Rule-based expert system for detection of coffee rust warnings in colombian crops

Abstract

The coffee rust is a devastating disease that causes large economic losses across the world. The severity of this disease changes over time so the farmers are not fully aware of the economic importance of the rust disease in the coffee crops. From a computational science perspective, several investigations have been proposed to decrease the effects caused by the coffee rust appearance from Expert systems based on machine learning techniques. However, because samples about coffee rust incidence are few, the rules created from machine learning techniques do not contain enough information to consider the diversity of scenarios for detecting coffee rust. This paper proposes an expert system based on rules, where the rules are created considering the expert knowledge of specialists and technical reports about the behavior of the disease during a crop year. As far as we know, this is the first expert system proposed using not only expert knowledge but also technical reports in the coffee rust problem. The Buchanan methodology is used to design the proposed system. Experiment results present an average accuracy of 66,67% to detect a correct warning of coffee rust levels.

Keywords

Decision support system crops disease agriculture hemileia vastatrix

1. Introduction

Coffee rust is a devastating disease that causes large economic losses across the world. The causal agent is the fungus Hemileia vastatrix. The severity of this disease changes over time, so the farmers are not fully aware of the economic importance of the rust disease in the coffee crops. Especially, when the weather is appropriate and the fruit load is high, the coffee rust causes big losses [1]. In the period between 2008 and 2012, the weather conditions favored the coffee rust growth causing that an appropriate control of the crops to prevent the appearance of the disease was not possible, affecting the production in several countries, including Colombia (its production descended 31%), El Salvador (its production descended 54%) and many places from Central America [2]. As a consequence, several research works have been conducted to detect the factors that promote the growth of this disease [3]. The authors in [4] present a literature review for detecting and forecasting pests and diseases in several crops. Authors conclude that coffee is one of the most studied crops, in countries like Colombia, due to its economic importance.

Several modeling approaches applied to coffee rust have been tested to detect the disease in the time. Some algorithms of “black box” as Support Vector Regression (SVR) [5], neuronal networks (NN) [5, 6], Bayesian networks [7], Error Correction Output and SVM [6 –8] have been used to detect the disease due to this type of modeling approaches have a high accurate; the different research works have demonstrated that these algorithms have an small alert error probability for detecting coffee rust. Algorithms with graphic representation also have been used to detect coffee rust, since these approaches allow to know the characterizes of the models built and analyze it to take important decisions [5–7 , 10]. On the other hand, Corrales (2015) concludes the ensemble methods approach can improve the performance of coffee rust detection, given that it has been demonstrated that the combination of several classifiers can improve the results in a classification/regression process [5]. In the coffee context, researchers have used ensemble methods using neural networks, SVR and regression trees [11] and random forest [6].

Although the studies mentioned above address the coffee rust detection, the experiment results use a low number of instances to predict the coffee rust incidence (0% – 100%). When there are few samples, the data are not able to represent the study population, and then the classifiers are not accurate [5, 11]. In the coffee rust context, the number of samples always will be small, since the collection method of coffee rust infection is done monthly and on some days of the month [13]. To increase the volume of data (amount of coffee rust samples), which a limiting factor as indicated by Corrales 2015, two studies propose the use of machine learning techniques. Authors in [12] propose the cubic spline interpolation to increase the measurements of Incidence of Rust (IR). Authors in [14] propose a guideline for generating coffee rust samples by applying machine learning techniques.

Although these research works propose methods to increase the volume of coffee rust data, the researchers only consider the algorithm behaviors. However, to detect appropriately the coffee rust, it is necessary to aggregate information from the main factors that cause this disease.

Due the lack of coffee rust samples, an interesting alternative is to leverage the knowledge of the coffee farmers and agronomists to tackle the coffee rust disease. This paper proposes a rule-based expert system able to trigger a warning about the level of rust in the coffee crops, considering not only the weather conditions but also the production behavior in Colombian coffee zones. The paper is organized as follows: Section 2 describes the main concepts involve in the study and the related works with expert systems in several agricultural domains. Section 3 develops the expert system for detection of coffee rust warning. In Section 4 the results obtained in the evaluation of expert system with specialists in coffee crops are discussed. Finally, Section 5 shows conclusions and future works.

2. Materials and methods

This section introduces the definition about coffee rust and expert systems and present related works.

2.1. Background

2.1.1. Coffee rust

The coffee rust disease is caused by the fungus Hemileia vastatrix. Among the cultivated species, Coffea arabica is the most severely attacked. The disease begins with the appearance of yellowish spots underside the leaves. When the disease is acute, the coffee rust causes defoliation, the death of branches and heavy crop losses [15]. The expression and intensity of the disease are the result of interactions between [16] the crop characteristics, the pathogen, the environment, and crop management, that all can vary particularly weather.

