Comparison of maximum likelihood, support vector machines, and random forest techniques in satellite images classification

Abstract

Context: Nowadays, the images of the Earth surface and the algorithms for their classification are widely available. In particular, the algorithms are promising in the differentiating of cotton crops stages, but it is necessary to establish the capabilities of the different algorithms in order to identify their advantages, and disadvantages. Method: This paper describes the assessment process in which the Support Vector Machines (SVM) and random-forest technique (decision trees) are compared with the maximum likelihood estimation when differentiating the stages of cotton crops. A RapidEye satellite image of a geographic area in the municipality of San Pelayo, Cordoba (Colombia), is used for the study. Using a set of sampling polygons, a random sample of 6000 pixels was taken (2000 training and 4000 for validating the classifications.) Confusion matrices, and R (data processing and analysis software) were used during the validation process Results: The maximun likelihood estimation presented a correct classification percentage of 68.95%. SVM correctly classified 81.325% of the cases and the decision trees correctly classified 78.925%. The confidence test for the classifications showed non-overlapping intervals, and SVM obtained the highest values. Conclusions: It was possible to confirm the superiority of the technique based on support vector machines for the proposed verification zones. However, this technique requires a number of classes that comprehensively represent the variations of the image (in order to guarantee a minimum number of support vectors) to avoid confusion in the classification of non-sampled areas. This was less evident in the other two classification techniques analysed.

Keywords

• confidence test
• confusion matrix
• decision tree
• random forest
• software R
• support vector machine