Deep Ensemble Network for Handling Class-Imbalance Problem in Land-Cover Classification

In this manuscript, a bagging based random undersampling technique using the ensemble of deep convolutional neural networks (DCNNs) has been investigated to deal with the issue of the class-imbalance problem in land-cover classification (LCC). The proposed method is intended to address a scenario in object-level LCC where some land-cover classes outnumber other classes in a specific region, and as a result, the training set contains fewer samples from some (minority) classes compared to other (majority) classes. Here, initially, some bags consisting of balanced classes from the dataset is acquired using random undersampling technique. After that, samples from each bag are used to independently fine-tune a pre-trained convolutional neural network (CNN). Finally, the decisions of all trained finetuned CNN models are combined to assign the class-label for each unlabeled test sample. In order to validate the efficiency of the proposed scheme, experiments have been conducted on the University of California, Aerial Image Dataset, and a new remotely sensed aerial image dataset obtained across the Asian subcontinent region (ASCD). The results obtained over these three different aerial image datasets provide superior performance for the proposed scheme as compared with the other state-of-the-art techniques.