Colombian researchers describe the progress of the disease through polynomial curves, each curve being different depending on the region of production [13]. In general, the progress of the coffee rust depends on four factors [17]:

– The host: coffee crops can be susceptible or resistant. The Castillo variety is highly resistant to rust. Whereas the Tipica, Borbón and Caturra are susceptible to rust. The fruit load is also a characteristic of the host that can change its physiological susceptibility: the higher the fruits load, the higher the susceptibility.

– The pathogen:Hemileia vastatrix life cycle has four main stages: dispersal of spores, germination of spore with leads to infection, colonization stage where the fungus intracellularly colonizes the leaves, and sporulation where new spores are produced. In this final stage the fungus begins the cycle again.

– Weather conditions: the spread of the disease depends strongly on the rainy season; in dry periods rust incidence is low because germination and infection are free water dependent. However, each stage of the rust life cycle has its own microclimate requirements. Colonization is particularly dependent on temperatures (optimum of 23°C) and not on humidity, because the fungus is inside the leaf at this stage. Dispersal is favored by wind and raindrops impacting and shaking leaves, as well as by workers when they circulate within the plantation.

– Agronomic practices: is referred to crop properties (plant spacing, the percentage of the shade, etc.), application of fungicides and fertilizers on coffee crops.

2.1.2. Expert system

The expert systems (ES) collect information from specialists into a specific area and generate solutions to a problem doing use of the knowledge collected. The systems based on rules are an appropriate tool to solve problems. The knowledge base contains the variables and the set of rules that define the problem; on the other hand, the inference engine obtains the conclusions applying the logic of the rules [18]. A rule is a logical proposition that explains the relationship between two or more variables and includes two parts: the premise and the conclusion. A rule is written usually like this: “if premise, then conclusion” [19]. An example of a rule shows as follows:

As example, considering the following expressions:

Zone = 1 → p.

Trimester = 4 → q.

Amplitude of temperature = small → r.

Month = wet → s.

Humidity = greater 80% → t.

Then, the formula of a rule into an expert system could be as follows: $p \land q \land r \land s \land t = medium level$ (1)

The expression above represents a rule into an expert system, when a user enters that:

The main coffee harvest in the region is in the first semester of the year (zone = 1).

During the fourth semester of the year data have been collected.

The amplitude of temperature is less than 12°C (small).

The month of the data collected is wet.

And the relative humidity is greater than 80%.

So, the expert system concludes that the level of coffee rust with those conditions is “medium”.

2.2. Related works

Several studies have used ES to detect or to diagnose pests and diseases in the agricultural context. A systematic mapping was made to find research works related with the agriculture expert systems. Two exclusion criterions were used to select the investigations found: “recent papers (less than 10 years)” and “literature reviews” The systematic mapping based its search on the following research question:

“¿What kinds of agriculture expert systems are used to detect pests and diseases?”

In the systematic mapping 14 papers were found. However, only 9 papers were selected. Five excluded papers were focused on literature reviews. The papers selected were found in the following sources of information: IEEEXplore (4), ScienceDirect (2) and SpringerLink (3). We defined the next search query: “(“expert system” OR “expert systems”) AND (“disease” OR “diseases”) AND (“agriculture” OR “coffee”)”. These results showed two types of expert systems: 1) systems focused on the diagnosis of pests and diseases, and 2) systems focused on the detection of pests and diseases. The papers found in the systematic mapping are described below:

2.2.1. Expert systems to diagnose and detect pests and diseases

The expert systems found propose different ways to create decision rules. Some authors use machine learning techniques to generate rules according to the behavior of their data [9 , 21], these research works use graph pattern, criteria decision, weighting techniques and regression and classification algorithms to extract rules and define guidelines. Other authors use image databases to save examples of different diseases and using computer vision approach and a set of rules they are able to build an efficient expert system [22, 23]. Finally, some research works consider the opinion of people with high knowledge in a specific topic to create the rules. A rule-based expert system for rice and wheat crop pest diagnosis and management is developed in [24]. The system presents a user graphical interface to show comparisons between the user and the system responses. In [25] an expert system based on a web application is presented; the objective of this system is to diagnose pests and diseases in rice plants. The knowledge in the system is created by using production rules and forward chaining to analyze the symptoms from attacked rice plant.

In Malaysia, a set of rules are created to develop an expert system for the diagnosis of oyster fungal disease. This ES assists farmers in the crop management decisions [26]. For controlling and diagnosing diseases in crops like legumes, an expert system called PulsExpert has been developed [27]. This system uses rules defined from the knowledge obtained by several experts on the topic. In China, researchers designed an Internet-Based Expert System to detect pests and diseases. This system is based on organizing the knowledge related to the agricultural pest species/category and their characteristics. The idea is to build a dynamic expert system where several experts of the diseases and pests in the agriculture update information about diseases and pests to improve the knowledge rules [28]. The studies focused in the diagnosis and detection of pests and diseases are summarized in the Table 1.

