Research S.A., Athens, Greece) at a constant temperature of 60°C for 48 h. Using this procedure, the dry weight of all species became constant, and then we measured the aboveground weed biomass per unit area using a digital balance with three decimal places (KF-H2, Zenith S.A., Athens, Greece). Durum wheat was harvested manually with a scythe when the grains were hard and could not be pressed in with a thumbnail (Zadoks 92). Harvest dates were 16 June 2021, and 13 June 2022, in Velestino and 27 June 2021, and 30 June 2022, in Kozani. Two central points were harvested in each plot, which were determined using a 1 m² wooden quadrat. Samples were placed in numbered gunny sacks and taken to the laboratory to measure the number of spikes per unit area (no. m⁻²), the number of grains per spike (grains spike⁻¹), and the weight (g) of 1000 grains (using the digital balance mentioned above) to determine the final grain yield of durum wheat (at a grain moisture content of 13% after air drying).

2.4. Statistical Analysis

Normal distribution of all data was confirmed with the Shapiro–Wilk test [28], while homoscedasticity was validated with Levene’s test [29]. All data were first subjected to a two-way analysis of variance (ANOVA) in which site-years and herbicide treatments were considered fixed effects and blocks (replicates) were considered random effects. If site-year effects or site-year-herbicide treatment interactions were significant (p ≤ 0.05) for the parameters examined, data were analyzed again using a one-way analysis ANOVA to compare herbicide treatments in each site-year. If site-year effects were not significant or no significant interactions were found between site-years and treatments (p ≤ 0.05), data were pooled across site-years and analyzed again by one-way analysis ANOVA to compare herbicide treatments. Means of the initial two-way analysis and subsequent one-way ANOVAs were compared using the Fisher’s Least Significance Difference (LSD) test. All analyses were performed at a significance level of a = 0.05. Linear regressions were also run for each site year between (i) weed NDVI and biomass and (ii) crop NDVI and grain yield according to the linear model

$$y=a+bx,$$

(2)

where the dependent variable y is weed biomass or grain yield, the independent variable x is weed NDVI or the crop NDVI, a is the intercept, and b is the slope of the regression line.

3. Results

Site-years affected weed NDVI, weed biomass, and crop NDVI (p ≤ 0.001). The same was observed for durum wheat grain yield (p ≤ 0.01) while all studied parameters were influenced by herbicide treatments (p ≤ 0.001). Significant site-year by treatment interactions were observed for weed NDVI, weed biomass, and grain yield (p ≤ 0.001) but not for crop NDVI (p ≥ 0.05). In both experimental years, weed NDVI was higher in the field at Kozani compared to the field at Velestino while the reverse trend was noticed regarding crop NDVI. Weed biomass was highest in Kozani in 2020–2021 and lowest in Kozani and Velestino in 2021–2022. The intermediate values corresponded to the second year of the experiment in Velestino. Higher grain yields were observed in Velestino than in Kozani; this was confirmed in both growing seasons. However, because the effects of site-years were significant on all parameters studied, herbicide treatments were compared in each site-year.

Across all environments, T4 resulted in the lowest weed NDVI, while the highest values corresponded to the untreated control (T1). In addition, weed NDVI was higher in T2 plots than in T5 and T3 plots. Crop NDVI was significantly lower in T3 plots than in plots of all other treatments. On average over four site-years, T4 reduced weed biomass by 80 and 94% compared to T2 and T1, respectively. Treatments T2, T3, and T5 reduced weed dry weight per unit area by 80, 88, and 89%, respectively, as compared to T1. In addition, T3 and T5 caused a 38 and 42% reduction, respectively, as compared to T2. Treatment T4 resulted in lower weed biomass than T3 and T5, but these differences were not statistically significant. Grain yield exceeded 5100 kg ha⁻¹ in plots treated with the pyroxsulam + florasulam mixture (T4), while it was less than 3500 kg ha⁻¹ in the untreated plots (T1). Grain yield of T5 was comparable to that of T4. Compared to T2 and T3, T4 and T5 significantly increased durum wheat productivity. Another observation was that T3 and T2 resulted in increases of 14 and 17%, respectively, as compared to the untreated control (T1) when pooled across the four site-years studied. An overview of the site-year means across all herbicide treatments and the herbicide treatment means across all four environments is provided in the following table (

Table 3. The effects of site-years, treatments, and site-year by treatment interactions on weed NDVI, weed biomass, crop NDVI, and grain yield. Values for weed NDVI, weed biomass (g m⁻²), crop NDVI, and grain yield (kg ha⁻¹) were compared between the four site-years (across herbicide treatments) and the five herbicide treatments (across site-years). In the upper part of the table, the P values are listed. The remaining parts of the table show the mean values across site-years and herbicide treatments are presented with standard errors of measurements cited in brackets.

