Forecasting Alternaria Leaf Spot in Apple with Spatial-Temporal Meteorological and Mobile Internet-Based Disease Survey Data

Abstract

Early warning of plant diseases and pests is critical to ensuring food safety and production for economic crops. Data sources such as the occurrence, frequency, and infection locations are crucial in forecasting plant diseases and pests. However, at present, acquiring such data relies on fixed-point observations or field experiments run by agricultural institutions. Thus, insufficient data and low rates of regional representative are among the major problems affecting the performance of forecasting models. In recent years, the development of mobile internet technology and conveniently accessible multi-source agricultural information bring new ideas to plant diseases’ and pests’ forecasting. This study proposed a forecasting model of Alternaria Leaf Spot (ALS) disease in apple that is based on mobile internet disease survey data and high resolution spatial-temporal meteorological data. Firstly, a mobile internet-based questionnaire was designed to collect disease survey data efficiently. A specific data clean procedure was proposed to mitigate the noise in the data. Next, a sensitivity analysis was performed on the temperature and humidity data, to identify disease-sensitive meteorological factors as model inputs. Finally, the disease forecasting model of the apple ALS was established using four machine learning algorithms: Logistic regression(LR); Fisher linear discriminant analysis(FLDA); Support vector machine(SVM); and K-Nearest Neighbors (KNN). The KNN algorithm is recommended in this study, which produced an overall accuracy of 88%, and Kappa of 0.53. This paper shows that through mobile internet disease survey and a proper data clean approach, it is possible to collect necessary data for disease forecasting in a short time. With the aid of high resolution spatial-temporal meteorological data and machine learning approaches, it is able to achieve disease forecast at a regional scale, which will facilitate efficient disease prevention practices.

1. Introduction

Plant diseases and pests are the key threats to the quality and yield of economic crops. According to the early warning of diseases and pests, guidance on prevention will improve control efficiency, reduce the use of pesticides, and ensure environmental safety. At present, much research has focused on clarifying the relationship between the factors influencing the occurrence and degree of plant diseases and pests and a corresponding mathematical model to help forecast their emergence. Among these factors, the meteorological factors are frequently used as model inputs, given their significant influence on the occurrence of diseases and pests, data availability, and a high level of data standardization. For example, it was found that the occurrence of cotton bollworm and black spot can be forecasted by temperature, humidity, rainfall, wind speed, sunshine duration, evaporation, and 74 atmospheric circulation indices [1]. In forecasting apple scab disease, Wrzesień et al. [2] reported that the rainfall, humidity, temperature, and wind speed may well be associated with the disease occurrence. Lee et al. [3] modeled the probability of inter-annual occurrence of pine nematode disease with temperature, rainfall, altitude, slope, and land-use types. Based on the random forest (RF) and Maxent algorithms, a forecasting accuracy of 76% is achieved. Besides, Bhardwaj et al. [4] established a forecasting model between temperature, relative humidity, sunshine, rainfall, and the incidence of powdery mildew on oats with the LR method.

In constructing a forecasting model for plant diseases and pests, besides the forecasting factors, the availability of the occurrence/incidence/severity data of the disease or pest is crucial and is usually a restrictive point, given the scarcity of such survey data [5,6]. Although the traditional field campaign and sampling method is relatively accurate [7,8], it is difficult to meet the needs of forecasting due to insufficient data and lack of representativeness to the region. However, with the continuous development of mobile internet technology, the electronic survey is able to be efficiently deployed to collect information directly from growers, which thus greatly facilitates the collection of field survey data. For example, Laurett et al. [9] collected data on agricultural production, farm ecological environment, and farmers’ information (i.e., gender, income, education background) from 300 family farmers using an electronic survey, and used the data to identify factors reflecting sustainable development in agriculture. In addition, Rana and Moniruzzaman [10] used web-based surveys and in-depth interviews to collect data on socio-economic conditions, planting patterns, agricultural productivity, and perceptions of climate change from farmers, which yielded data for studying the relationship between agroforestry patterns, rural livelihoods, and climate change. Therefore, owing to the convenience for dissemination and flexibility of question setting, the web-based questionnaire method is an efficient way in collecting abundant data for analyzing and modeling purpose. However, given the fact the different levels of the participants’ knowledge, experience, and attention (i.e., in some cases, casual answers may mislead the modeling process) etc., it is inevitable that the data would contain some noise and subjectiveness information [11,12,13]. Therefore, it is important to properly design the questionnaire and clean the data prior to using them for constructing a disease forecasting model. However, corresponding research is lacking.