Table 1
Expert systems focused in to diagnose and detect pests and diseases

Work Year Kind of expert system used Type of crop

[9 , 21] 2015, 2017, 2011 Expert systems based on rules. The rules were built using machine learning techniques. Pomegranate, plants in general, chickpea, fruit plants, and coffee

[22, 23] 2010, 2009 Expert systems based on rules. The rules were built using an image databases. Corn and olive

[24 –28] 2014, 2017, 2012, 2011, 2010 Expert systems based on rules considering the expert knowledge of people. Rice, wheat, oyster and coffee

Work	Year	Kind of expert system used	Type of crop
[9 , 21]	2015, 2017, 2011	Expert systems based on rules. The rules were built using machine learning techniques.	Pomegranate, plants in general, chickpea, fruit plants, and coffee
[22, 23]	2010, 2009	Expert systems based on rules. The rules were built using an image databases.	Corn and olive
[24 –28]	2014, 2017, 2012, 2011, 2010	Expert systems based on rules considering the expert knowledge of people.	Rice, wheat, oyster and coffee

Several research works used expert systems based on rules to detect or diagnose pests and diseases. The rules in this kind of expert systems are created based on several types of information, such as datasets with disease samples, expert knowledge from specialists and technical reports, and from machine learning techniques used to build rules based on context information. This paper proposes an expert system to prevent the coffee rust appearance considering some key factors that explain the behavior of the disease during a crop year. Expert systems to diagnose crop pests and diseases are not good options to this purpose because the main objective of this kind of systems is identifying (not always with good results) pests and diseases in crops through rules based on information of the databases, that contain samples (images mainly) of different diseases and pests in the crops studied. Expert systems proposed to detect pests and diseases using rules created from machine learning techniques and validated by expert knowledge are more appropriate to our purpose as their objective is to prevent the appearance of the disease. However, because samples about coffee rust incidence are few [4], the rules created from machine learning techniques do not contain enough information to consider the diversity of scenarios for detecting coffee rust. For this reason, this paper proposes an expert system based on rules, where the rules are created considering the expert knowledge of specialists and technical reports about the behavior of the disease during a crop year. As far as we know, this is the first expert system proposed using not only expert knowledge but also technical reports in the coffee rustproblem.

3. Our proposal: Expert system for detection of coffee rust warning

This section presents the design of the expert system proposed. The Buchanan methodology [29] was used to build the ES for detection of coffee rust warning. This methodology is based on a hierarchical life cycle, which defines five modules or steps: identification, conceptualization, formalization, implementation, and validation or test. These steps are described as follows:

3.1. Identification

This step identifies the actors to build the system and their roles in the construction. In this stage, the resources and knowledge sources available are also identified. The study has identified the next aspects:

Problem: Since the small farmers are not able to make an appropriate diagnosis about the rust behavior in a certain period of time, they cannot make suitable decisions to prevent the rust in the crops.

Solution: Design an ES to detect the coffee rust using information about the weather conditions and the crop variables.

To understand the problem context, the following knowledge sources are used:

Technical reports from Centro Nacional de Investigaciones del Café (Cenicafé)

Several datasets from the central zone of Colombia with samples of the percentage of rust incidence

Knowledge from specialists in coffee crops and biological sciences

Once the resources available are identified, several interviews with experts in coffee diseases and experts in computational systems are carried out. Considering the previous factors, the most important tasks of the expert system are:

The system allows that the users enter the weather variables for which they want to predict the level of coffee rust in their crops

The system returns an early warning about the rust level in their crops

3.2. Conceptualization

The main task in this stage is to describe the expert knowledge obtained for the project. The weathervariables and the rust progress curves for the different coffee zones from Colombia are described.

In this case, we can obtain two different types of information: qualitative and quantitative:

Qualitative information

Target variable: level of rust (high, medium, low).

Progress curves: these curves represent the “normal” behavior of the coffee rust with no control in a year. In Colombia two progress curves can be differentiated: one of them represents the main harvest in the first semester of the year (Fig. 1a), and another curve represents the main harvest in the second semester of the year (Fig. 1b) [1]. The levels of coffee rust incidence usually increase during the flowering period and the highest levels appear in the main harvest period.

Fig.1

Progress curves of coffee rust without control.

Quantitative information

Weather variables (per day): average minimum temperature, average maximum temperature, relative humidity and rainfall.

Temporal variables: date for each sample analyzed.

Farm variables: altitude of the farm, main harvest date.

3.3. Formalization

In this step, the knowledge base of the ES is created: first, setting the rules and then, defining the engine knowledge.

Description of variables

Study areas: considering the progress curves of the coffee rust in Colombia, it is necessary that the expert system can identify the zone where the data come from. In this paper two areas are defined (Table 2).

Table 2
Zones identifies for the expert system

Region Description

Zone 1 Main harvest in the first semester of the year in Quindío, Norte de Santander and Cauca.

Zone 2 Main harvest in the second semester of the year in Antioquia, Caldas and Risaralda.

