Efficient Supervised Features Learning for Remote Sensing Image Classification

Abstract

The features extracted from the fully connected (FC) layers of a convolutional neural network (ConvNet or CNN) can provide accurate classification results as long as the labelled datasets are large enough. On the other end, high accuracy remote sensing image (RSI) classification is demanded various implementations such as urban planning, environmental monitoring, and geographic image retrieval. Many studies have been presented in this domain; however, satisfactory classification accuracy is yet to be achieved. In this study, the proposed method of fine-tuning the pre-trained ConvNets (GoogleNet, VGG16, and ResNet50) on RSI, extracting features from the last fine-tuned FC layer of these networks and reprocess the extracted features for classification by SVM, produced high classification accuracy. Extensive experiments have been conducted on three RSI datasets: the NWPU, AID, and PatternNet. Comparative results over the selected datasets demonstrate that our method considerably outperforms the state-of-the-art best-stated results. Also, the overall accuracy (OA) and confusion matrix report quantitative evaluation. Our best outcomes from the first part were 99.54%, 94.60%, and 94.83% on the PatternNet, NWPU, and AID datasets, respectively, achieved by fine-tuned ResNet50. Moreover, the best classification accuracies with training ratios 20% and 50% on the AID dataset, 10% and 20% on the NWPU dataset, and the 10%, 20%, 50% and 80% on PatternNet dataset were 95.72%, 97.53%, 96.19%, 96.85%, 99.60%, 99.56%, 99.75% and 99.80% respectively. The classification performance of each class was estimated using a confusion matrix for the three datasets.