Genetic algorithm-based feature selection for classification of land cover changes using combined LANDSAT and ENVISAT images

Abstract

Recent monsoon failures and reduced rain falls urge the environmental and ecology researchers to concentrate on the land cover changes. Significant and efficient way to monitor the land cover changes is satellite image classification. This work describes the combination of remotely sensed data, LANDSAT and ENVISAT images, to improve the classification accuracy. Instead of predictor space, embedding space is considered in the proposed KNNES and SVMES methods and applied for the classification of combined LANDSAT and ENVISAT datasets. Genetic algorithm-based (GA) feature selection is adopted to enhance the proposed classification methods. Classification of land cover changes of the study area are identified as used land, unused land, forest and vegetation. Proposed methods are evaluated by an accuracy analysis which follows good practice recommendations. Accuracy is quantified by reporting standard errors, i.e., producer accuracy, user accuracy, omission error and commission error. Performance of the proposed SVM and KNN-based methods using GA-based feature selection for combined dataset is improved significantly and provide overall accuracy 80% and 76% respectively.