Identification of Optimal Starting Time Instance to Forecast Net Blotch Density in Spring Barley with Meteorological Data in Finland

Abstract

The performance of meteorological data-based methods to forecast plant diseases strongly depends on temporal weather information. In this paper, a data analysis procedure is presented for finding the optimal starting time for forecasting net blotch density in spring barley based on meteorological data. For this purpose, changes in the information content of typically measured weather variables were systemically quantified in sliding time windows and with additionally generated mathematical transformations, namely with features. Signal-to-noise statistics were applied in a novel way as a metric for identifying the optimal starting time instance and the most important features to successfully distinguish between two net blotch densities during springtime itself. According to the results, the information content of meteorological data used in classifying between nine years with and four years without net blotch reached its maximum in Finnish weather conditions on the 41st day from the beginning of the growing season. Specifically, utilising weather data at 41–55 days from the beginning of the growing season maximises successful forecasting potential of net blotch density. It also seems that this time instance enables a linear classification task with a selected feature subset, since the averages of the metrics in two data groups differ statistically with a minimum 68% confidence level for nine days in a 14-day time window.

1. Introduction

The performance of the modern data-based forecasting tools for plant diseases is highly dependent on the methods applied and on the representativeness of the available data. The temporal characteristics of the information content in meteorological data and its effect on the classification potential of net blotch risk levels have recently been discovered [1]. In this paper, the optimal starting time instance for net blotch risk forecasting in Finnish weather conditions is studied with an analysis framework including signal-to-noise statistics for meteorological data with feature generation proceeding in sliding data windows.

Barley, Hordeum vulgare L., is one of the largest grain crops and in 2020 it was grown on 51.6 million hectares globally [2]. Barley is grown for example as animal fodder and as a source of malt for beverages. It is also common in food products such as breads, soups and stews and in health products. There are several biotic and abiotic pressure factors that challenge barley production. Barley net blotch is one of the most common fungal diseases in barley and is caused by the two following pathogens: Pyrenophora teres f. teres (net form) and Pyrenophora teres f. maculata (spot form). In Finland, net blotch was present in 86% of barley fields investigated in 2009 [3]. According to Jalli et al. [4], leaf blotch diseases, with a severity >50% at DC 73–77, cause an average of 1114 kg ha⁻¹ yield-loss in spring barley in the long term in the Nordic countries. An assessment in 2015 showed that 40% of fields in South Tigray, in Ethiopia, had net blotch and 60% of them had its relative, spot blotch [5]. This can decrease the barley crop by 10–20% of the annual average yield [6,7], but yield losses as high as 40% have been reported [8].

There are several means to combat net blotch and also other foliar diseases: using clean seed from a resistant cultivar, utilising crop rotation and controlling nitrogen. Usually, chemical and biochemical means are additionally needed [9]. Chemical protection saves the crop, but the overuse of pesticides should be avoided. In [10], the benefits and hazards of pesticides are discussed with many examples. The authors look at pesticide use from different aspects, namely the exposure to pesticides of production workers, formulators, sprayers, mixers, loaders and farm workers. Furthermore, the impact of pesticide residues through food commodities are examined widely [10]. The authors summarise this as “Pesticides have contaminated almost every part of our environment” and advocate for finding ways to protect people against the adverse effects of pesticides. There must be a balance between chemical crop protection and the risks caused by pesticides. In the European Union, IPM (Integrated Pest Management) is codified into the form of a directive which needs to be followed by farmers. According to the directive, chemical protection needs to be justified and well-documented [11]. The main idea is to avoid the negative impacts of agrochemicals and use chemical protection only when absolutely necessary. Another driving force is the fact that the European Commission has adopted the proposal to restore damaged ecosystems and nature by 2050 and to halve the use of pesticides by 2030 [12]. This strongly dictates the reduction of pesticides and will lead to more sustainable food systems in the future.

Forecasting is an important tool for the early detection of plant diseases and in evaluating the risks connected to them. It can help in choosing and implementing disease management strategies. The increased amount of information and improved possibilities to process it has made forecasting tools viable for everyday use. A couple of reviews have shown increased interest in these applications [13,14,15].