Factor	DF	Weed NDVI	Weed Biomass	Crop NDVI	Grain Yield
Block	3	0.8064	0.7220	0.5987	0.8162
Site-Year	3	<0.0001	<0.0001	0.0001	0.0010
Error (a) 2	9
Treatment	4	<0.0001	<0.0001	<0.0001	<0.0001
Site-year × Treatment	12	<0.0001	<0.0001	0.1593	<0.0001
Error (b) 3	48
Total	79
Site-Year		Weed NDVI	Weed Biomass (g m−2)	Crop NDVI	Grain Yield (kg ha−1)
Kozani-2020–2021		0.572 a (0.019)	162.848 a (16.584)	0.605 b (0.012)	4219.826 b (145.299)
Kozani-2021–2022		0.565 a (0.021)	99.384 c (8.309)	0.585 c (0.017)	4065.811 b (167.805)
Velestino-2020–2021		0.452 b (0.023)	91.384 c (12.908)	0.643 a (0.011)	4526.013 a (207.419)
Velestino-2021–2022		0.449 b (0.020)	135.241 b (15.403)	0.638 a (0.014)	4641.287 a (173.058)
LSD0.05		0.035	19.023	0.017	230.067
Treatment		Weed NDVI	Weed Biomass (g m−2)	Crop NDVI	Grain Yield (kg ha−1)
T1		0.831 a (0.001)	412.367 a (24.519)	0.632 a (0.010)	3468.847 c (96.219)
T2		0.518 b (0.023)	79.311 b (13.905)	0.612 b (0.016)	4223.834 b (207.206)
T3		0.443 c (0.022)	48.523 c (8.415)	0.581 c (0.020)	4074.025 b (99.657)
T4		0.334 d (0.022)	25.352 c (7.847)	0.635 a (0.001)	5119.613 a (236.104)
T5		0.421 c (0.028)	45.468 c (11.819)	0.628 ab (0.014)	4929.869 a (227.792)
LSD0.05		0.032	24.194	0.020	285.235

DF: Degrees of Freedom; ² Error (a): Block × Site-Year; ³ Error (b): Block × Treatment (Site-Year).

).

3.1. Weed NDVI and Weed Biomass

At Kozani, weed NDVI was lowest in plots treated with pyroxsulam + florasulam (T4) in 2020–2021 (0.31) and 2021–2022 (0.33). Mesosulfuron-methyl + iodosulfuron-methyl-sodium (T5) reduced NDVI compared to 2,4-D treatments (T2, T3) and the untreated control (T1) in both growing seasons. In comparison to T5, T4 resulted in lower weed NDVI in 2020–2021, but no significant differences were observed in 2021–2022. NDVI was lower in plots where 2,4-D was applied at a rate of 300 g a.e. ha⁻¹ (T2) than in T1 plots, while levels were even lower when 2,4-D was applied at a rate of 600 g a.e. ha⁻¹ (T3). These differences were similar in the two experimental years at this site (Figure 1a,b).

At Velestino, the sulfonylurea prepackaged mixture (T5) reduced weed NDVI compared to the untreated control (T1) in both experimental runs. The triazolopyrimidine herbicide mixture (T4) tended to reduce weed NDVI compared to T5 in the first growing season, and this difference was statistically significant in 2021–2022. In 2020–2021, weed NDVI decreased most in plots treated with 2,4-D at the highest recommended rate (T3; 600 g a.e. ha⁻¹). Treatment 2 (T2) 2,4-D at a rate of 300 g a.e. ha⁻¹ resulted in slightly higher levels than T3. No differences were observed between these above herbicide treatments, namely T2, T3, and T4, in 2020–2021 (Figure 1c,d).

