End-to-End Deep Reconstruction and Clustering of Satellite Image Time Series for Crop Group Mapping

Classification of annual crops according to their seasonality (crop group classification) is beneficial for various applications, for instance, drought prevention. Using satellite image time series (SITS) allows access to the temporal patterns that correspond to the phenology of the crops. On one hand, training data for such applications may be scarce, especially when relatively wide scenes are considered, thus prompting the need for unsupervised approaches. On the other hand, optical SITS usually exhibit missing data, e.g., due to clouds. One common practice to address unsupervised classification in SITS with missing values is to perform gap filling prior to the clustering task. The choice of methods for both tasks is often performed independently, which may generally influence the effectiveness of the resulting clustering solution. In this article, we propose a neural architecture that performs gap filling and clustering simultaneously in an end-to-end framework for crop group classification. In comparison to previous deep clustering methods that conduct representation learning and cluster assignment separately, these tasks are also addressed jointly by the proposed approach, together with the gap-filling task. For this purpose, the developed method combines a generative component, based on a conditional generative adversarial network, with an information-theoretic clustering component. Experimental results on two datasets suggest the effectiveness of the proposed method as well as the relevance of all of its architectural components.