Evaluation of different classification techniques for the detection of glass and plastic greenhouses from WorldView-2 satellite imagery

Abstract

Greenhouse detection using remote sensing technologies is a challenging task for urban and rural planning, agricultural yield estimation, natural resource management, and sustainable development. The main objectives of this research are (1) to compare the supervised classification techniques including maximum likelihood (ML), random forest (RF), and support vector machines (SVM) for land cover classification with emphasis on greenhouse detection and (2) to investigate the utility of WorldView-2 satellite imagery for detecting and discriminating plastic and glass greenhouses. The study was implemented in an area selected from Muratpasa district of Antalya, Turkey, which includes both glass and plastic greenhouses extensively. Overall, the computed classification accuracies are quite high using all classification techniques. However, the most accurate classification result was obtained using SVM classifier, with overall accuracy of 93.88%. The performance of RF classification is slightly lower than SVM classification, with overall accuracy of 91.73%. On the other hand, SVM and RF classifiers significantly outperformed ML classifier, providing higher overall accuracies. When plastic greenhouse detection accuracies are considered, the obtained results revealed that all of the used classification techniques are quite successful. However, for glass greenhouse detection, the SVM and RF classifiers provided significantly higher accuracies than ML.