Region	Description
Zone 1	Main harvest in the first semester of the year in Quindío, Norte de Santander and Cauca.
Zone 2	Main harvest in the second semester of the year in Antioquia, Caldas and Risaralda.

Weather variables: based on several information sources and taking into account the opinion of experts in the coffee diseases, two kinds of variables can be identified: first, the basic variables which are inserted by the user (Table 3); second the derived processed variables, which are created by the system as drivers of rust development (Table 4).

Table 3

Basic variables that the user enter in the expert system

Variable	Description
Zone	This variable identifies the region of the data according to its main harvest.
Zone 1: Quindío, Norte de Santander and Cauca.
Zone 2: Antioquia, Caldas and Risaralda.
Maximum temperature	Maximum temperature (°C).
Minimum temperature	Minimum temperature (°C).
Rainfall	Level of precipitation accumulates (milimeters).
Relative humidity	Percentage of relative humidity.
Month	The month of the year where the sample was taken.
Altitude	Height above sea level of the study area.

Table 4

Processed variables that belong the rules of the expert system

Variable	Description
Zone	Integer number that describes the region of the data according to its main harvest. According to the zone, the system knows which behavior analyzing considering the progress curves.
	1: identifies the zone 1.
	2: identifies the zone 2.
Amplitude of temperature	Numeric value calculated according to the following equation:
	ΔT = maximum temperature – minimum temperature
	According to experts, it is possible to divide the amplitude of temperature as a small difference or large difference:
	ΔT < = 12°C: small.
	ΔT >12°C: large.
Month	Value that specifies if the previous 28 days were wet, dry or very wet.
Month = (dry, wet, very wet).
	A month is dry if the evapotranspiration (eto) value is less than level of rainfall {dry: eto < rainfall}
	A month is wet if eto value is greater than rainfall level and it is less than or equal to 2*eto {wet: rainfall < = eto < =2eto}
	A month is very wet if rainfall value is greater than 2eto {very wet: rainfall> = 2eto}Eto(per day) = 4,37 exp(– 0,0002altitude) [30, 31].
Relative humidity	In constant periods of time the relative humidity can be divided in two categories according to experts:
	Greater or equal to 80%.
	Less than or equal to 80%.
Trimester	Value that represent the trimester where a variable come from.
	January, February and March = 1.
	April, May and June = 2.
	July, August and September = 3.
	October, November and December = 4.
Target variable	Level of rust estimated by the system.
	Output = (low level, medium level, high level).

Rules: we propose to create a series of rules that allow identifying the coffee rust level (low, medium, high) based on the values of the processed variables. To establish the rules it is necessary first of all to know the crop area since in each region the disease has a different behavior during the year (zone 1 or zone 2). Then, the trimester where the data come from has to be defined, because the expected levels of the disease are depends on the zone; therefore, the interpretation of the weather conditions must be analyzed in different ways in each trimester into a specific region.

In the other hand, analyzing the weather variables independently, it can be said that:

The amplitude of temperature (ΔT) less than 12°C the temperature is never too high or too low for rust (particularly colonization which is temperature dependent). A short amplitude with means very low or very high is not favorable.

In general, the dry and very wet months do not favor the development of the disease. In a dry month there is no germination, so no infection; and a very wet month usually the rainy wash the leaves, so the spores of the fungus are reduced.

The average monthly of the humidity greater than 80% favor the development of the disease.

However, these considerations can be change taking into account the joint analysis of the weather variables. For example, a value of ΔT small (<12°C) and a month very wet do not favor the development of the disease because constant rains clean the leaves of the crops, but a value of ΔT large can maintain high constant temperatures and eliminate the excess of water in the leaves, providing a level of humidity appropriate to development of the disease into a month very wet.

Considering the all above, it is created a rule of the expert system in the following order:

To identify the zone (crop area)

To identify into the zone which trimester the data come from

To analyze the weather conditions and concluding a coffee rust level

Taking account the next expressions:

Zone [1] → Z, Zone [2]→Z’

Trimester [1]→D, Trimester [2]→D’, Trimester [3] →D”’, Trimester [4] →D””

Amplitude of temperature[small] →T, Amplitude of temperature[large] →T’

Month[wet] →M, Month[very wet] →M’, Month[dry] →M”

Humidity[greater80%]→H, Humidity[smaller 80%]→H’

Then, the formulas of a rule into the expert system are: $Z^{D^{‴}}^{'}^T^{'}^M^{'}^H = M e d i u m l e v e l$ (2)

$Z^D^T^M^{″}^H = H i g h l e v e l$ (3) $Z^{'}^D^{'}^T^M^{″}^H = L o w l e v e l$ (4)

The expert system contains 96 rules divided as follows in the Table 5.

