USE OF MULTIVARIATE ANALYSIS OF VARIANCE AND DISCRIMINANT ANALYSIS TO EVALUATE CHANGES IN WEED COMMUNITY IN WINTER WHEAT DEPENDING ON HERBICIDES DOSE

Agnieszka Kubik-Komar¹, Maria Jędruszczak², Mirosława Wesołowska-Janczarek^1
1 Department of Applied Mathematics, University of Agriculture in Lublin, Poland
² Department of Soil Tillage and Plant Cultivation, University of Agriculture in Lublin, Poland

ABSTRACT

The paper is an attempt at evaluating changes in the weed community in winter wheat grown in short-term monoculture as a result of the application of 100, 75, 50, 25 and 0% of the recommended dose of Puma 069 EW (fenoxaprop-P-ethyl) and Aminopielik 450 SL (2,4D + dicamba). A two-factor field experiment carried out over 1997-2000, which involved tillage method and herbicides doses, provided data analyzed in the present paper, limited to the reaction of weeds to the herbicide doses only. Relative abundance of weeds, representing a given weed community was determined, based on species frequency and relative density of individuals which occur in test area (Ra index). The experimental data were verified with multivariate analysis of variance and discriminant analysis. The data were analyzed separately both in successive research experiment years and field weed infestation evaluation dates, namely I – prior to herbicide application, II – about 15 days after application and III – prior to winter wheat harvest. Significant differences between the average Ra values resulting from herbicides doses were found at all the dates analyzed, except for the spring date (II) in 1999. Besides, with the Mahalanobis distance square (M²) revealed that 75% of the recommended dose caused the greatest changes in the average relative weed abundance throughout the research period. However, with such inconsiderable dose differences, the graphic representation of the discriminant analysis results was hardly clear. Therefore three doses of herbicides (zero, low and high) were applied and at high M² values species strongly related to the dose at respective dates were demonstrated. Dose-specific species pattern was changing depending on the observation date; most frequently such changes were found in APESV, LAMAM, LAMPU, CHEAL, VIOAR, VERAR. There also species connected with one dose only (e.g. CAPBP and MATIN – zero, GALAP, MYOAR – low and STEME – high). In successive experiment years the dynamics of changes in the average relative abundance (Ra) increased. The multivariate statistical methods used are applicable to evaluate changes in the weed community affected by the herbicide dose. The more diverse the levels of the factor studied, the more clear the representation of changes.

Key words: herbicides, multivariate analysis of variance and discriminant analysis, weed control.

INTRODUCTION

In general, when exposed to simplified tillage, which means eliminating some tillage treatments, decreasing the plough depth or applying direct sowing, the number and dry matter of weeds are observed to increase, and therefore it is necessary to use herbicides [10]. In theory, herbicides application is the simplest and the most effective weed control method, however, it has some drawbacks, e.g. a low effectiveness when exposed to unfavorable weather (drought, excessive rainfall) upon application or the time of herbicides activity, which can not be avoided. For example, Antoszek [2] reports on the herbicides effectiveness ranging from 38 to 70%. Yet another drawback can be the type of the biologically active substance, showing a selective effect on some weed species only, thus creating favorable conditions to growth and development of other weeds. Besides, a long-term application of herbicides can disturb the ecol ogical balance. Jędruszczak [8] as well as Zawiœlak and Adamiak [24] found that the long-term application of herbicides in winter wheat decreases the diversity of segetal species and causes domination of resistant species, thus having a negative effect on crops since a domination of a single or a few weed species results in much greater competition than multi-species communities [17,18].

To alleviate a negative effect of herbicides, the use of herbicide mixtures of a greater effectiveness or applying lower herbicides doses are often recommended [1,2], which is justified from the ecological perspective by decreasing both the changes in the biodiversity of the weed community and weed infestation of the crop. Applying lower herbicides doses enhances not only the ecological balance but also economic results. Herbicides are expensive [3] and frequently the production costs go up considerably. Jędruszczak et al. [9] found that in 1999 seed dressing and a single weed control treatment were the only cost-effective measures. Thus researching the effect of changing weed community due to different herbicides doses is justifiable and interesting from the practical point of view. It can be expected that it will make it possible to define the weed species which show a specific reaction to the herbicide dose applied and to foresee the consequences of such reaction, mostly competition-wise (defined competition) as well as the future effects of reproduction (defined reproduction) and applicable statistical methods provide tools to evaluate the character of this dependence.

