Tai, X.; Wang, G.; Grecos, C., and Ren, P., 2020. Coastal image classification under noisy labels. In: Jung, H.-S.; Lee, S.; Ryu, J.-H., and Cui, T. (eds.), Advances in Geospatial Research of Coastal Environments. Journal of Coastal Research, Special Issue No. 102, pp. 151-156. Coconut Creek (Florida), ISSN 0749-0208.
Remote sensing coastal images commonly suffer from noisy labels, which tend to reduce the benefits of training coastal image classifiers and thus deteriorate the classification performance. This article explores the feasibility of classifying coastal images under noisy labels. The state-of-the-art co-teaching paradigm in computer vision is exploited for processing remote sensing coastal images that are partially incorrectly labeled. The data with noisy labels are hard to learn and induce large loss values for a classifier. In the light of this observation, the co-teaching paradigm trains two identically structured but randomly initialized convolutional neural network (CNN) classifiers with stochastic optimization. Training of the two CNN classifiers is performed on mini-batches in an interactive and iterative manner. Specifically, the data with small loss values of one CNN classifier are used for updating the other CNN classifier in the next training iteration. The two CNN classifiers turn robust to noisy labels through the iterative process of teaching each other. Experimental evaluations on a coastal image dataset, i.e., the Moye Island remote sensing dataset, validate the effectiveness of the proposed coastal image classification strategy.