In practice, the aim of forecasting applications is to avoid routine pesticide sprayings and to help farmers in decision making when planning their chemical crop protection strategy. Risk assessment is often based on pathogen- and plant-specific factors, selected weather parameters, agronomic variables e.g., cultivar resistance or disease pressure, and in some cases on earlier infection data and geological location. Three different risk models and their use in several test fields in five different countries in the Nordic–Baltic region were studied in [16]. In their paper, the models discussed were the Danish decision support Crop Protection Online (CPO), the Danish Humidity Model (HM) and the Finnish net blotch and scald model WisuEnnuste. The authors compared the models’ suitability to predict barley leaf blotch diseases. In the CPO system, the risk assessment for all relevant barley diseases is computed by the number of days with precipitation over 1 mm, information about cultivar resistance and disease data [17,18]. The Danish Humidity Model originally estimated the risk for Septoria tritici blotch in winter wheat and is based on rain events such as the relative humidity or leaf wetness [19]. The Finnish WisuEnnuste [20] has been developed to estimate the field-specific disease risk based on information about the previous crop, the tillage method, the cultivar resistance and certain weather parameters. In [21] the Fourier transform is used in studying the effects of intra-day meteorological changes to Septoria net blotch in winter wheat.

Some examples of plant disease prediction tools for decision making and crop protection are documented and discussed in [22]. The authors focused on Fusarium head blight, which is the major fungal disease that causes losses in wheat and barley production in Canada. The use of fungicides in addition to evolutionary factors have led to more virulent forms of Fusarium head blight. Fernando et al. discuss the utilization of modern prediction tools for plant diseases as well as potential plant defence mechanisms and resistance breeding as a means for plant disease management.

Some other prediction systems for Fusarium head blight are presented in [23] and [24]. One web-based platform that allows Fusarium risk assessment based on parameters such as the geographic location, crop type and weather is the Fusarium Head Blight Prediction Centre in the United States [25,26,27,28]. Generally, the need for an accurate forecasting system for crop protection has been recognized and several applications have been developed around the world.

In [29], a prediction system for barley net blotch is presented and discussed with detail. The computation of the risk index for barley net blotch utilises selected weather variables in different growing zones in Finland. Instead of using original variable values in forecasting, the forecasting accuracy was increased by using features generated from the original data. The feature selection utilised the two-sample t-test. The data originated from the open weather data of the Finnish Meteorological Institute and long-term observations of plant disease severity in different growing zones in Finland; forecasting was performed without field-specific measurements. In that study, the forecasting of barley net blotch densities was carried out with advanced data fusion applied to two different data sets.

The accuracy of data-based forecasting depends on the method applied and on the information content of the utilized data. In [1], it is demonstrated that the amount of information content in data is time-dependent. This means that the accuracy of plant disease forecasting may vary during the growing season. Three different data window sizes (7, 14 and 21 days) were studied in the paper, while the starting point of the prediction varied between zero and 50 days from the beginning of the growing season.

The previously mentioned two papers [1,29] show that feature generation improves the forecasting accuracy and helps to avoid additional field tests and the forecasting accuracy depends on choosing the correct time sample, especially the starting point of forecasting. The research problem in this paper is how to identify and define the optimal starting time instance for the net blotch forecasting in Finnish conditions, so enhancing the performance of plant disease forecasting methods that utilise meteorological data. Based on the earlier results [1], this study uses the window size of 14 days in forecasting. The target is to be able to define automatically the optimal starting point for forecasting from the history database by evaluating the information content of data for every time-step from the beginning of the growing season. The signal-to-noise ratio us used as the metric for the information content of the data.

2. Materials and Methods

2.1. Data

Weather data from the open database of the Finnish Meteorological Institute (FMI) and field observations of net blotch density from the official variety trials database of the Natural Resources Institute Finland (Luke) were utilised. The net blotch data was collected during the period 1991–2017 and the test fields were located in Central and Southern Finland. Net blotch density is divided into two categories:

• Category 1 (very low net blotch density, maximum net blotch severity value of 0.5%);
• Category 2 (net blotch appears in the selected observation fields in these years, severity value of 0.6−5%).

One example of labelling the intensity of plant disease in cereals is presented in [30]. The locations of the test fields and the years of the selected weather data by category are presented in

Table 1. Location of test fields and the years of utilised weather data by category.

Location of Test Fields	Mynämäki N = 6,732,402.033 E = 218,702.907	Jokioinen N = 6,746,822.331 E = 308,359.757	Seinäjoki N = 6,986,750.229 E = 271,138.563	Siikajoki N = 7,174,584.799 E = 408,818.353	Years in Total
Years of observations Category 1	2011	2013	2011	2010	4
Years of observations Category 2	2013, 2014, 2016	2014, 2015	2016	2012, 2014, 2015	9

.

A description of the weather data and pre-processing is presented in detail in [29]. Information on the weather stations related to the data used can be found in Appendix B. In this study, the weather variables analysed were:

• Atmospheric pressure (hPa);
• Relative humidity (RH %);
• Temperature (°C);
• Dew point temperature (°C).

