Machine learning applied to the identification of nutritional deficiencies in banana trees

Abstract

The nutritional evaluation of plants is done through chemical analysis or visual diagnosis and it is necessary to know the characteristic patterns of nutritional deficiency of each element. Digital image processing (PDI) is an example of the use of technology in agriculture. In this work the PDI is applied to identify the symptomatology of deficiency in digital images of banana leaves, induced to the deficiency of N, P and K nutrients. The experiment was carried out in two stages: 1) in a greenhouse, with seedlings submitted to a randomized complete block design in a 5x5 factorial scheme, with three replications. The factors were nutritional variation (complete solution, individual omissions of N, P, K and a control with soil cultivation), and sampling time at 0, 30, 60, 90 and 120 days after application of the treatments; 2) PDI experiment, applied in four phases: sample collection and digitization, extraction of histograms, selection of attributes and classification, performed with a database for each time studied (0, 30, 60, 90 and 120 days). The highest accuracy rates of the experiment were presented by classifiers with artificial neural networks (ANNs), equal to 66.7%, 62%, 76.7%, 62.3%, 68.3%, in the 0, 30, 60, 90 and 120, respectively. A good performance was found by the classifiers with RNA, verified by specificity (90%, 98%, 97%, 97% and 98%) sensitivity (93%, 77%, 93%, 75% and 82%) in the 0, 30,60,90 and 120, respectively, of the models.