Patch spraying of weeds in spring cereals: Simulated influences of threshold level and spraying resolution on spraying errors and potential herbicide reduction

Abstract

A major obstacle to patch spraying of broad-leaved weeds in cereals is a cost-effective method to assess within-field heterogeneity of the weed population. One method could be a camera mounted in front of the spraying vehicle, online image analysis, and field sprayer shifting between ‘on’ and ‘off’ as the predefined weed damage threshold level is reached. Because such a camera will capture a very limited area (<1 m²) compared to the sprayer width (several m), success of this method requires that spraying decisions vary little within boom width, thereby causing few spraying errors. This approach was evaluated by simulations for varying boom widths and three levels of a weed damage threshold model. Potential herbicide reductions compared to blanket application were also simulated. The average potential herbicide reductions estimated as proportions of fields below the threshold, were 60%, 64% and 53% for the original, 25% increased and 25% reduced threshold levels, respectively. The simulated herbicide reductions were not influenced by boom width, but varied significantly between fields, and between threshold levels. As evaluated by spraying errors, the suitability of the suggested approach will increase by decreasing boom width, vary between fields, and in some fields vary between the threshold levels. For boom widths of 15 m and 24 m, the spraying errors were about 10% and 15%, respectively, where omission of spraying areas above the threshold constituted 5% and 8%, respectively.

Keywords:

• Herbicide reduction
• precision farming
• site-specific weed management

Introduction

Weeds generally occur in patches or clusters within agricultural fields (e.g. Dessaint & Caussanel, 1994; Gerhards et al., 1997; Wallinga et al., 1998; Dieleman & Mortensen, 1999; Cousens et al., 2002) which represents a potential for treating only the patches. Lack of cost-effective weed detection methods on a field scale has so far prevented implementation of patch spraying (PS) in cereal crops.

In Europe, cereals are normally grown with a row spacing about 12–13 cm, and herbicide spraying against seed-propagated broad-leaved weeds is conducted when the crop has 3–4 leaves. At this stage crop and weed plants have nearly the same spectral reflectance characteristics. Together with the short row spacing, this has prevented development of image analysis algorithms based on spectral reflectance characteristics, which has proved promising for row crops (Tian & Slaughter, 1998; Steward & Tian, 1999; Onyango & Marchant, 2003). Therefore, other than spectral characteristics, e.g. shape and contour of plants for image analysis have been explored (e.g. Andreasen et al., 1997; Clausen et al., 2000; Pérez et al., 2000; S⊘gaard & Olsen, 2003). Promising results have been achieved for single species recognition without overlapping leaves (Gerhards & Christensen, 2003; S⊘gaard, 2005).

There are economic, environmental and governmental pressures for reducing herbicide inputs (Anon., 2004). The average reduction of herbicide consumption from site-specific weed control in spring cereals ranged from 40 to 54%, and from 50 to 61% in winter cereals (Heisel et al., 1999; Nordmeyer et al., 2003; Christensen, 2005). In Scandinavia cereals are the most widespread arable crop, and in Norway about 90% of the total pesticide use is in cereals, the main proportion being herbicides (Bjugstad, 2000). The potential for reducing pesticide use through PS in cereals should therefore be significant.

Reports on spatial variation of weed density within cereal fields in Norway are few (Fykse & Wærnhus, 1994; Fykse, 2004). Fykse (2004) plotted total weed density along 180 m transects from four cereal fields. Spatial variation in weed density was high, both within field and between fields. Interestingly, there was clear autocorrelation in weed density: positions with high density were neighbours with positions with high density, and vice versa.