China is the largest apple producing country accounting for about 50% of the total apple production globally [14,15]. Recently, diseases and pests are major threats to apple production. Alternaria leaf spot (ALS) is one of the main diseases in apple production, which occurs under high temperatures and humidity [16,17]. The disease affects the apple trees’ leaves and fruits, resulting in early defoliation of leaves, weakened trees, and reduced fruit yield and quality [18]. Until recently, effective methods to forecast the occurrence of ALS at a regional scale have been lacking.

To answer this call, we combined data on apple ALS obtained by a mobile internetbased questionnaire and high spatial-temporal resolution meteorological data to forecast the disease. The main tasks were as follows: (1) to design a mobile internet-based questionnaire and develop a data clean procedure to obtain disease occurrence data that is able to support a disease forecasting model at a regional scale; (2) to analyze the sensitivity of meteorological factors (i.e., temperature, humidity) on the disease occurrence, and identify the appropriate feature setting for disease forecasting; (3) to construct and validate a forecasting model of apple ALS using machine learning approaches.

2. Materials and Methods

2.1. Study Area

The study was conducted at Linyi County (110.732405° E, 34.196514° N) in Shanxi province and Qixia County (120.781994° E, 37.379757° N) in Shandong province, which are major apple producing areas in China (Figure 1). The area under apple production in Linyi County is about 47,000 hectares, and 85,000 hectares in Qixia County. The two study areas have different local environmental and ecological characteristics. Linyi County is characterized by flat terrain, winter rain and sparse snow, and summer rainfall, while Qixia County is located in a hilly mountainous area, with rain and heat appearing in the same season. Therefore, the two study areas provided diverse and representative conditions for studying and forecasting the occurrence of apple ALS.

2.2. Data Collection

2.2.1. Meteorological Data

The biological characteristics of the apple ALS indicate that the disease incidence is closely associated with temperature and relative humidity conditions [19]. Therefore, both meteorological parameters were selected as modeling features for disease forecasting. To indicate the spatial and temporal variation of the two parameters, a high resolution (i.e., spatial temporal (i.e., 1 km, 1 h) reanalysis meteorological data, the HRCLDAS-V1.0 product was used in the analysis (Figure 2). The data covered both regions and the years 2018–2020 (Figure 3). The HRCLDAS-V1.0 product uses data fusion, assimilation, and terrain correction techniques to integrate ground, satellite data, and numerical models [20]. Compared with conventional meteorological data set on coarse resolution, the HRCLDAS-V1.0 product significantly enhances the spatial temporal resolutions and is able to delineate variations of the meteorological factors within a region, which is necessary for regional disease forecasting.

2.2.2. Geographical Data

In order to present the geographical location and scope of the study area and provide the land-use type of orchard location for the subsequent data clean, the boundary data of Chinese administrative regions and 30-meter land cover data, GlobeLand30, were used in this study (obtained from National Geomatics Center of China on 10 January 2021, https://www.ngcc.cn/). The GlobeLand30 product includes 10 land cover types: farmland, forest, grassland, shrubland, wetland, water body, tundra, artificial surface, bare area, glacier and firn, which were used as a background map in this study [21,22].

2.2.3. Design of the Mobile Internet Based Questionnaire for Apple ALS Survey

Data on apple ALS were collected from orchard growers using a web-based questionnaire. This disease electronic survey was conducted from 15 July 2020 to 31 July 2020. The questionnaire was designed to collect information about the occurrence of ALS in apple orchards from 2018 to 2020 as well as the information about orchards’ management practices.

Table 1. Summary of questions in the mobile internet-based survey questionnaire.

