Aerial Scene Classification through Fine-Tuning with Adaptive Learning Rates and Label Smoothing

Abstract

Remote Sensing (RS) image classification has recently attracted great attention for its application in different tasks, including environmental monitoring, battlefield surveillance, and geospatial object detection. The best practices for these tasks often involve transfer learning from pre-trained Convolutional Neural Networks (CNNs). A common approach in the literature is employing CNNs for feature extraction, and subsequently train classifiers exploiting such features. In this paper, we propose the adoption of transfer learning by fine-tuning pre-trained CNNs for end-to-end aerial image classification. Our approach performs feature extraction from the fine-tuned neural networks and remote sensing image classification with a Support Vector Machine (SVM) model with linear and Radial Basis Function (RBF) kernels. To tune the learning rate hyperparameter, we employ a linear decay learning rate scheduler as well as cyclical learning rates. Moreover, in order to mitigate the overfitting problem of pre-trained models, we apply label smoothing regularization. For the fine-tuning and feature extraction process, we adopt the Inception-v3 and Xception inception-based CNNs, as well the residual-based networks ResNet50 and DenseNet121. We present extensive experiments on two real-world remote sensing image datasets: AID and NWPU-RESISC45. The results show that the proposed method exhibits classification accuracy of up to 98%, outperforming other state-of-the-art methods.

Towards Detecting Pneumonia Progression in COVID-19 Patients by Monitoring Sleep Disturbance Using Data Streams of Non-Invasive Sensor Networks

“…Choices are abundant for the machine learning algorithms, and the state-of-the-art and future challenges and research directions for medical datasets were described in Kalantari et al [ 46 ]. Almost all algorithms require some hyper-parameter tuning [ 47 ], and their performances can significantly vary on the choice of the cross-validation approach. In the case of medical datasets, which often exhibit a high class imbalance, a leave-one-subject approach is preferable, leading to an overall reliable estimation of the classifier performance [ 48 ].…”

Multi-Horizon Air Pollution Forecasting with Deep Neural Networks

“…Similarly, in [ 51 ], an attention-based approach was applied to all available weather and pollution information. Alternative studies in the literature exploit feature extraction as a preprocessing step for the predictive task [ 52 , 53 , 54 , 55 ].…”

Modules and Techniques for Motion Planning: An Industrial Perspective

“…A second family of strategies, i.e., one-stage detectors, are described by Redmonn et al [ 20 ] with YOLO and its variants. Convolutional neural networks have also been used to solve the problem of high-resolution aerial scene classification, as the one presented by Petrovska et al [ 21 ].…”