Improving deep learning algorithm performance on remote sensing scene image analysis tasks for practical applications remains a challenge due to high within-class diversity and between-class similarity. To address this problem, we propose an approach that involves conditionally learning deep features directly over neighbor scene images. The approach utilizes a siamese network architecture to improve discriminative capability of Convolutional Neural Networks. It works by exploiting semantic similarity between adjacent input image pairs for enriching the feature vector of the image patch for which we want to predict a label. Our experimental results are consistent with performance improvements reported by previous work. For example, our model improved overall accuracy by 1 percentage point and dropped the mean squared error value by 0.02 over the baseline, on a disease density estimation task. These results point to potential of our approach for improving model performance by exploiting spatial dependence between scene images. They also demonstrate that deep learning algorithms are useful for estimating disease density using housing crowding signals extracted from remote sensing scene imagery. Keywords: Siamese convolutional neural networks, disease density estimation, remote sensing scene imagery, urban setting, developing countries.

Keywords: Siamese convolutional neural networks, disease density estimation, remote sensing scene imagery, urban setting, developing countries.