Soft computational techniques to identify cotton leaf damage

Abstract

The principal objective of agriculture is the production of a high yield of healthy crops. This yield may be improved by the automatic detection of diseases and the consequent reduction in the use of pesticides. A digital processing system for images was thus developed and used to identify lesions on the leaves of cotton plants. A collection of 60,659 images of sub-metric resolution showing samples of soil and both healthy and damaged leaves was obtained and processed with an algorithm for the extraction of texture from 102x102-pixel samples. Then they analyzed with a neuro-fuzzy classifier trained to discriminate the three types of regions (soil, healthy leaf, and lesioned leaf). The algorithm developed was able to recognize the three classes. It generated a great amount of information on recognition of background which was more consistent than leaf damage areas. Therefore, it surpassed the performance of areas of healthy leaves. A similar trend was found for sensitivity. The overall accuracy of the system was 71.2%, suggesting that the unbalanced data of the different classes had skewed the results of the algorithm, as the number of false positives for the less well represented classes was greater. The analysis of unbalance (F-Score) showed that, independent of the volume of data, the attributes of texture utilized yielded better results for the images containing areas of damage in relation to overall accuracy. Therefore, given the challenges involved in the automatic identification of lesions in agricultural crops, such as variations in illumination, color, and texture, as well as obstruction, overlapping, and complexity of the region of which the image was taken, the behavior of the model was deemed satisfactory. Given the hybrid nature of the model, it should contribute to the state of the art in the use of intelligent systems in agriculture. This algorithm is available at https://github.com/rafaeufg/Cotton-diseases