No.	Question	Types	Options	Notes
1	Gender, age and contact information of the respondents.	Gap filling
2	Education level of the respondents.	Multiple choice	Middle school or below, Undergraduate, Graduate or above
3	Where is the orchard?	Gap filling
4	What is the area of the orchard?	Gap filling		Unit: hectare
5	What the varieties of apple are planted?	Gap filling
6	What is the age of the apple tree at present?	Gap filling
7	What is the annual output of apples (2018–2020)?	Gap filling		Unit: kg/ha
8	Did apple ALS occur in orchards (2018–2020)?	Multiple choice	Yes, No
9	Whether the orchard is subject to disease and pest prevention and control (2018–2020)?	Multiple choice	Yes, No
10	If you carry out disease and pest prevention and control, how do you control it (2018–2020)?	Multiple choice	Spraying pesticide, other conditions

summarizes the questions about growers’ information, geographical location of the orchards, occurrence of apple ALS, varieties and age of the apple trees, annual unit yield per orchard, disease prevention and control practices, etc. The meteorological data corresponding to the samples were extracted according to the geographic coordinates of each questionnaire record, to form a dataset for constructing the forecasting model of the apple ALS.

2.2.4. Multilevel Data Clean Strategy (MDCS)

Although the mobile internet-based questionnaire can collect data efficiently, the data is prone to mix with noise, which would affect the performance of the model. Therefore, it is necessary to conduct a data clean procedure prior to using them for modeling. Here, a four-step procedure is proposed to conduct such a data clean procedure:

Step 1—preliminary screening: conduct the preliminary screening to ensure the basic information in the questionnaires is correct. In this step, any questionnaires containing the same orchard address and filled out in less than three minutes are considered as an invalid record and are discarded;

Step 2—geographic cross-check: The geographic locations of the questionnaire records are cross-checked and filtered in this step. The orchard locations are checked referring to the Globaland30 land use map. The questionnaires corresponding to orchards located on some apparently impossible classes (e.g., artificial surface, water body, etc.) are defined as invalid records and are discarded;

Step 3—economic analysis: It is assumed that the disease-infected orchards would suffer a certain degree of yield loss. Therefore, the orchards that recorded a serious disease infestation but have higher yields (by comparing the average yield of those disease-free orchards) are considered as a logical paradox. The corresponding questionnaires are discarded;

Step 4—spatial aggregation analysis: In case the incidence of the apple ALS is frequent in some parts of the region, it is indicated that the meteorological and environmental conditions in these areas are suitable for disease occurrence and epidemic. However, there are still some uninfected orchards in the same region due to the differences in varieties and control strategies. Given that these disease-free orchards shared a similar environment with those diseased orchards, it is anticipated that the inclusion of these data would inevitably interfere with the development of the disease forecasting model. To avoid such impact, by using a 3 km × 3 km grids, according to the criteria of disease incidence >60%, the disease aggregated grids were sorted out and the disease-free points were discarded. This step was different from steps 1–3 that aimed at controlling the data quality, as this step is designed to avoid possible information confusion.

A total of 231 survey samples were obtained from 2018 to 2020. Among them, 15, 12, 3 and 47 samples were eliminated during the first, second, third, and fourth steps of the data clean process. Thus, 154 (66.67%) samples with valid data were retained. To evaluate the influence of the data clean procedure on the model performance, the model calibrated with the cleaned data is compared with the model calibrated with the original data.

2.3. Feature Selection of Meteorological Data

ALS mainly infects tender leaves and spring shoots of the apple tree leaves at the leaf sprouting and spreading stages [23,24]. The fruit bagging, pruning, pesticide spraying, and other disease and pest control practices are usually conducted at fruit expansion stage to promote the growth of apple trees and development of fruits [25,26]. Therefore, the temperature and relative humidity during the apple leaf spreading stage (March) to the fruit expanding period (June) were used as candidate input data for constructing the forecasting model of apple ALS.