Despite several proposed economic threshold models for annual weeds in cereals (e.g. Gerowitt & Heitefuss, 1990; Fykse, 1991; Black & Dyson, 1993), the practical use has been limited because cost-effective methods to assess the weed populations have been missing (Wilkerson et al., 2002). Image analysis is probably an appropriate solution, and from the above literature study, the necessary algorithms will probably be available within a few years. To reduce costs, one digital camera and the ordinary sprayers already in use should be the basis for PS. Therefore we investigated, by means of simulations, the suitability for real-time on/off PS in cereal fields based on the information from one camera. We simulated one camera mounted in front of the tractor taking images along a narrow path in the middle of the sprayer boom, and the nozzles switched on and off as the weed infestation was above or below the damage threshold. To reduce spraying errors, this design requires that spraying decisions are relatively strongly correlated within the actual boom widths. The objectives of this study were to: 1) assess the suitability of using one camera per boom to control ordinary field sprayers on/off for patch spraying of seed-propagated broad-leaved weeds in spring cereals; 2) estimate potential herbicide reductions; and 3) test whether results were dependent on the actual weed damage threshold level.

Material and methods

Weed data were collected in 1993 at three spring cereal fields in Norway (Table I), originally for another type of investigation (Fykse & Wærnhus, 1994). The weeds were counted within quadrats sized 0.5×0.5 m when the crop was at Zadoks 13–14 (Zadoks et al., 1974), the normal time for post-emergence herbicide application. The quadrats were separated by 10 to 18 m (Table I).

Table I. Field and sampling characteristics.

	Voll	Landvik	Kvithamar
Position	59°41′N,10°46′E	58°20′N,8°31′E	63°30′N,10°52′E
Precipitation, mm^1	600	1230	892
Growing days (>5°C)^1	169	202	182
Temperature sum (>5°C), degree days^1	1091	1442	1064
Sampling date in 1993	June 3–4	June 7	June 28
Number of quadrats	190	84	124
Metres between quadrats,
N-S direction:	10, 10.5, 13, 14	11, 12	12, 13, 14
E-W direction:	15, 16, 18	10	10, 13, 14, 15
^1 Normal period 1961–1990.

Spraying decisions ‘spray’/‘not spray’ were based on a threshold model (Fykse, 1991) developed for seed-propagated broad-leaved weeds in spring cereals at Zadoks 13–14 (Table II). If the total weed density exceeded 175 plants m⁻², or the threshold density for at least one of five strongly competitive weed species was passed, the decision was ‘spray’. Furthermore, if the sum of two or more of these five species passed the threshold density for the species with the highest threshold value, the decision was also ‘spray’ (Table II). Presence of Galium aparine often leads to lodging, and the threshold value (1 plant m⁻²) expresses the problems connected with harvest and drying, and not its yield-depressing effect. As an attempt to apply the weed densities from the original threshold model with more confidence, the original density levels were also increased and reduced by 25% (Table II). The original threshold model included also relative crop/weed ground coverage, but as this parameter was not recorded in the current fields, it was omitted in this study.

Table II. Definition and threshold values (weeds m⁻²) of the model after Fykse (1991). Decision is ‘spray’ if criterion XX are met.

Criteria	Definition	Original threshold level	±25% modified threshold levels^1
A	Total density of seed-propagated broad-leaved weed species	175	131–219
B	Density of Galeopsis ssp. L.	25	19–31
C	Density of Chenopodium album L.	45	34–56
D	Density of Brassica rapa ssp. oleifera (DC.) Metzger	20	15–25
E	Density of Stellaria media (L.) Vill.	45	34–56
F	Density of Galium aparine L.	1	1
X	Density of at least two of the species B–F higher than the threshold for the species with the highest value
XX	‘Full model’, one of the criteria A–X fulfilled
^1 For G. aparine not relevant, explained in text.

Spraying decisions (‘spray’/‘not spray’) were taken per quadrat and plotted for each field. These maps were the basis for manual simulations of PS with varying sprayer boom widths based on the spraying decision of one quadrat per boom width. Increased boom widths were simulated for the N-S direction and the E-W direction by including adjacent quadrats. For a given boom width, spraying decisions could either be conspecific of type ‘spray’, conspecific of type ‘not spray’, or heterogeneous including both ‘spray’ and ‘not spray’ quadrats. Proportions of the sum of conspecific spraying decisions per given boom width, hereafter termed ‘conspecific booms’, were plotted against boom width.

