Carlos Alberto Alves Meira

Análise da epidemia da ferrugem do cafeeiro com árvore de decisão

Uma arvore de decisao foi desenvolvida com o objetivo de auxiliar na compreensao de manifestacoes epidemicas da ferrugem do cafeeiro causada por Hemileia vastatrix. Taxas de infeccao calculadas a partir de avaliacoes mensais de incidencia da ferrugem foram agrupadas em tres classes: reducao ou estagnacao - TX1; crescimento moderado (ate 5p.p.) - TX2; e crescimento acelerado (acima de 5p.p.)- TX3. Dados meteorologicos, carga pendente de frutos do cafeeiro (Coffea arabica) e espacamento entre plantas foram usados como variaveis explicativas das classes de taxa de infeccao. A arvore de decisao foi treinada com 364 exemplos preparados a partir de dados coletados em lavouras de cafe em producao, de outubro de 1998 a outubro de 2006. Ela classificou corretamente 78% do conjunto de treinamento e a sua acuracia foi estimada em 73% para a classificacao de novos exemplos. O acerto do modelo foi de 88%, 57% e 79% dos exemplos, respectivamente, para as classes de taxa de infeccao TX1, TX2 e TX3. As variaveis explicativas mais importantes foram a temperatura media nos periodos de molhamento foliar, a carga pendente de frutos, a media das temperaturas maximas diarias no periodo de incubacao e a umidade relativa do ar. A arvore de decisao demonstrou seu potencial como modelo de representacao simbolica e interpretavel, permitindo a identificacao das fronteiras de decisao existentes nos dados e da logica contida neles, auxiliando na compreensao de quais variaveis e como as interacoes dessas variaveis conduziram as epidemias da ferrugem do cafeeiro no campo.

The use of fuzzy decision trees for coffee rust warning in Brazilian crops

This paper proposes the use of fuzzy decision trees for coffee rust warning, the most economically important coffee disease in the world. The models were induced using field data collected during 8 years. Using different subsets of attributes from the original data, three distinct datasets were constructed. The class attribute, representing the monthly infection rate, was used to construct six datasets according to two distinct infection rates. Induced models can be used to trigger alerts when estimated monthly disease infection rates reach one of the two thresholds. The first threshold allows applying preventive actions, whereas the second one requires a curative action. The fuzzy decision tree models were compared to the ones induced by a classic decision tree algorithm, taking into account the accuracy and the syntactic complexity of the models, as well as its quality according to an expert opinion. The fuzzy models showed better accuracy power and interpretability.

Modelos de alerta para o controle da ferrugem-do-cafeeiro em lavouras com alta carga pendente

The objective of this work was to develop decision trees as warning models of coffee (Coffea arabica L.) rust in growing areas with large fruit load. Monthly data of disease incidence in the field collected during eight years were transformed into binary values considering limits of 5 and 10 percentage points in the infection rate. Models were generated from meteorological data and space between plants for each binary infection rate. The warning is indicated when the infection rate is expected to reach or exceed the respective limit in a month. The accuracy obtained by cross-validating the model to the limit of 5 percentage points was 81%, reaching up to 89% according to an optimistic estimate. This model showed good results for other important evaluation measures, such as sensitivity (80%), specificity (83%), positive reliability (79%), and negative reliability (84%). The model for the limit of 10 percentage points had a 79% accuracy and did not show the same balance among the other evaluation measures. Together, these two models may support the decisions about coffee rust control in the field. The decision tree induction is a viable alternative to conventional modeling techniques, thus facilitating the comprehension of the models.