Table 5

Categories of the rules in the expert system

Zone	Trimester	Weather condition (number of options)	Number of rules
Zone 1	I	ΔT (2), month (3), humidity (2)	12
	II	ΔT (2), month (3), humidity (2)	12
	III	ΔT (2), month (3), humidity (2)	12
	IV	ΔT (2), month (3), humidity (2)	12
Zone 2	I	ΔT (2), month (3), humidity (2)	12
	II	ΔT (2), month (3), humidity (2)	12
	III	ΔT (2), month (3), humidity (2)	12
	IV	ΔT (2), month (3), humidity (2)	12

3.4. Implementation

This step converts the knowledge into formal software documentation which defines the software tools used in the expert system, the user interface, and the system architecture.

Uses cases: a use case is a diagram that defines the interactions between a role (human or external system) and a system to achieve a goal. The expert system has a unique use case called “get rust level” (Table 6).

Fig.2

Expert system architecture.

Table 6

Use case “get rust level”

Use case	Get rust level
Code	UC01
Objective	The user introduces the basic variables and after that, an answer is returned with the rust level.
Pre-conditions	The user must enter all the basic variables required.
Post-conditions	It does not apply.
Actors	User.
Principal flow	1. The user enters the basic variables in the UI.
	2. The user pushes the execution button.
	3. The system calculates the processed variables and enters the responses to the engine knowledge.
	4. The system applies the rules and returns an answer to the user with a message about the level of rust.
Extensions	At less a variable did not be entered. The system does a request to enter all basic variables.

Considering the interviews with the experts, the interface of the system has two options for entering data; one of them is a manual entry, where the user has specific values to analyze the basic variables. The second option provides the possibility of entering a dataset, in this second option the ES analyzes the last 28 days and calculates the averages of the numeric basic variables, then, the processed variablesare built.

Development tools: The rules were created in Python version 3.6 with pyknow library [32]. The User Graphical Interface was created with pyqt library [33] and the QtDesigner environment [34]; other libraries as pandas and numpy were used to data preprocessing and mathematical calculations. The modules described and the development environment spyder are integrated through Anaconda3 [35].

System Architecture: The architecture of the system is composed by the modules described below (Fig. 2).

User Interface: users inform the basic variables in the interface, and these are sent to the processing module. The variables can be entered by a dataset or manually. The user interface is built using the pyqt library which provides a constant synchronization between the elements into the interface and the model of the system. When the user modifies a value in the UI, the attributes involved change at the same time. Another tool used in this module is QtDesigner environment, an IDE multi-platform to create Graphic User Interfaces (GUI).

Processing basic variables module: this module carries out the necessary mathematical calculations to get the processed variables from basic variables. In this module, the pandas library (an extension of numpy library) gives an appropriate format to data first, and then the numpy library which uses vectors and matrixes to do efficient calculations allows creating better internal methods to generate the processed variables.

Pyknow module: this module uses the Pyknow library to manage the tasks of the ES. The Pyknow library has three components in its architecture: the working memory, the set of rules and the interpreter. The working memory contains the information to manage the system by means of facts. The facts are the basic unit of information in Pyknow library since this information is used by the system to reason about the problem. The set of rules has the knowledge of the system (the rules). Rules are the callable into the system and they have two components: LHS (left-hand-side) which describes the conditions where a specific rule should be executed, and the RHS (right-hand-side) component which contains a set of actions to carry out when the rule is fired. Finally, the interpreter manages the execution of the tasks into the expert system considering the information in the working memory and the set of rules.

Levels module: once a conclusion is triggered, an answer is shown to the user.

3.5. Test

The expert system rules were evaluated with experts in the coffee context. The experts that interacted with the expert system are a PhD in Biology (subject 1) and the operational manager in Supracafé Colombia (subject 2). The protocol of the test is described as follow:

The premise

The experts made the test based on the following premise:

“Each value added by the expert in the system is the average value of the 28 days before”

The input variables

Each expert must analyze the following input variables:

ID: identification of the test (10 different cases are evaluated with each expert)

MAXT: average maximum temperature in °C

MINT: average minimum temperature in °C

Pl: rainfall in millimeters

H: relative humidity in percentage

Zone: zone 1 or 2, it depends on the main harvest period

Month: month in which the samples were taken

Altitude: height above sea level in meters

The answer provided by the expert

Response: the expert proposes low level, medium level or high level of coffee rust considering the previous input variables.

The system response

The experts add the different values of the input variables and run the expert system.

The expert system returns an answer of low level, medium level or high level according to the application of its rules.

Discussion

Each expert compares his answers with the answers proposed by the expert system, and then they make an analysis to get feedback about the system. The results are described in the next section.

4. Results

Once the expert system was built in a desktopapplication, the experts used the system to evaluate the system responses. The subjects used the system taking into account the protocol described in Section 3.5. The measures used to evaluate the system responses are:

Accuracy: precision of the system to predict a correct value.

$Accurancy = \frac{TP + TN}{TP + TN + FP + FN}$ (5)

Precision: measure to determinate the proportion of positive values that are true positives.