Since plants in the community interact, investigating the plant communities requires multivariate analyses which incorporate interactions between plants, e.g. multivariate analysis of variance and discriminate analysis. The first one defines whether or not significant changes of a given character occurred as a result of the factor applied in the experiment, while the latter one is used provided that a positive answer is given to that question; this method gives more detailed information about the directions and dynamics of the changes. Such a comprehensive analysis is more and more frequently available in the literature on plant community biodiversity [13], however many papers use univariate analysis of variance, not factoring in interactions between plants, instead of multivariable analysis [19,20]. Earlier reports included only discriminant analysis [3,6] and graphic representations and conclusions based on the results sometimes seem to be insufficiently justified, see Derksen et al. [5].

With the above in mind, a working hypothesis has assumed that varied herbicides doses do not cause significant changes in the weed community and when the working hypothesis has been rejected and alternative one accepted, an attempt has been made at presenting the direction and dynamics of these changes and determining the weed species groups significantly ‘related to’ a given dose, namely non-susceptible to the dose, finally determining the weed infestation of the field in the future. Similarly the applicability of multivariate analysis of variance and discriminate analysis to evaluate the changes was verified.

MATERIAL AND METHODS

The experimental data was provided by the 2-factor experiment carried out by the Department of Soil Tillage and Plant Cultivation, University of Agriculture in Lublin, which investigated the effect of the tillage system (a) and herbicide doses (b) on yielding and weed infestation of ‘Kobra’ winter wheat grown in short-term (3-year) monoculture. The paper uses the date on the reaction of weeds to the herbicide doses, irrespective of the tillage (conventional and three simplified systems which were favorable to weed infestation of the field) since neither of the years nor any of the evaluation dates recorded significant interactions between the weed infestation characteristics (species composition, number and dry matter of weeds) and the experimental factors. Besides with years of monoculture in each tillage system there was found an increased number of weeds [2]. The experiment was performed on the lessive loess soil (34% of silt and clay fraction, slightly acidic, 2.4% of humus) and of good wheat complex, in split-plot design in four replications. Oats was cultivated as a forecrop in the first year of the experiment. The sowing rate of winter wheat was 250 kg·ha^-1 and NPK fertilization was 280 kg NPK·ha^-1 following the proportional composition guidelines. To protect the wheat plants from fungal diseases, Siarkol K 1000 S.C. (2.5 dm³·ha^-1) and Tilt CB 37.7 WP (1 kg·ha^-1) were applied over vegetation.

Tillering wheat plants were sprayed with Puma Super 069 EW (a.i. fenoxaprop P-ethyl) – against monocotyledonous weeds and Aminopielik D 450 SL (a.i. 2.4 D + dicamba) – against dicotyledonous weeds [21]. Monocotyledonous weeds were sprayed early in spring and the dicotyledonous ones at least a week later, except for Puma Super 069 EW in first year of the experiment due to unfavorable weather conditions. Dry eastern wind and low air humidity made early application of Puma Super 069 EW impossible. The herbicides doses in dm³·ha^-1 were as follows:

The weed infestation of wheat canopy was evaluated three times:

1. Before herbicides application (April 24, 1998, March 30-31, 1999 and April 11-12, 2000).

2. On average 15 days after the last herbicide application (May 18, 1998, May 11, 1999, May 4-5, 2000).

3. Before winter wheat harvest (July 20-21, 1998, July 15-16, 1999 and July 4-7, 2000).

Weed infestation of 4 randomly selected 0.25 m² testing areas for each plot was evaluated with the quantitative-and-gravimetric method [14].

The weather conditions evaluation includes temperature and rainfall. Experiment periods differed in their weather conditions. The 1997/1998 growing season was warm and moist, while 1998/1999 rather cool and dry, except for June and July, which in 1999 were exceptionally hot and moist. The 1999/2000 weather conditions were similar to those observed in 1998/1999, although higher temperature was recorded in March and in April and much higher rainfall – in July. In all the growing seasons, the mean temperature and rainfall were higher than the multi-year mean (1966-1995).

To avoid a non-uniform distribution of weed plants of each species, their relative abundance (Ra) was determined following the Conn and Delapp [4] formula:

where: rd (relative density) is the number of weeds of each species which occur in 4 test areas of the plot divided by the total number of weeds obtained from these areas; rf (relative frequency) was calculated as a ratio of the number of test areas where a given species occurred to the number of test areas in which species in total occurred.