The following PS approaches, with constant detection resolution, but decreasing spraying resolution, were simulated: 1) PS-1: spraying decisions assigned to each quadrat independently, simulating a patch sprayer with independently working sections and one camera per section, or alternatively a very narrow sprayer guided by one camera. This would yield the most flexible spraying, and represented the best estimate of potential herbicide reductions through PS for the dataset; 2) PS-2: the same spraying decision assigned to two adjacent quadrats based on weed conditions at the quadrat viewed by the simulated camera. This simulated a wider boom (with no possibility for independent actions by the boom sections) and one camera per boom; 3) PS-3: the same spraying decision assigned to three adjacent quadrats based on the weed status of the middle one, simulating an even wider boom (with no possibility for independent actions by the boom sections) and one camera per boom. Compared to PS-1, using PS-2 and PS-3 would inevitably result in ‘spraying errors’. These were either failures caused by not spraying quadrats with infestations above the threshold (error type 1), or spraying quadrats with infestations below the threshold (error type 2).

Differences between directions, threshold levels and fields were tested with analysis of variance. Tukey tests were applied to detect the actual differences. Linear regression models were fitted to the simulation results, and two-sample t-tests were used to test differences between regression coefficients. Minitab (Minitab Inc., 2003) was used in all statistical analysis.

Results

Mean total weed density per field varied more than five-fold, from 53 to 295 plants m⁻² (Table III). In terms of standard deviation and coefficient of variation, the within-field variation of total weed density was high (Table III). No field had random distribution of total weed density (p<0.005, Table III). The frequency distributions of total weed density were clearly positively skewed at Voll and Kvithamar, but only slightly skewed at Landvik (Figure 1). Landvik showed flatter than normal distribution; elsewhere the distributions were sharper than normal, with high (Kvithamar) and very high (Voll) positive kurtosis (Table III).

Figure 1.  Frequency distributions of total weed density (weeds m⁻²) per field.

Table III. Summary statistics of total weed density and threshold model species per field.

	Voll	Landvik	Kvithamar
Mean total weed density, weeds m^−2	53	295	131
Minimum total weed density, weeds m^−2	0	0	8
Maximum total weed density, weeds m^−2	432	844	628
Proportion of quadrats without weeds, %	5	1	0
Standard deviation of mean density, weeds m^−2	59	244	115
Coefficient of variation, %	112	83	88
Skewness	2.76	0.66	1.99
Kurtosis	11.46	−0.63	4.72
Total weed density randomly distributed^1	10.85	2.19	6.55
Mean density of Galeopsis ssp., plants m^−2	3	0	11
Mean density of C. album, plants m^−2	7	82	18
Mean density of B. rapa ssp. oleifera, plants m^−2	–	1	0
Mean density of S. media, plants m^−2	1	19	5
Mean density of G. aparine plants m^−2	1	–	–
^1 Anderson-Darling normality test statistic. p value < 0.005 for all fields.

Conspecific spraying decisions were more or less aggregated at all fields, and for all threshold levels (Figure 2). Typically, weed patches defined as areas consisting of nearest neighbour ‘spray’ quadrats using the original threshold level increased and/or merged when the threshold level was reduced, and shrank and/or divided if the threshold level was increased (Figure 2).

Figure 2.  Spraying decision maps for PS-1. Filled (▪) and open (□) squares denote ‘spray’ and ‘not spray’ quadrats for original threshold level. Circles (○) mark additionally ‘spray’ quadrats if the threshold level was reduced by 25%. Diamonds (⋄) mark quadrats that changed to ‘not spray’ if the threshold level was increased by 25%.

