End-to-end learning for image-based air quality level estimation

Air quality estimation is an important and fundamental problem in environmental protection. Several efforts have been made in the past decades using expensive sensor-based or indirect methods like based on social networks; however, image-based air pollution estimation is still far from solved. This paper devises an effective convolutional neural network (CNN) to estimate air quality based on images. Our method is comprised of three ingredients: We first design an ensemble CNN for air quality estimation which is expected to obtain more accurate and stable results than a single classifier. Second, three ordinal classifiers, namely negative log-log ordinal classifier, cauchit ordinal classifier and complementary log-log ordinal classifier, are devised in the last layer of each CNN, to improve the ordinal discriminative ability of the model. Third, as a variant of the rectified linear units, an adjusted activation function is introduced. We collect open air images with corresponding air quality levels from an official agency as the ground truth. Experimental results demonstrate the effectiveness of our method on the real-world dataset.