The results were verified using two multivariate statistical methods: analysis of variance and discriminant analysis. The first one was used to verify the hypothesis that the herbicide doses applied did not differentiate the mean relative weed abundance. If this hypothesis was to be rejected, the alternative hypothesis was assumed which involved the other method. The results of the discriminant analysis made it possible to calculate the Mahalanobis squared (M²) distances between the groups determined by the herbicide doses, which in turn facilitated defining the mean Ra differences. The significance of M² was determined with the F test [3]. Depending on the number of groups significantly distant from one another, an attempt was made to plot discriminant points scatterplot with biplot of weed species. The plots were used to represent the relationship between species and herbicide doses, which is seen from the sense of the species vectors towards one of the group determined by the herbicides doses [7,16]. Besides, the length of the vector could provide information about the individual contribution of respective species to group discrimination. Since the herbicides doses did not differ much, it was not possible to plot a diagram of this type which would be legible and easily interpretable. M² values were too low and mostly non-significant. Small distances concerned mostly the weed groups when 25, 50 and 100% of the recommended herbicide doses were applied and hence the proposal of the division into 3 groups determined by: 1 – no herbicides, 2 – low dose e.g. 25 and 50% of the recommended herbicides dose and 3 – high dose e.g. 75 and 100% of the recommended herbicides dose. In most cases the new division of the herbicides doses increased the value of M² and allowed plotting the diagrams.

The data were analyzed separately for each year and date of weed infestation evaluation at α = 0.01. Besides, although the experiment was set up in a split-plot design, the analysis of variance was based on a univariate model which is also applied in discriminant analysis and which made the results of both analyses compatible.

Before the analysis the data were transformed as follows: z = arcsin , mostly used for values expressed as percentage [6,22], to enhance the variance homogeneity in groups and a normal distribution of the data. Despite that the number of weed species analyzed was limited because some of them did not meet the requirements of basic assumptions of the methods applied [12,15].

Of 46 weed species in the winter wheat canopy over 3-years study, in 1998 the analysis included 31 species, while in the other years – 20 species each (Tables 1-4). The weed species of a very low Ra index or species whose number of non-zero observations was minimal were eliminated from further study. Besides, the results of the 1999 and 2000 analysis did not cover the species which occurred in 1998 only.

The calculations were made with STATISTICA® while the diagrams were plotted in Excel which used procedures developed by the authors in VBA language.

RESULTS

1998

No average Ra differences between herbicide doses in the first date of weed evaluation were analyzed since it was the first year of the experiment and there was no effect of the herbicides used. The results of the multivariate analysis of variance showed a significant variation in the average relative abundance of the weed community due to the doses of herbicides at the other dates of this year (Tables 3-5). The highest differences in weed infestation following the herbicides used at 0 and 75% of the recommended dose were observed at the second date of observation, while the smallest Mahalanobis distances were found between the groups of 50 and 100% of the recommended herbicide dose (Table 6). The discriminant points scattering diagram for the second date is hardly legible due to no significant differences between the group determined by 50% dose and the other doses (Fig. 1). Also, when divided into 3 groups which included no herbicides, a low dose (25 and 50%) and a high dose (75 and 100% of the recommended one), the groups were not distant enough to be plotted and to make a final interpretation of weed species biplot (Fig. 2).

At the third date of observations, average Ra values differed mostly between 0 and 50% of the recommended dose. Non-significant differences between groups determined by 25, 50 and 100% of the recommended dose (Tables 4 and 6) resulted in overlapping of the discriminant points (Fig. 3). In this case the three group division was favorable as the distance between the groups increased enough to make the weed species biplot legible and easy to interpret (Fig. 4). The greatest individual effect on the group discrimination was shown by APESV, whose vector points to the relationship with the no-herbicide group. CAPBP, VICHI, VERPE, VIOAR coincided mainly with the low dose. GALAP vector located on the border of the groups, indicate a considerably lower average relative abundance of this weed on no-herbicides plots than on the other plots, where no herbicides selection occurs. Out of the high dose-related species, POLLL demonstrated the longest vector, followed by POAAN, TAROF, GASPA, CHEAL and STEME.

1999

The results of the multivariate analysis of variance for the 1999 data show no significant difference in average Ra values in the first date of observation only, before herbicides application (Table 5). One can therefore state that the effect of various herbicide doses on the changes in average relative abundance of weed community was not yet observed in the successive year of the experiment. Because of non-significant differences in average Ra values between herbicide doses recorded at that evaluation date, the discriminant analysis was made for the second and third date of observations only. The second date of observation (after herbicides application) revealed the greatest significant differences in the average relative abundance of weed community following the application of 0 and 75% of the recommended dose (Table 7). Besides there were also observed significant differences in average Ra values between no-herbicides dose and others as well as following the use of 100 and 75% of the recommended dose. The lowest differences were found between groups of 25 and 50% of the recommended dose. The biplot shows clearly the discriminant points scattering of the groups of no-herbicides and 75% of the recommended dose (Figs 5 and 6). Points representing the other groups overlap. When applying the three-group division, the distances between groups were greater (Fig. 6), however, there was no clear border between groups determined by the low and high dose. Plotting the weed species biplot of the discriminant point scattering would be, in this case, pointless since it would be difficult to classify clearly the species whose vectors would be directed towards them.