For proportions of ‘conspecific booms’ (Figure 3), field (p≤0.001), threshold level (p=0.002) and boom width (p≤0.001) were main effects, but not the direction within field. The average reduction in the proportion of ‘conspecific booms’ per metre boom was equal at Voll and Landvik (p=0.918), and less than the reduction at Kvithamar (p≤0.001). The intercept for Voll and Landvik was higher (p=0.008) than the intercept for Kvithamar (Figure 3). For simulated boom widths covering 2, 3 and 4 quadrats, there is a probability of 50%, 25% and 12.5%, respectively, to obtain conspecific decisions if the decisions are randomly distributed. All simulations at Voll and Landvik were well above these limits of randomness, whereas a few simulations were below at Kvithamar (Figure 3). Increased threshold levels caused a significantly higher proportion of ‘conspecific booms’ than reduced threshold levels, both on an average of all fields (p=0.002), and at Voll (p≤0.001) and Kvithamar (p=0.020), but not at Landvik (Figure 3). The proportions of ‘conspecific booms’ decreased linearly with increasing boom widths, and the mean proportions of ‘conspecific booms’ for 15 m and 24 m booms were 69% and 61%, respectively.

Figure 3.  Proportions of booms with conspecific spraying decisions versus simulated boom width. Labels on datapoints are numbers of quadrats per boom width. Horizontal lines mark limits for random distributions of spraying decisions for 2, 3 and 4 quadrats per boom.

The proportions of potential reduction in herbicide use (Figure 4) varied significantly between fields (p≤0.001) and threshold levels (p≤0.001), but not by boom width. Using resolution PS-1, 86%, 36% and 59% of the quadrats were below the original threshold level at Voll, Landvik and Kvithamar, respectively (Figure 4), corresponding to a mean of 60%. Estimates obtained by resolution PS-1 for the increased and decreased threshold levels were 64% and 53%, respectively, but not statistically different from the estimate using the original level. The magnitude of the tested threshold levels was of no influence on reduction at Landvik (Figure 4), whereas at Voll the reduced level gave less herbicide reduction than the increased level (p=0.017), and at Kvithamar the reduced level gave less than both the increased and the original level (p≤0.009). By applying wider booms (PS-2 and PS-3) both smaller and greater individual savings than by PS-1 were observed (Figure 4), but neither linear slope fitted to the mean of all fields and all threshold levels, nor fitted linear slopes per threshold level, were significantly different from zero (p≥0.051). At field level, Landvik was the only one with fitted slope different from zero (−0.4% m⁻¹, p=0.040).

Figure 4.  Proportions of field simulated unsprayed versus simulated boom width. Diamonds are values for PS-1 and dots are values for PS-2 and PS-3. Legend: see Figure 3.

At every field all spraying error types increased linearly with boom width, but less than 1% m⁻¹ (Figure 5). For 15 m and 24 m booms, predicted average total spraying errors (95% confidence intervals in brackets) were about 10% (9–11%) and 15% (14–17%), respectively. Predicted type 1 and type 2 spraying errors for these widths were 5% (4–5%) and 8% (7–9%), respectively.

Figure 5.  Proportions of field with spraying errors versus simulated boom width. Upper row: error type 1, middle row: error type 2, bottom row: sum of error type 1 and error type 2. Legend: see Figure 3.

Generally, Kvithamar had the highest spraying errors, Voll the second highest and Landvik the lowest errors (Figure 5). This ranking was significant for total and type 1 spraying errors (p≤0.046). Main effect of threshold level on spraying errors was observed at Voll and Kvithamar, but not Landvik. At Voll, the reduced level gave higher errors than the increased level for total (p=0.017) and type 1 error (p=0.032). For total error the reduced level also caused higher (p = 0.034) error rates than the original level (Figure 5, bottom row). At Kvithamar, both the reduced (p=0.002) and the original (p≤0.001) level gave higher total error than the increased level.

