Detection of coffee berry necrosis by digital image processing of landsat 8 oli satellite imagery

Abstract

Coffee berry necrosis is a fungal disease that, at a high level, significantly affects coffee productivity. With the advent of surface mapping satellites, it was possible to obtain information about the spectral signature of the crop on a time scale pertinent to the monitoring and detection of plant phenological changes. The objective of this paper was to define the best machine learning algorithm that is able to classify the incidence CBN as a function of Landsat 8 OLI images in different atmospheric correction methods. Landsat 8 OLI images were acquired at the dates closest to sampling anthracnose field data at three times corresponding to grain filling period and were submitted to atmospheric corrections by DOS, ATCOR, and 6SV methods. The images classified by the algorithms of machine learning, Random Forest, Multilayer Perceptron and Naive Bayes were tested 30 times in random sampling. Given the overall accuracy of each test, the algorithms were evaluated using the Friedman and Nemenyi tests to identify the statistical difference in the treatments. The obtained results indicated that the overall accuracy and the balanced accuracy index were on an average around 0.55 and 0.45, respectively, for the Naive Bayes and Multilayer Perceptron algorithms in the ATCOR atmospheric correction. According to the Friedman and Nemenyi tests, both algorithms were defined as the best classifiers. These results demonstrate that Landsat 8 OLI images were able to identify an incidence of the coffee berry necrosis by means of machine learning techniques, a fact that cannot be observed by the Pearson correlation.