To mitigate random data fluctuation of the meteorological data and retain its general pattern, we calculated the ten-day averages of temperature and relative humidity. To assess the sensitivity of the meteorological factors to the disease incidence at different stages, a t-test [27] was performed. The factors with p-value < 0.01 were identified as sensitivity factors. Further, to further eliminate redundancy among features, a Pearson Cross-Correlation Analysis (PCCA) [28] was performed on those sensitivity factors. By traversing all pairs of the factors, for pairs of factors with a correlation coefficient (R) higher than 0.8, the relatively less sensitive factor was removed in each pair till the correlation coefficient of all pairs of factors below 0.8. The retained features were then used for construction of the disease forecasting model.

2.4. Development of Disease Forecasting Model

Considering there are a number of machine learning algorithms, it is impossible to try every algorithm. Therefore, the forecasting model of apple ALS with temperature and humidity was established using four representative algorithms, with different principles and characteristics selected, including Logistic Regression (LR), Fisher Linear Discriminant Analysis (FLDA), Support Vector Machine (SVM) and K-Nearest Neighbors classifier (KNN). Among them, the LR [29] is a classical method that is based on statistical theory and has the ability to provide probabilities for classification. The FLDA [30] is a linear classifier that projects a p-dimensional feature vector onto a hyperplane, and has a strong explanatory trait. The SVM [31,32] is a learning model that is effective in high dimensional spaces by transforming the data with kernel functions. The KNN [33] is a nonlinear learning algorithm that adopts an easy-to-understand distance criteria in classification. All the above algorithms have relatively simple principles and low computational complexity, which are frequently used in constructing agricultural forecasting models.

Prior to model construction, the datasets were randomly assigned into five equal groups (including 4 groups with 31 samples and 30 samples in the other group), three of which were used as training sets and the other two as validation sets. By traversing all 10 splits of data, the accuracy of the forecasting model was evaluated using four indexes [34,35]: overall accuracy (OA); Kappa; false-positive rate (FPR); and false-negative rate (FNR). Their definitions are provided as follows:

$$\mathrm{OA}=\frac{\mathrm{TN}+\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{TN}+\mathrm{FN}}$$

(1)

$$Kappa=\frac{{{\displaystyle \sum }}_{\mathrm{i} =1}^{\mathrm{r}}{\mathrm{X}}_{\mathrm{ii}}-{{\displaystyle \sum }}_{\mathrm{i} =1}^{\mathrm{r}}({\mathrm{X}}_{\mathrm{i}+}\times {\mathrm{X}}_{+\mathrm{i}})}{{\mathrm{N}}^2-{{\displaystyle \sum }}_{\mathrm{i} =1}^{\mathrm{r}}({\mathrm{X}}_{\mathrm{i}+}\times {\mathrm{X}}_{+\mathrm{i}})}$$

(2)

$$\mathrm{FPR}=\frac{\mathrm{FP}}{\mathrm{FP}+\mathrm{TN}}$$

(3)

$$\mathrm{FNR}=\frac{\mathrm{FN}}{\mathrm{FN}+\mathrm{TP}}$$

(4)

where TP is the number of correctly judged disease-free orchards; FP is the number of diseased orchards that were mistakenly judged as disease-free orchards; TN is the number of correctly judged diseased orchards; FN is the number of disease-free orchards that were mistakenly judged as diseased orchards; r represents the number of rows and columns in the confusion matrix; X_ii represents the number of samples in row i and column i; X_i+ represents marginal total of row i; X_i+ represents marginal total of column i; and N represents the total number of samples. The workflow of the present study is illustrated in Figure 4.

3. Results

3.1. Sensitive Features for Disease Forecasting

Even at a county level, a certain degree of temperature difference can be observed from its spatial distribution map (Figure 5). The average temperature from March to June in the ALS infected orchards is between 14–16 °C, which is significantly higher than that of the disease-free orchards (12 °C).

According to the sensitivity analysis, the temperature and humidity on multiple stages showed significant differences between the normal and diseased samples (p-value < 0.01). Thus, 11 features were selected for temperature and 10 for relative humidity (

Table 2. The t-test sensitivity analysis of ten-day average temperature and ten-day average relative humidity between healthy and diseased samples from March to June.