Discussion

Suitability of simulated PS approach due to spraying errors

Potential for PS of the three fields was indicated by the variation in total weed density (Figure 1), visual aggregation of conspecific spraying decisions (Figure 2) and proportions of the fields below the threshold (Figure 4). These were field-scale results. Suitability of the simulated PS approach requires conspecific spraying decisions to be relatively strongly autocorrelated at the boom-scale too. In Norway, sprayer booms have traditionally been less than 15 m, but up to 24 m wide booms are of growing interest. The predicted average proportions of ‘conspecific booms’ (Figure 3) for boom widths 15 m and 24 m were relatively high: 69% and 61%, respectively. The associated total spraying errors (Figure 5) were about 10% (15 m) and 15% (24 m). During operational PS, it is probably harder for farmers to accept omission of spraying patches above the threshold than accepting spraying areas below the threshold. As the former error type was relatively low (about 5% (15 m) and 8% (24 m)), the simulated PS approach may be acceptable, particularly among farmers using the 15 m wide booms. The expected short- and long-term herbicide cost reductions compared to total costs, and the farmers’ personal experiences and interests, will influence acceptable error rates.

Suitability of the simulated PS approach is likely to vary between fields due to differences in spraying error levels. In the present study Landvik had the lowest errors, and Kvithamar the highest (Figure 5). This could be explained by the spatial distributions of ‘spray’ and ‘not spray’ quadrats at the boom width scale. Whereas Landvik had relatively elongated and well-defined patches of neighbouring ‘spray’ quadrats (Figure 2) and relatively strongly autocorrelated conspecific spraying decisions (Figure 3), Kvithamar had comparatively frequent shifts between ‘spray’ and ‘not spray’ quadrats (Figure 2) and conspecific decisions were less autocorrelated (Figure 3). Generally, a field where weed patches fit the width of the sprayer boom will have fewer spraying errors than a field in which the majority of the patches are narrower than the boom width and/or are very irregular in shape.

As the spraying errors decreased with decreasing boom width (Figure 5), suitability of the simulated PS concept will generally increase as boom width decreases. Based on our results, using selective control of individual 2 m (Stafford & Miller, 1993) and 3 m (Gerhards et al., 2002) boom sections would have reduced the average total spraying errors to about 1% and 2%, respectively. These low error rates require high mapping and spraying resolutions which necessarily involve relatively complex systems that probably are less robust and more expensive (Paice et al., 1998) than our simulated PS approach. In a simulation study, Barroso et al. (2004) found that the critical parameter that determined the economic viability of patch mapping and spraying resolution was the technology costs. Whether very high detection and spraying resolution would facilitate adoption of PS in cereal fields therefore remains an open question.

Is there a relationship between the level of spraying error of a field and the general weed infestation level? Analysing these variables, it seemed that if the average field density of at least one of the threshold criteria A–F (Table II) was close to its threshold value, the field was likely to have relatively high spraying errors. This is because: 1) an average density near the threshold value indicates that a relatively large proportion of the field has weed densities near that value (cf. Figure 1 for criterion A); and 2) weed densities within fields are generally autocorrelated (e.g. Cardina et al., 1997). Consequently, if a quadrat's density is far below or far above the threshold, it is likely that the neighbouring quadrat is also far below or far above, and hence these neighbours are likely to have the same spraying decision. However, if a quadrat's density is near the threshold (a little above or a little below), it is likely that the neighbour is also near the threshold, but it is more a matter of chance whether these quadrats have the same spraying decisions compared to neighbours with densities far from the threshold. Hence, areas with densities near the threshold are likely to cause higher quadrat-to-quadrat variation in spraying decisions than areas having densities far from the threshold. As it is this quadrat-to-quadrat variation that generates the spraying errors, a field with weed densities close to one or more of the threshold criteria is likely to have relatively high spraying errors. For example, at Kvithamar the average weed densities of criteria A, B and C (Table III) were relatively close to the threshold densities (Table II), and this field had the highest spraying errors (Figure 5).

Potential herbicide reduction

The average potential herbicide reductions estimated by PS-1 were substantial, 53–64%, depending on threshold level. These were of the same order as reported from practical PS of seed-propagated broad-leaved weeds in spring barley in Denmark (Heisel et al., 1999) and in winter cereals in Germany (Nordmeyer et al., 2003; Gerhards et al., 2005).