In Kozani, 2,4-D applied at 300 g a.e. ha⁻¹ (T2) reduced weed biomass by 64 and 72% as compared to the untreated control (T1) in 2020–2021 and 2021–2022, respectively. Application of 2,4-D at a rate of 600 g a.e. ha⁻¹ (T3) further reduced weed biomass by 72% and 84% as compared to T1 in the first and second years of the experiment, respectively. Treatments based on ALS-inhibiting herbicides, namely T4 and T5, resulted in very low levels of weed biomass in 2020–2021 (14–16 g m⁻²) and 2020-2021 (19–22 g m⁻²) and were by far more effective than T2 and T3. Observations were different at Velestino, where T3 tended to be more effective than pyroxsulam + florasulam (T4) and reduced weed biomass by 92% compared to mesosulfuron-methyl + iodosulfuron-methyl-sodium (T5) in 2020–2021. In 2021–2022, weed dry weight per unit area was significantly lower (87%) in the T3 plots than in the T5 plots. No significant differences were found between T3 and T4. Treatment T5 reduced weed biomass by 83% in both experimental years at this site when compared to the untreated plots. In comparison to T1 plots, even further reductions were observed in T2 plots, but T2 was not significantly different from T4 and T5. Although the differences were not significant, increasing the application rate of 2,4-D from 300 g a.e. ha⁻¹ (T2) to 600 g a.e. ha⁻¹ (T3) reduced weed biomass by 81% in 2020–2021 and 75% in 2021–2022 (Figure 2a–d).

Furthermore, approximately 67% and 73% of the variability in weed biomass in 2020–2021 and 2021–2022, respectively, in Kozani could be explained from weed NDVI values. Weed NDVI and weed biomass were also strongly correlated at Velestino, with R² exceed-ing 90% in both growing seasons (Figure 3a–d).

3.2. Crop NDVI and Grain Yield

There was no difference in crop NDVI between the herbicide treatments and the untreated control in Velestino in either the first or second year of the trial. However, application of 2,4-D at the highest recommended rate of 600 g a.e. ha⁻¹ (T3) resulted in significantly lower durum wheat NDVI in Kozani in both 2020–2021 (0.56) and 2021–2022 (0.51) when compared to the other treatments. For 2,4-D, 300 g a.e. ha⁻¹ (T2) tended to cause some reduction in crop NDVI when compared to the untreated control (T1), but the differences were not statistically significant. Treatments with ALS-inhibitors (T4 and T5) did not reduce crop NDVI (Figure 4a–d).

At the end of the 2020–2021 growing season in Kozani, grain yield was 30, 38, and 40% higher in plots treated with pyroxsulam + florasulam (T4) than in plots treated with 2,4-D at 300 g a.e. ha⁻¹ (T2), untreated plots (T1), and plots treated with 2,4-D at a rate of 600 g a.e. ha⁻¹ (T3), respectively. Results were similar in 2021–2022; T4 resulted in 32, 36, and 42% higher yields than T1, T2, and T3, respectively. Mesosulfuron-methyl + iodosulfuron-methyl-sodium (T5) gave similar yields to T4 in both years. Treatment T3 reduced durum wheat grain yield when compared to T2 in 2020–2021 and to T1 in 2021–2022. No significant differences were found between T3 and T1 in 2020–2021 and between T3 and T2 in 2021–2022. Treatments T1 and T2 did not differ in either year of the experiment at this site. In Velestino, all herbicide treatments improved crop yields compared to the untreated control (T1) in both growing seasons. In T1 plots, grain yield was 25, 27, 27, and 29% lower than in T2, T4, T5, and T3 plots, respectively, in 2020–2021. Similar results were obtained in 2021–2022. In general, grain yield was highest in T3 plots, where it remained above 5000 kg ha⁻¹ in both years. However, no statistically significant differences were observed between T3 and the other herbicide treatments (Figure 5a–d).

Significant relationships were observed between crop NDVI and grain yield in Kozani (p ≤ 0.05), but not in Velestino (p ≥ 0.05). These results were the same in both experimental years (Figure 6a–d).

