Exploring Recurrent and Feedback CNNs for Multi-Spectral Satellite Image Classification

Abstract The emergence of deep learning applications such as convolutional neural networks (CNNs) have resulted in huge improvements on computer vision applications in a wide variety of fields. However, several works demonstrated that low-quality or noisy data (even including perceptually not visible noises) may have a huge impact on the accuracy of CNN models. Therefore, as inspired by biological perception systems, some recent works proposed the use of recurrent and feedback features in CNNs as an improvement to the existing feed-forward CNNs. These recent works on the integration of recurrence and/or feedback to CNNs mostly tested deep networks on natural scenes with relatively perceptually good resolution color images. In this work, we explored the effectiveness of CNNs with recurrent and feedback features for the solar-power plant classification task on mid-resolution (1 pixel - 30×30 square meters per pixel) multi-spectral satellite images. Experiments show promising results when using top-down signals (especially recurrent and feedback features together) on CNNs for multi-spectral image classification tasks, outperforming the baseline CNN model without any recurrent and feedback structure and other approaches in the literature including deep models.