DeepCropClustering: A deep unsupervised clustering approach by adopting nearest and farthest neighbors for crop mapping

Abstract

Existing crop type maps usually rely on extensive ground truth, limiting the potential applicability in regions without any crop labels. Unsupervised clustering offers a promising approach for crop mapping in regions lacking labeled crop samples. However, due to the high-dimensional complexity and pronounced temporal dependencies of crop time series, existing unsupervised clustering methods are inadequate for effectively capturing deep semantic representations. In this study, we developed a novel deep unsupervised clustering approach, named DeepCropClustering (DCC), for crop mapping without any crop label information. This approach includes a generating cluster feature space component to acquire the semantically meaning features via contractive learning and a learnable deep clustering component for unsupervised clustering using the nearest-farthest neighbor information derived from feature spaces sorting. DCC achieved an average Overall Accuracy (OA) of 70.9% across the four sites, surpassing the average OA of K-means and GMM by 7.0% and 8.6% respectively. Evaluation results of the cluster feature space indicated that the generated feature space contained reliable far-neighbor and near-neighbor samples, providing highly discriminative feature representations. By monitoring the clustering confidence during each training iteration, we found that clustering reliability increased progressively throughout the learning process, gradually converging to appropriate clusters. DCC does not require any crop labels during the clustering process, offering a new option for crop mapping in regions without crop labels and has the potential to become a new method for large-scale crop mapping.