Time (Ten Days)	Early March	Mid March	Late March	Early April	Mid April	Late April	Early May	Mid May	Late May	Early June	Mid June	Late June
p-value (Temperature)	**	**	**	**	**	**	**	**	**	**		**
p-value (Humidity)	**	**	**	**	*	**	**	**	**	**		**

Notes: * indicates p-value <0.05, ** indicates p-value <0.01.

). The correlation analysis among these features revealed that some of them are highly correlated (correlation coefficient exceeds 0.8, Figure 6). After removing the redundant features, the temperature in mid-May, late May, and early June and the relative humidity in early April, late April, early May, mid-May, and late June were retained (Figure 7), which were used as input variables in constructing the forecasting model of apple ALS.

3.2. The Forecasting Model of Apple ALS

The forecasting results of apple ALS based on the four algorithms are presented in

Table 3. Forecasting results of apple ALS under four machine learning algorithms.

Split	Logistic		FLDA		SVM		KNN
	OA	Kappa	OA	Kappa	OA	Kappa	OA	Kappa
1	0.89	0.69	0.87	0.63	0.92	0.69	0.87	0.63
2	0.84	0.56	0.87	0.67	0.9	0.72	0.92	0.73
3	0.89	0.65	0.87	0.67	0.92	0.76	0.90	0.71
4	0.90	0.71	0.89	0.70	0.93	0.8	0.90	0.67
5	0.85	0.35	0.87	0.67	0.90	0.65	0.85	0.40
6	0.85	0.35	0.85	0.64	0.90	0.65	0.87	0.45
7	0.84	0.24	0.89	0.70	0.90	0.65	0.87	0.45
8	0.90	0.65	0.85	0.64	0.94	0.78	0.85	0.40
9	0.85	0.40	0.89	0.70	0.87	0.49	0.85	0.40
10	0.84	0.24	0.87	0.67	0.90	0.67	0.87	0.45
Mean	0.87	0.48	0.87	0.67	0.91	0.69	0.88	0.53
Mean FPR	0.49		0.00		0.31		0.49
Mean FNR	0.05		0.16		0.04		0.03

Notes: An n-fold (n = 5) cross-validation is adopted which generates 10 results (with different splits of training and validation samples).

and Figure 8. Based on the spatial distribution of the projected disease incidence (Figure 8), Linyi County had the more severe incidence of apple ALS than Qixia County. Across the three year period, the disease occurred most seriously in 2018, which is in agreement with actual disease survey results. Among the forecasting algorithms (Table 3), the accuracy of LR was the lowest, with the Kappa coefficient fluctuating between 0.24–0.71 under different training and validation sample divisions. For FLDA (Table 3, Figure 8b), despite the error rate of both types being low (i.e., FPR = 0.00, FNR = 0.16), the spatial distribution of the forecasting results in each year did not reflect the actual inter-annual pattern of apple ALS occurrence. For the SVM algorithm (Table 3), OA accuracy and Kappa were the highest among all the algorithms. For the KNN algorithm (Table 3), the OA accuracy under different training and validation sample divisions was higher than 85%, while Kappa was above 0.4.

4. Discussion

In the present study, the disease survey data on apple ALS were obtained over a short period of time with the aid of the mobile internet-based questionnaire approach. However, due to a difference in agricultural knowledge/experience, subjective attitude, and possible memory bias of the survey participants, a certain degree of noise was inevitable. In this study, based on original data, the model yielded OA between 64–83% and the highest Kappa of 0.31. Although the OA accuracy of the original data was high, almost all the diseased orchards were missed by the forecasting model, which may be caused by the noise in the original data. To account for this issue, a multi-step data clean procedure was proposed to enhance both quality and self-consistency of the survey data. The model based on the cleaned data yielded significantly higher accuracy, with an OA of 91% and Kappa of 0.69. Therefore, through the mobile internet-based questionnaire and data clean procedure, it is possible to generate a dataset that is able to support the development of a disease forecasting model.