In Kozani, significant linear relationships were observed between the dependent (grain yield) and independent (crop NDVI) variables, but the R² values were quite low, i.e., 0.1720 in 2020–2021 and 0.2925 in 2021–2022. However, it was found that low crop NDVI values in T3 plots corresponded to low grain yield values at the end of the two growing seasons. In contrast, no correlations were found between crop NDVI and grain yield in Velestino, where crop NDVI did not differ between herbicide treatments in either growing season.

4. Discussion

Measurement of weed NDVI in the field was a reliable estimate of herbicide efficacy shortly after treatment, as a strong linear relationship between weed NDVI and weed biomass was observed in the four site-years. This is consistent with previous studies conducted in durum wheat and also in perennial orchards and vineyards [17,30,31]. Low NDVI values recorded by the handheld optoelectronic sensor were due to herbicide symptoms occurring in the leaf area of treated weeds, indicating that the herbicide was effective and resulting in lower weed biomass later in the season; this led to lower weed pressure on the crop. Similar results were obtained by Lewis et al. [32], who also found strong linear relationships between white clover (Trifolium repens L.) NDVI and biomass in tall fescue (Festuca arundinacea Schreb.) stands.

At Kozani, weed NDVI and weed biomass were lower in plots treated with mixtures of ALS-inhibitors than in plots treated with 2,4-D. This was due to the composition of the weed flora and, in particular, to the presence of A. sterilis in the field, which was not affected by the synthetic auxin herbicide (2,4-D) targeting only broadleaf weeds in winter cereals. On the contrary, both the triazolopyrimidine (mesosulfuron-methyl + iodosulfuron-methyl-sodium) and sulfonylurea (pyroxsulam + florasulam) herbicide mixtures have demonstrated their potential to control a broader range of weed species with a single application and at lower application rates [33,34]. It should be noted here that the good performance of pyroxsulam + florasulam is encouraging, as this new mixture may provide solutions when populations with resistance to mesosulfuron-methyl-sodium potentially increase in the future. This is likely considering that this sulfonylurea mixture has been used for some time to control grass weed populations resistant to ACCase-inhibitors in winter cereals in Greece and other Mediterranean countries [6,26,35,36]. However, in Velestino, where the field was dominated by broadleaf weeds, 2,4-D resulted in the lowest NDVI and weed biomass values, as expected given the high efficacy of this synthetic auxin herbicide against broadleaf species [37]. In any case, it has been shown that herbicide efficacy can be site specific. The composition of species in the weed flora at a particular site should determine the choice of herbicides used to achieve optimal weed control. Emphasis should be placed on controlling noxious competitive weeds that affect crop yield and quality. In this context, sustainable cultural practices such as crop rotation, cover cropping, intercropping, false and stale seedbeds, etc., should also be used regularly to manipulate weed flora, avoid the formation of weed communities dominated by only a few noxious weed species, and maintain diversified weed communities that tend to be less competitive [38].

However, 2,4-D decreased crop NDVI at Kozani; the effects were more pronounced when the herbicide was applied at the highest recommended rate of 600 g a.e. ha⁻¹. The low crop NDVI values indicate that the durum wheat plants were injured by the application of 2,4-D. The unacceptable level of injury to the crop plants, especially in 2021–2022, was confirmed at the end of the growing season when significant losses in grain yield were observed in these plots. This was not observed in Velestino and the differences could be attributed to the different climatic conditions at the two sites. In particular, it is very possible that the severe freezing temperatures at Kozani and the large temperature fluctuations between day and night before and after herbicide application stressed the durum wheat plants and affected plant metabolism. Therefore, by reducing crop plant metabolism and growth rate, the cold conditions increased the sensitivity of durum wheat to 2,4-D, resulting in some degree of damage to the crop plants and loss of grain yield later in the season, as shown by relevant studies on bread wheat (Triticum aestivum L.) under adverse climatic conditions [39,40,41,42]. This is also supported by the fact that crop NDVI and grain yield continued to further decline in 2021–2022, when temperatures at this location were lower than in 2020–2021. In any case, the results show that crop NDVI can be useful to evaluate herbicide selectivity or herbicide injury symptoms under real field conditions, as also shown by previous studies in winter cereals and other economically important field crops [19,22].