On average, boom width was of no main influence on the potential herbicide reduction, meaning that savings of the same order as estimated by PS-1 can also be expected for wider booms. However, as both smaller and greater reductions than by PS-1 occurred from resolutions PS-2 and PS-3 (Figure 4), the actual saving obtained one for field at least a relatively small one, might depend on the relative camera position and starting point of spraying. In the long term, i.e. several years and/or larger fields, there will be no boom width influence on herbicide savings. The significant, but weakly decreasing reduction by the increasing boom width seen at Landvik (Figure 4) was probably due to the rectangular-like pattern of conspecific spraying decisions (Figure 2) and the limited number of possible simulations. Effects of boom width on the amount of herbicide reduction would probably happen more often in fields where the patterns of conspecific spraying decisions are shaped as rectangles and/or relatively small-sized fields, than in fields with weed infestations of more irregular pattern and/or larger fields.

The relatively high, but expected field-to-field variation in potential herbicide reduction (Figure 4) is explained by different weed infestation levels of the fields (Table III), and is in agreement with, e.g. Gerhards et al. (2005) who reported variation of 20–80% in herbicide savings by PS against broad-leaved weeds in German spring barley. Within a six-year period of PS of dicotyledonous weeds in German winter cereals, mean herbicide reduction could be less than 30% in one year and more than 60% in another (Nordmeyer & Zwerger, 2005). Year-to-year variation in savings must also be expected.

The proportion of a field being above the damage threshold affects the suitability of PS. Barroso et al. (2004) found that the profitability of site-specific Avena sterilis control in Spanish winter barley generally increased as the proportion of the field infested decreased. This is very logical and probably also valid for our simulated PS approach and data. Hence Voll would be the most, and Landvik the least, profitable field to patch spray (in the year of investigation) (cf. Figure 4).

Influence of the tested threshold levels

At Landvik neither proportions of ‘conspecific booms’ (Figure 3), potential herbicide reduction (Figure 4) nor spraying errors (Figure 5) were affected by different threshold levels, whereas at Voll and Kvithamar these variables were influenced. This happened because at Landvik the proportions of ‘spray’ quadrats were relatively stable for different threshold levels (cf. Figure 2): 68% at reduced level and 63% at increased level, leading to a relative difference of only 8% (100*(68%–63%)/63%) between the modified levels. At Voll and Kvithamar, on the other hand, the relative differences in proportions of ‘spray’ quadrats between the modified threshold levels were much higher: 54% and 74%, respectively, and high enough to yield significant effects of the threshold levels on the tested variables. In addition to the expected increase in herbicide savings due to reduced threshold level seen at Voll, and particularly at Kvithamar (Figure 4), higher proportions of ‘conspecific booms’ (Figure 3) and reduced total spraying errors (Figure 5) were observed for the increased compared to the reduced level. The two latter variables, which both depend on the local (boom-scale) distribution of conspecific spraying decisions, were affected because the local patterns between the modified levels were significantly different due to the substantial relative differences in ‘spray’ quadrats. At Landvik, however, the pattern differed only marginally between the different threshold levels (cf. Figure 2).

Because the threshold level affected the spraying errors significantly in some fields, the suitability of the simulated PS approach may vary with the applied threshold level. Even if the 25% increased threshold level generally posed least spraying errors in the present study, there are no logical reasons to believe this is a general rule. Whether the increased or reduced threshold level will give the fewest spraying errors depends on the field-specific pattern and level of weed infestations.

Conclusion

The simulated PS approach caused relatively small spraying errors for the 15 m (10%) and 24 m (15%) booms. Therefore, the simulated approach appears to be a worthwhile method for spring cereal fields. To support this statement, future studies should however include more fields, finer sampling grids, and image datasets should be the basis for the simulations.

Acknowledgements

The Norwegian Institute for Agricultural and Environmental Research (BIOFORSK) is gratefully acknowledged for financial support. The authors also thank Kjell Wærnhus and Kristin Eide, BIOFORSK, for performing the fieldwork.