In the data analysis, the daily minimum, daily maximum and daily average values of the above variables were considered.

2.2. General Structure of Data Analysis

The weather variables were first normalised with linear scaling between 0 and 1 based on Equation (1):

$$x^{\prime }=\frac{x-x_{min}}{x_{max}-x_{min}}$$

(1)

where the minimum value (x_min) and the maximum value (x_max) were found from the Category 1 data (low net blotch density). The Category 2 (high net blotch density) data were normalised with their corresponding minimum and maximum values. The beginning of the growing season varies according to the year and observation field because of the varying climate conditions and the geological position. This has been considered in the analysis by selecting the starting point of each data set at the beginning of the growing season instead of a fixed calendar date, as explained below in the Section 2.3.

The main idea of the analysis is to search and rank the time windows where yearly weather data is grouped within the categories, but to at the same time separate Categories 1 and 2. The general concept for analysing the classification properties of net blotch densities based on temporal weather conditions is presented in Figure 1. In addition to the original weather measurements, the feature prototypes (Appendix A), namely mathematical transformations of the original variables, were incorporated into the analysis.

The feature subset values were calculated using the feature prototypes (see Section 2.4) at each time-step from the beginning of the growing season for fifty days onwards, proceeding in sliding windows of fourteen days. The feature values were thus the sum of daily signal-to-noise ratios over the sliding window (Equation (2)). After all the time-steps and feature combinations had been calculated, the optimal starting point was determined as the time instance related to the highest obtained feature values. The 14-day window for data analysis was then repeated 50 times, with the first time window starting from the beginning of the growing season. The recurrence of data windows in the analysis is illustrated in Figure 2.

2.3. Starting Date of Growing Season, Automatic Calculation

In this research, the starting date of the growing season is calculated using weather measurements. The beginning of the growing season is determined as the time when the mean outdoor temperature remains over +5 °C for 10 consecutive days. The estimated time for the beginning of the growing season instead of a certain calendar date, for example the sowing date, enables here the spatially-independent comparison of data sets that may also exhibit different weather conditions related to the measurement location. This especially results in a standard and automatic procedure for triggering the data analysis. This method thus differs considerably from the typical usage of the sowing date, requiring the manual and field-specific insertion of the date into present disease prediction tools.

The data sets applied in this study exhibited differences between the sowing date and the beginning of the growing season. The actual sowing date is on the y-axis at zero value, and the year- and field-related difference (in days) from the beginning of the growing season is represented with the bars in Figure 3.

2.4. Feature Generation

Feature generation was utilised to extract more information for classification between net blotch densities than was available with the original weather variables [29]. Some examples of feature generation techniques are presented in [31,32,33,34,35]. In this study, the feature generation method presented in [36] (p. 50) was applied, and the considered feature prototypes are listed in Appendix A.

The feature subset value of every tested feature was computed with every possible variable combination in every fifty selected time-steps. The classification ability of the generated features was studied with signal-to-noise statistics (see Section 2.5). The generated features were normalised before applying the signal-to-noise statistics to ensure the comparability of the D_sn value (see Equations (2) and (3) below).

The total number of tested feature prototypes was 115 in 715 different combinations in groups of four variables, including generated features from a single to three variables. The number of generated features tested in each time window was 1,973,400. The features were generated as combinations of the minimum, maximum and average of the available weather variables (4) and the calculated Leaf Wetness Duration (LWD) that was computed here on an hourly basis as presented in [37] with rules and their inference as follows:

• If the relative humidity is >87%, then the leaf is humid → LWD = 1;
• If the relative humidity is >70%–<87% and increasing >3% per 30 min, then the leaf is humid → LWD = 1;
• If the relative humidity is >70%–<87% and decreasing >2% per 30 min, then the leaf is dry → LWD = 0;
• If the relative humidity <70%, then the leaf is dry → LWD = 0.

The calculated daily LWD was thus the sum of 24-hourly estimates according to the rule inference above.

2.5. Metrics

In this study, the signal-to-noise statistics were examined for the vectors (here time-series) applied for the classification of the different weather data sets according to the net blotch severity. For example, in [38] the authors have successfully utilised signal-to-noise statistics in the prediction of embryonal tumour outcomes in the central nervous system based on gene expression. The authors developed a classification system based on DNA microarray gene expression data and predicted the risk of selected tumour outcomes.