In the modeling process, the delineation of the spatial difference of meteorological parameters is crucial. The high resolution spatial continuous meteorological products can reflect the spatial variation of temperature and humidity at a regional scale. For example, the temperature in Linyi County had a decreasing trend from southwest to northeast, and relatively high temperatures and a low level of humidity was found in some areas in southwest Linyi County (Figure 5). Such a pattern is important to explore the relationship between meteorological parameters and disease occurrence.

The apple ALS usually overwinters as mycelium on injured leaves, damaged branches, and dormant buds [18]. In the following spring, the overwintered mycelium infects the young leaves and spring shoots of apple trees through conidia formation [25]. Therefore, if the relative humidity is high in April and May and the temperature is suitable for fungal growth, the spores will germinate and spread rapidly through airflow and rain, leading to an epidemic [19]. Following the infection, the rise in temperatures in May and June subsequently gradually shortened the disease incubation period, reaching the incidence peak of apple ALS [24]. The sensitive factors that were screened by the present study (i.e., temperatures in May and June and humidity in April and May) embodied the infection mechanism. According to the forecasting results, the disease incidence in Linyi County is more severe than that in Qixia County. The temperatures and humidity in Linyi County were higher than those in Qixia County in May and June, which is conducive to the development and epidemic of apple ALS.

In the modeling process, the four algorithms returned different results. The results yielded by the LR, SVM, and KNN algorithms indicated an interannual change in the trend of the disease. The FPR of the LR and KNN algorithms was 0.49, implying a high level of commission error. Unlike the SVM, the KNN produced better accuracy than the LR algorithm, and showed stable results under different data splits. In addition, compared to other algorithms, the forecasting results of FLDA yielded an obvious omission error. Therefore, comprehensively considering the model accuracy and generalizing capability, KNN is recommended in this study to establish a forecasting model of apple ALS.

This paper shows that mobile internet disease survey data with reasonable data clean procedure can effectively bridge the gap between the survey data of plant diseases and pests and the forecasting models. Under the condition of a sufficient quantity of data with controllable quality, as well as high resolution spatial-temporal meteorological data, the quantitative forecasting models can be established efficiently. Such a working scheme thus greatly facilitates the deployment of these forecasting models in various scenarios and keeps updating the models to improve their performance [36,37]. The outcomes of the disease forecasting models can provide important information in guiding the control practices of crop diseases, and help reduce the input of pesticides to mitigate negative impacts on the environment [38].

Regarding the future research plan, improvements are expected in data acquisition and modeling approaches. For example, the way of conducting a disease survey can be improved. Besides the questionnaire, the disease incidence can be indicated using image-based or spectral-based approaches, to mitigate possible subjective error [39]. In addition, some observation data from wireless sensors networks (WSNs) and satellite remote sensing can also be included in the forecasting models to indicate environmental conditions and the growing status of host plants [40,41]. For model structure, it is worth attempting to introduce some mechanism-based model (e.g., disease development model) to enhance the robustness of the forecasting [42,43]. Such efforts are important to promote the green control of orchards’ diseases and pests.

5. Conclusions

Aiming at the problem of insufficient disease occurrence data and lack of regional representativeness, this study designed a mobile internet-based questionnaire and developed a data clean procedure to obtain disease occurrence data. Then, by combining meteorological factors and web-based disease survey data, the LR, FLDA, SVM and KNN machine learning algorithms were used to construct and validate forecasting model of apple ALS. The main conclusions are as follows:

(1)

Based on the disease survey data that were obtained by the web survey, the noise is expected to be mitigated according to a purposely developed multilevel data clean strategy;

(2)

In analyzing the relationship between the occurrence of apple ALS and high-resolution meteorological data, the temperatures in mid to late May, early to late June, and the humidity in early to late April, early to mid May, and late June were found to be sensitive, which were used as input variables in constructing the forecasting model of apple ALS;

(3)

With the preprocessed disease survey data and sensitive meteorological data, four machine learning algorithms (i.e., Logistic regression, Support Vector Machine, Fisher Linear Discriminant Analysis, and K-Nearest Neighbors) were tested and compared for disease forecasting. Given that the KNN exhibited relatively high accuracy and strong robustness in model validation, it is thus recommended as appropriate modeling approach in forecasting of apple ALS in this study.