The absence of injury to the crop plants in plots treated with either pyroxsulam + florasulam or mesosulfuron-methyl + iodosulfuron-methyl-sodium can be explained by the fact that both commercial products contain safeners that prevent the occurrence of phytotoxic symptoms. Specifically, cloquintocet-mexyl was the safener in the commercial product containing pyroxsulam + florasulam as active ingredients, while mefenpyr-diethyl was the safener in the mesosulfuron-methyl + iodosulfuron-methyl-sodium mixture. Safeners reduce herbicide sensitivity in safener-responsive plant species by accelerating the metabolism of herbicides into less effective or inactive compounds [43]. Cloquintocet-mexyl increases the rate of hydroxylation of pyroxsulam in wheat, without appreciable effects on weeds, thus improving the selectivity of the herbicide and preventing the occurrence of phytotoxic symptoms in the crop [44]. As for mefenpyr-diethyl, the conclusions of the recent study by Yuan et al. [45] were that the safener enhanced the expression of glutathione S-transferase (GST) and other molecules that facilitate herbicide detoxification and enhanced the glycosylation of sulfonylurea herbicides in wheat crop plants.

In each site-year, the herbicides that were most effective in reducing weed NDVI and weed biomass while being safe for the crop resulted in the highest grain yields, as in previous studies [17]. At this point, it should be mentioned that the initial site-year by treatment analysis of variance revealed some differences between site-years in weed biomass and grain yield. On average among the herbicide treatments, weed biomass was higher and grain yield was lower in Kozani. This may be attributed to the higher precipitation at this location, which caused greater weed emergence and weed pressure to the crop plants, resulting in yield losses due to weed competition. However, the highest yielding plots in Kozani (T4, T5) showed higher grain yields than the highest yielding plots in Velestino (T2–T5). This could be due to the application of pyraclostrobin in Kozani, a strobilurin fungicide that has been reported to improve durum wheat physiology and have positive effects on grain yield components [23]. Strobilurin fungicides are known to prolong the time durum wheat plants remain green, probably due to the reduction in ethylene synthesis and the protective effect of the antioxidant system, which delays chlorophyll degradation and leaf yellowing [46]. Therefore, the period of grain filling is prolonged and higher grain yields are obtained, suggesting that this is a beneficial cultural practice even in the absence of fungal infections [47].

In any case, the novelty of this work is the potential of a practical and easy-to-use sensor that provides estimates of herbicide efficacy and selectivity in the field and recommends to the farmer which herbicide to use in the same growing season and according to the specific needs of each field. Regarding the limitations of the method, it should be noted that there can sometimes be many potential reasons for recording high weed NDVI values after treatment with a particular herbicide. High NDVI values indicate low herbicide efficacy, suggesting the presence of potentially herbicide-resistant weed biotypes in the field. However, this should be further validated by pot experiments under controlled conditions [17]. It should be recalled that possible reasons for high weed NDVI levels include application errors such as incorrect weed growth stage, application rate or pressure, unfavorable environmental conditions, etc. In addition, the time period in which weed NDVI levels become low enough to allow estimation of herbicide efficacy may depend on the site of action (SoA) of the herbicide and climatic conditions. Regarding selectivity, low crop NDVI values are not always evidence that an herbicide was phytotoxic to the crop, because they could also be due to other external factors such as frost stress, fungal infections or insect infestations, prolonged droughts, etc.

5. Conclusions

The current research has shown that measuring weed and crop NDVI is a smart precision agriculture (PA) practice that can provide non-subjective evidence of herbicide efficacy or and selectivity, respectively, in the field and shortly after treatment. In addition, herbicide mixtures containing safeners appear to be preferable in colder regions with higher elevation to avoid crop injury from single 2,4-D applications at higher application rates. Further field trials are needed to evaluate the methodology used in this work in other crops and under different soil and climatic conditions to overcome any limitations and optimize the method. Such precision agriculture (PA) tools and practices are considered essential for the development of development of Integrated Weed Management (IWM) systems in arable crops and agriculture in general. It should also always be remembered that regardless of herbicide efficacy, thresholds for intervention should be developed to define the actual need for application of a particular herbicide at a particular site. This can help reduce herbicide use in arable crops and align agricultural production with the EU Green Deal goals of reducing pesticide use by at least 50% by 2030.