$Precision = \frac{TP}{TP + FP}$ (6)

Recall: measure to determinate the proportion of positives that are identified correctly. $Recall = \frac{TP}{TP + FP}$ (7)

Where:

True Positives (TP): the expert predicted a positive answer (high, medium or low) and the answer of the ES predicted a positive answer.

True Negatives (TN): when the ES predicted NO (not low, not medium, not high), and the expert agrees.

False Positives (FP): the ES predicted YES (low, medium, high), but the expert disagrees.

False Negatives (FN): the ES predicted NO (not low, not medium, not high), but the expertdisagrees.

The system responses compared with the prediction of each expert allowed generating a confusion matrix for each expert. Expert system easily identified the true negatives responses with subject 1 (Fig. 3). In other words, the system identified better the responses that do not belong to any of the coffee rust levels. This mainly happens because it is difficult for the system to find a difference between a high and medium level of coffee rust.

Figure 4 shows the results for subject 2. In this case, high and medium levels of coffee rust have the same behavior. For this reason, the false positives and false negatives measures are almost equal to true positives and true negatives values. On the other hand, this expert did not find any case with a low level of coffee rust. The Table 7 shows a summary of the expert system results with each expert.

Fig.3

Expert system results (%) of the levels of coffee rust warning with the subject 1.

Fig.4

Expert system results (%) of the levels of coffee rust warning with the subject 1.

Table 7

Summary of expert system results

Expert	Coffee rust level	Precision	Recall	Accuracy
Subject 1	High	50%	66.67%	73.3%
	Medium	60%	60%
	Low	100%	50%
Subject 2	High	50%	40%	60%
	Medium	40%	40%
	Low	0%	0%

According to our results, the system has a good accuracy; the system predicts a correct value of the coffee rust levels 66,7% of the times. However, the precision is affected by the ability of the system to distinguish between a high level and a medium level of coffee rust obtaining an average value of 50%.

5. Conclusions

This research presents an expert system for detecting a warning of coffee rust levels in crops. The use of methodologies to build the expert system allowed designing a system that considers weather properties and expert knowledge about the progress of the disease in different places of Colombia to create 96 rules and trigger preventive warnings about coffee rust levels. The expert system has an accuracy to trigger a correct warning of 66,7%. The system needs to be expanded and their rules updated to consider new scenes and improve its accuracy. It is needed to consider:

Specific zone: besides to consider if the main harvest is in the first or second semester of the year, it is necessary to take into account the month in which the harvest is the highest. With that information it is possible to calculate the period where crops are more susceptible to the rust disease.

Temperature interpretation: the amplitude of temperature has a different meaning in different crops areas, and with different maximum and minimum temperature values, for example, a ΔT = 10°C could be the result of MINT = 10°C and MAXT = 20°C, but also of MINT = 20°C and MAXT = 30°C. Both examples result in equal amplitude of temperature, but the effect of each case is different into the rules. To improve the interpretation of this variable, it is necessary to assign weights to the rules depending on the behavior of the amplitude of temperature considering their minimum and maximum temperature.

The Colombian researchers use machine learning algorithms to build models able to detect the coffee rust incidence rate or the growth rate of the disease into some experimental farms. These models have been created using datasets with information about weather variables of the crop zone and some samples of coffee rust; some works have information about crop properties, these studies are limited by the lack of data on coffee rust. This paper proposed an expert system that can trigger a warning about the coffee rust level in the crops taking account information about weather variables and crop zone properties. The proposed rules can work even when data are not available. However, it is necessary to increase the analysis to build new rules to improve the system results considering the specific zones of coffee crops and dates of the coffee crops, besides the temperature interpretation must be analyzed deeply.

As future works we propose first, to update the rules taking account information about the flowering period dates and the month where the harvest is most high, second, to use time series for weather variables to deduce coffee rust behavior at short term (few days), to work with ontologies in order to represent the coffee rust knowledge and this way to do semantics inferences, and finally build an expert system considering the case based reasoning.

Footnotes

PhD program coordinator in Ambiental sciences of the University of Cauca. Colciencias, “CvLAC Apolinar Figueroa,” 2017. [Online]. Available: .

The operational manager of Supracafé Colombia. LinkedIn, “Professional profile Wilton Benitez,” 2017. [Online]. Available:

Acknowledgments

We thank to the Telematics Engineering Group (GIT) of the University of Cauca for the technical support, the Tropical Agricultural Research and Higher Education Center (CATIE). Besides, we are grateful with the experts Apolinar Figueroa 1 and Wilton Benitez" href="#fn2" id="a-84"> 2 to assess the Expert System. Finally, this work has been partially supported by Colciencias through PhD scholarship granted to MsC. David Camilo Corrales and Cauca Innovacción Nucleus Group for the Young investigator program to Ing. Edwar Javier Girón. This work has also been supported by:

Project: “Alternativas Innovadoras de Agricultura Inteligente para sistemas productivos agrícolas del departamento del Cauca soportado en entornos de IoT - ID 4633” financed by Convocatoria 04C–2018 “Banco de Proyectos Conjuntos UEES-Sostenibilidad” of Project “Red de formación de talento humano para la innovación social y productiva en el Departamento del Cauca InnovAcción Cauca”.

