Investigations on Deep Learning Pre-trained Model Inception-V3 Using Transfer Learning for Remote Sensing Image Classification on Benchmark Datasets

Remote sensing image classification is the efficient execution of image categorization of high spatial resolution images for large remote sensing archives. High performance of image categorization is directly based on efficient image feature extraction. Before the widespread adoption of deep learning in remote sensing image classification, the feature extraction stage relied on manually designed low level features mainly focusing on basic features color, shape, and texture. However, these traditional handcrafted methods due to its lower performance were eventually replaced by convolutional neural networks which efficiently extracted abstract features. In the domain of remote sensing, significant results in image classification could be achieved through convolution neural networks combined with transfer learning. In this letter to enrich the accuracy of image classification using transfer learning pre-trained deep learning model Inception-V3 has been used as feature extractor for image classification on four benchmark datasets: UCMerced, AID, NWPU-RESISC45, and PatterNet dedicated for remote sensing scene classification. The proposed Inception-V3 combined with transfer learning produced improved accuracy of 87% on UCMerced, 75% on AID, 89% on PatterNet, and 90% on NWPU-RESISC-45. The results demonstrate that the NWPU-RESISC45 benchmark dataset achieved the highest accuracy score of 90% surpassing UCMerced, AID, and PatterNet.