For signal-to-noise statistics, D_sn, the distance between two vectors a and b, Category 1 and 2, respectively, with their mean values µ_a and µ_b and standard deviations δ_a and δ_b are computed according to Equation (2) [38]:

$$D_{sn}=\frac{({\mu }_a-{\mu }_b)}{\left({\delta }_a+{\delta }_b\right)}$$

(2)

Equation (2) is applied in this study. The calculation of D_sn for the identification of the optimal starting time instance for net blotch prediction proceeds in sliding windows from the beginning of the estimated growing season and the following 50 days, step by step in data windows of 14 days for every generated feature n as a sum of the calculated D_sn daily values:

$$D_{sn}={\displaystyle \sum }_{j=1}^{14}\frac{\left|{\overline{x}}_{nj}\left[{\mathrm{M}}_{C1}\left(1\right),{\mathrm{M}}_{C1}\left(4\right)\right]-{\overline{x}}_{nj}\left[{\mathrm{M}}_{C2}\left(1\right),{\mathrm{M}}_{C2}\left(9\right)\right]\right|}{s_{nj}\left[{\mathrm{M}}_{C1}\left(1\right),{\mathrm{M}}_{C1}\left(4\right)\right]+s_{nj}\left[{\mathrm{M}}_{C2}\left(1\right),{\mathrm{M}}_{C2}\left(9\right)\right]}$$

(3)

where M_C1 and M_C2 are feature matrices generated from scalar observations of weather variables related to the years with data of Categories 1 and 2 (see Table 1), ${\overline{x}}_{nj}$ is the average of the data of the years in question and s_nj is the standard deviation of the same data. In Figure 4, the behaviour of the D_sn index is illustrated. There, Feature number 1 would be ranked as a more plausible candidate than Feature number 2 by comparing their calculated D_sn values in classification of the two groups (o and x). According to Figure 4 and Equation (2) with the same notation of statistical quantities, the resulting value of D_sn for Feature 1 would be much higher than for Feature 2 in this case.

3. Results and Discussion

In the following figures, the resulting D_sn values in the fourteen-day time window are presented for each time step. The daily average, minimum and maximum of the studied variables were tested and the D_sn values were calculated accordingly to every analysed feature subset (Equation (3)).

In Figure 5, the calculated D_sn values for the average, minimum and maximum outdoor temperature are presented. The highest D_sn value, 52.5, is achieved on day 40 from the beginning of the growing season with the daily average outdoor temperature. The D_sn value in the case of the average outdoor temperature remains relatively high between days 37 and 40, but then the value falls rapidly to 28.2 on day 41. On the other hand, the D_sn value of the maximum outdoor temperature achieves its highest value, 43.3, on the same day while the highest D_sn value of the minimum outdoor temperature of 29.3 is achieved on day 29. The D_sn value of the average outdoor temperature higher than the D_sn values of the minimum and the maximum outdoor temperature during the studied period. This indicates that the information content of the outdoor temperature related to the appearance of barley net blotch is the highest during the two weeks starting on days 36–40 from the beginning of the growing season.

In Figure 6, the D_sn values of the average, minimum and maximum daily relative humidity are presented. The highest D_sn value, namely 57, is achieved on day 22 from the beginning of the growing season with the minimum daily relative humidity. The average of the daily relative humidity reaches its highest D_sn value of 45.9 at almost the same time, namely on day 25. The maximum daily relative humidity exhibits the highest D_sn values in the time window 42–49 days from the beginning of the growing season. As can be seen in Figure 5, the D_sn values are relatively high between days 17 and 30 with all three variables, but the maximum daily relative humidity does not achieve its highest value of 49.5 until day 49.

In Figure 7, the D_sn values of the daily average, minimum and maximum dew point temperature are presented. The highest D_sn value, 48.9, is achieved for this weather variable on day 39 from the beginning of the growing season (day 1) with the daily average dew point temperature. In the case of the daily maximum dew point temperature, the D_sn value (42.1) peaks in the same time window, whereas the highest D_sn value related to the minimum dew point temperature (39.6) is achieved on day 29. The classification potential of the dew point temperature to separate the two data sets related to different levels of net blotch risk increases during the growing season until day 40, which can be seen in the ascending trend of all three series in Figure 7.

In Figure 8, the D_sn values for the daily average, minimum and maximum atmospheric pressure are shown. The highest D_sn value, 74.4 is achieved here on day 14 when applying the measured minimum values of the atmospheric pressure. Another peak appears on day 18 and corresponds to a D_sn value of 71.4. Furthermore, the highest D_sn values of the maximum (49.1; day 13) and average atmospheric pressure (68.4; day 14) peak almost in the same starting time instance. During days 21–28, the D_sn values of all three statistical quantities for atmospheric pressure are relatively low. All the D_sn values increase slightly after day 30, but are still considerably lower than during days 11–19.