Spanish Ministry of Economy, Industry and Competitiveness (Projects TRA2015-63708-R and TRA2016-78886-C3-1-R).

References

G.A.

Alvarado , El café y la roya, Caldas, Cenicafé, 2011.

Avelino ,

Cristancho ,

Georgiou ,

Imbach ,

Aguilar ,

Bornemann ,

Läderach , F. Anzueto,

A.J.

Hruska and

Morales , The coffee rust crises in Colombia and Central America (2008-2013): Impacts, plausible causes and proposed solutions, Food Secur7(2) (2015), 303–321.

I.D.

López ,

Campo ,

S.A.

Ordoñez ,

J.C.

Corrales ,

Figueroa ,

Leon and

D.C.

Corrales , Plataforma para el seguimiento de variables meteorológicas y ambientales parael sector agropecuario, VII Congr Ibérico AgroIngeniería y Ciencias Hortícolas, 2013, p. ISBN - 10:84-695-8844-3.

D.C.

Corrales , Toward detecting crop diseases and pest by supervised learning, Ing y Univ19(1) (2015), 207.

D.C.

Corrales ,

Ledezma ,

A.J.Q.

Peña ,

Hoyos ,

Figueroa and

J.C.

Corrales , A new dataset for coffee rust detection in Colombian crops base on classifiers, Sist y Telematica12(29) (2014), 9.

C.A.A.

Di Girolamo ,

Rodrigues and

L.H.A.

Meira , Modelos de predição da ferrugem do cafeeiro (Hemileia vastatrix Berkeley & Broome) por técnicas de minerçãao de dados. - Portal Embrapa. [Online]. Available: https://www.embrapa.br/informatica-agropecuaria/busca-de-publicacoes/-/publicacao/991078/modelos-de-predicao-da-ferrugem-do-cafeeiro-hemileia-vastatrix-berkeley-broome-por-tecnicas-de-mineracao-de-dados.

C.B.

Pérez-Ariza ,

A.E.

Nicholson and

M.J.

Flores , Prediction of Coffee Rust Disease Using Bayesian Networks, Proc 6th Eur Work Probabilistic Graph Model PGM, 2012.

D.C.

Corrales ,

A.J.Q.

Peña ,

León ,

Figueroa and

J.C.

Corrales , Early warning system for coffee rust disease based on error correcting output codes: A proposal, Rev Ing Univ Medellin13(25) (2014), 57–64.

Lasso ,

T.T.

Thamada ,

C.A.A.

Meira and

J.C.

Corrales , Graph patterns as representation of rules extracted from decision trees for coffee rust detection, Commun Comput Inf Sci (2015), 405–414.

10.

M.E.

Cintra ,

C.A.A.

Meira ,

M.C.

Monard ,

H.A.

Camargo and

L.H.A.

Rodrigues , The use of fuzzy decision trees for coffee rust warning in Brazilian crops, in 2011 11th International Conference on Intelligent Systems Design and Applications, 2011, pp. 1347–1352.

11.

D.C.

Corrales ,

Figueroa ,

Ledezma and

J.C.

Corrales , An Empirical Multi-classifier for Coffee Rust Detection in Colombian Crops, Springer International Publishing, 2015, pp. 60–74.

12.

D.C.

Corrales ,

German ,

J.P.

Rodriguez ,

Agapito and

J.C.

Corrales , Lack of data: Is it enough estimating the coffee rust with meteorological time series?Comput Sci Its Appl -ICCSA 201710405 (2017), 3–16.

13.

Rivillas Osorio ,

Serna Giraldo ,

Cristancho Ardila and

Gaitáin Bustamante , La roya del cafeto en Colombia, Cenicafé, 2011.

14.

J.P.

Rodríguez ,

E.J.

Girón ,

D.C.

Corrales and

J.C.

Corrales , A Guideline for Building Large Coffee Rust Samples Applying Machine Learning Methods, Springer, Cham, 2018, pp. 97–110.

15.

Lasso ,

Valencia and

D.C.

Corrales , A Cloud-Based Platform for Decision Making Support in Colombian Agriculture: A Study Case in Coffee Rust, in Advances in information and Communication Technologies for Adapting Agriculture to Climate Change AACC17 2017, Springer, Cham, 2017, pp. 182–196.

16.

T.A.

Hernandez , Epidemiologia, fundamentos de analisis y manejo sisteimico de epidemias de cultivos tropicales. FEDAM, 2006.

17.

Rivillas Osorio ,

Serna Giraldo ,

Cristancho Ardila and

Gaitan Bustamante , La Roya del cafeto en Colombia, 2009.