Using nondeterministic learners to alert on coffee rust disease

Motivated by an agriculture case study, we discuss how to learn functions able to predict whether the value of a continuous target variable will be greater than a given threshold. In the application studied, the aim was to alert on high incidences of coffee rust, the main coffee crop disease in the world. The objective is to use chemical prevention of the disease only when necessary in order to obtain healthier quality products and reductions in costs and environmental impact. In this context, the costs of misclassifications are not symmetrical: false negative predictions may lead to the loss of coffee crops. The baseline approach for this problem is to learn a regressor from the variables that records the factors affecting the appearance and growth of the disease. However, the number of errors is too high to obtain a reliable alarm system. The approaches explored here try to learn hypotheses whose predictions are allowed to return intervals rather than single points. Thus, in addition to alarms and non-alarms, these predictors identify situations with uncertain classification, which we call warnings. We present three different implementations: one based on regression, and two more based on classifiers. These methods are compared using a framework where the costs of false negatives are higher than that of false positives, and both are higher than the cost of warning predictions.

Graph Patterns as Representation of Rules Extracted from Decision Trees for Coffee Rust Detection

Diseases in Agricultural Production Systems represent one of the biggest drivers of losses and poor quality products. In the case of coffee production, experts in this area believe that weather conditions, along with physical properties of the crop are the main variables that determine the development of a disease known as Coffee Rust. On the other hand, several Artificial Intelligence techniques allow the analysis of agricultural environment variables in order to obtain their relationship with specific problems, such as diseases in crops. In this paper an extraction of rules to detect rust in coffee from induction of decision trees and expert knowledge is addressed. Finally, a graph-based representation of these rules is submitted, in order to obtain a model with greater expressiveness and interpretability.

Viability of an alarm predictor for coffee rust disease using interval regression

We present a method to formulate predictions regarding continuous variables using regressors able to predict intervals rather than single points. They can be learned explicitly using the so-called insensitive zone of regression Support Vector Machines (SVM). The motivation for this research is the study of a real case; we discuss the feasibility of an alarm system for coffee rust, the main coffee crop disease in the world. The objective is to predict whether the percentage of infected coffee leaves (the incidence of the disease) will be above a given threshold. The requirements of such a system include avoiding false negatives, seeing as these would lead to not preventing the disease. The aim of reliable predictions, on the other hand, is to use chemical prevention of the disease only when necessary in order to obtain healthier products and reductions in costs and environmental impact. Although the breadth of the predicted intervals improves the reliability of predictions, it also increases the number of uncertain situations, i.e. those whose predictions include incidences both below and above the threshold. These cases would require deeper analysis. Our conclusion is that it is possible to reach a trade-off that makes the implementation of an alarm system for coffee rust disease feasible.

Weather-based coffee leaf rust apparent infection rate modeling

Brazil is the major coffee producer in the world, with 2 million hectares cropped, with 75% of this area with Coffea arabica and 25% with Coffea canephora. Coffee leaf rust (CLR) is one of the main diseases that cause yield losses by reducing healthy leaf area. As CLR is highly influenced by weather conditions, this study aimed to determine the best linearization model to estimate the CLR apparent infection rate, to correlate CLR infection rates with weather variables, and to develop and assess the performance of weather-based infection rate models to be used as a disease warning system. The CLR epidemic was analyzed for 88 site-seasons, while progress curves were assessed by linear, monomolecular, logistic, Gompertz, and exponential linearization models for apparent infection rate determination. Correlations between CLR infection rates and weather variables were conducted at different periods. From these correlations, multiple linear regressions were developed to estimate CLR infection rates, using the most weather-correlated variables. The Gompertz growth model had the best fit with CLR progress curves. Minimum temperature and relative humidity were the weather variables most correlated to infection rate and, therefore, chosen to compose a CLR forecast system. Among the models developed, the one for the condition of high coffee yield at a narrow row spacing was the best, with only 9.4% of false negative occurrences for all the months assessed.

Expert system for coffee rust detection based on supervised learning and graph pattern matching

Diseases in agricultural production systems represent one of the main reasons of losses and poor-quality products. For coffee production, experts in this area suggest that weather conditions and crop physical properties are the main variables that determine the development of coffee rust. This paper proposes an extraction of rules to detect coffee rust from induction of decision trees and expert knowledge. In order to obtain a model with greater expressiveness and interpretability, a graph-based representation is proposed. Finally, the extracted rules are evaluated using an expert system supported on graph pattern matching.