In Figure 9, the calculated D_sn values for leaf wetness duration (LWD) are presented. The highest D_sn value, 34.5 is achieved when the calculation is started on day 22 from the beginning of the growing season. Relatively high D_sn values also exist between days 27 and 32. The time window when the maximum D_sn values are achieved differs from the peaks presented in Figure 4, Figure 5, Figure 6 and Figure 7. Here, the maximum D_sn value of LWD is at a lower level than the D_sn values of the other analysed weather variables.

Figure 10 shows the boxplots of the D_sn values for the 715 best-ranked features calculated in data windows of 14 days at each starting time instance (day). On each daily boxplot, the central mark indicates the median of the 715 calculated D_sn values of the related features. The bottom and top edges of the boxplot indicate the 25th and 75th percentiles, respectively. The whiskers extend to the most extreme data points and the individual high D_sn values are plotted with ‘o’ markers.

The highest D_sn value, namely 285.8, is achieved on day 41 from the beginning of the growing season with the generated feature prototype number 115 (Appendix A), which is structured here as the combination of three weather variables:

$$\frac{\ln \left(maximum relative humidity\right)}{\ln \left(minimum dew point temperature\right)}\times \ln \left(minimum outdoor temperature\right).$$

(4)

These variables included in the feature are also generally known to affect the risk of net blotch infection. The high values are also present on days 20 (D_sn value 260.9), 40 (D_sn value 256.2) and 42 (D_sn value 276.9). The highest statistical median for the D_sn value of the plotted features appears on day 18, namely 92.9. It can also be concluded from Figure 10 in comparison to the values of single weather variables (Figure 5, Figure 6, Figure 7, Figure 8 and Figure 9) that the classification potential generally increases with the applied features. The highest D_sn values of analysed variables and of the best generated feature are presented in

Table 2. Highest D_sn values of analysed variables, related to Figure 5, Figure 6, Figure 7, Figure 8, Figure 9 and Figure 10.

Variable		Highest Dsn Value	Time of the Best Dsn Value	Related Figure
Daily outdoor temperature	Avg	52.5	40	5
	Min	29.3	29
	Max	43.3	40
Relative humidity	Avg	45.9	25	6
	Min	57	22
	Max	49.5	49
Dew point temperature	Avg	48.9	39	7
	Min	39.6	29
	Max	42.1	39
Atmospheric pressure	Avg	68.4	14	8
	Min	74.4	14
	Max	49.1	13
LWD		34.5	22	9
Feature with the highest Dsn value		285.8	41	10

.

In Figure 11 and Figure 12, the classification potential (sums of 14 days of D_sn values) of the best-ranked feature (Equation (2)) is compared at the starting time instances of days 41 and 18 from the beginning of the growing season. The daily D_sn values are presented with standard deviations. The markers ‘x’ and ‘o’ are the mean values of Category 1 (low net blotch density) and 2 (high net blotch density) data, respectively, and the whiskers describe standard deviation, namely the interval with a confidence level of 68%. In Figure 11, the D_sn values with Category 1 data are generally higher than those of Category 2 for every monitored day. Statistically, the categories differ from each other during nine days out of 14 with a confidence level of 68%, as can be seen from Figure 11.

Figure 12 shows that the mean D_sn values of Categories 1 and 2 are similar on five days out of 14 and the whiskers overlap in every case, namely with a confidence level of 68%. Thus, statistically the D_sn values for the best feature (Equation (2)) are similar in both data sets, leading to poor classification potential.

On the other hand, at least with the generated feature (Equation (4)), it seems that a linear classifier would be sufficient for the task if the analysis starts on day 41 after the beginning of the growing season in Finland. Generally, the results show that the starting time instance strongly affects the classification potential of net blotch risk levels based on meteorological data.

4. Conclusions

The results show that starting the analysis on day 41 from the beginning of the growing season while applying a 14-day data window would maximise the accuracy of forecasting net blotch risk levels spatially in Finland. The results also indicate that the starting date for forecasting can be identified automatically instead of utilising the sowing date. It can be further concluded that the utilization of features (mathematical transformation of variables) increases the net blotch forecasting potential considerably in comparison to the usage of raw weather variables, including leaf wetness duration. Importantly, it is shown that the selection of an appropriate starting time instance is the crucial factor in developing any forecasting methods for net blotch density, based on information exhibited in meteorological data.