18.

J.M.

Gutierrez , Sistemas Expertos Basados enReglas, 2008. [Online]. Available: http://personales.unican.es/gutierjm/cursos/expertos/Reglas.pdf.

19.

Jackson , Introduction to expert systems, 1998. [Online]. Available: https://es.wikipedia.org/wiki/Sistema_experto. [Accessed: 19-Dec-2017].

20.

Goodridge ,

Bernard ,

Jordan and

Rampersad , Intelligent diagnosis of diseases in plants using a hybrid Multi-Criteria decision making technique, Comput Electron Agric133 (2017), 80–87.

21.

Luaces ,

L.H.

Rodrigues ,

C.A.

Alves Meira and

Baha-monde , Using nondeterministic learners to alert on coffee rust disease, Expert Syst Appl38(11) (2011), 14276–14283.

22.

Lai ,

Ming ,

Li ,

Wang ,

Xie and

Gao , An image-based diagnostic expert system for corn diseases, Agric Sci China9(8) (2010), 1221–1229.

23.

J.L.

Gonzalez-Andujar , Expert system for pests, diseases and weeds identification in olive crops, Expert Syst Appl36(2) (2009), 3278–3283.

24.

Balleda ,

Satyanvesh ,

N.V.S.S.P.

Sampath ,

K.T.N.

Varma and

P.K.

Baruah , Agpest: An efficient rule-based expert system to prevent pest diseases of rice wheat crops, in 2014 IEEE 8th International Conference on Intelligent Systems and Control (ISCO), 2014, pp. 262–268.

25.

Agustina ,

Pratomo ,

A.D.

Wibawa and

Rahayu , Expert system for diagnosis pests and diseases of the rice plant using forward chaining and certainty factor method, in 2017 International Seminar on Intelligent Technology and Its Applications (ISITIA), 2017, pp. 266–270.

26.

Munirah ,

Rozlini and

Siti Mariam , Design and rules development of Expert System for Diagnosing Oyster Mushroom Diseases, in 2012 International Conference on Computer & Information Science (ICCIS), 2012, pp. 286–289.

27.

Devraj and

Jain , PulsExpert: An expert system for the diagnosis and control of diseases in pulse crops, Expert Syst Appl38(9) (2011), 11463–11471.

28.

Q.L.Q.

Li ,

Z.S.Z.

Shen ,

L.G.L.

Gao ,

H.Y.H.

Yang ,

W.L.W.

Liu and

J.Q.J.

Qiao , An Internet-Based expert system platform for assistant identification of agricultural pests, World Autom Congr(WAC), 2010, 2010, pp. 141–144.

29.

B.G.

Buchanan ,

Barstow ,

Bechtal ,

Bennet ,

Clancey ,

Kulikowsky ,

T and

D.A.

Waterman , Constructing an expert system, Addison We. Building Expert Systems, 1983.

30.

A.J.

Robledo and

Jaramillo , EvapotranspiraciÓn de referenda en la region andina de colombia, 57(4) (2006), 288–298.

31.

Á.

Jaramillo ,

V.H.

Ramírez and

Arcila , Patrones de distribucion de la lluvia en la zona cafetera, Av técnicos Cenicafé410 (2011), 12.

32.

R.A.

Martinez and

D.F.

Cuartero , pyknow Documentation, 2017. [Online]. Available: http://pyknow.readthedocs.io/en/stable/index.html.

33.

PyQt5 Reference Guide — PyQt 5.9.2 Reference Guide. [Online]. Available: http://pyqt.sourceforge.net/Docs/PyQt5/. [Accessed: 10-Dec-2017].

34.

Using Qt Designer — PyQt 5.9.2 Reference Guide. [Online]. Available: http://pyqt.sourceforge.net/Docs/PyQt5/designer.html. [Accessed: 10-Dec-2017].

35.

User guide — Conda documentation, 2017. [Online]. Available: https://conda.io/docs/user-guide/. [Accessed: 10-Dec-2017].

Rule-based expert system for detection of coffee rust warnings in colombian crops

Abstract

Keywords

1. Introduction

2. Materials and methods

2.1. Background

2.1.1. Coffee rust

2.1.2. Expert system

2.2.1. Expert systems to diagnose and detect pests and diseases

3.1. Identification

3.2. Conceptualization

Table 2 Zones identifies for the expert system Region Description Zone 1 Main harvest in the first semester of the year in Quindío, Norte de Santander and Cauca. Zone 2 Main harvest in the second semester of the year in Antioquia, Caldas and Risaralda.

4. Results

Footnotes

Acknowledgments

References

Table 2
Zones identifies for the expert system

Region Description

Zone 1 Main harvest in the first semester of the year in Quindío, Norte de Santander and Cauca.

Zone 2 Main harvest in the second semester of the year in Antioquia, Caldas and Risaralda.