Unsupervised Deep Clustering on Spatiotemporal Objects Extracted from 4D Point Clouds for Automatic Identification of Topographic Processes in Natural Environments

Abstract. Topographic processes, such as sediment erosion, accumulation, and transport are crucial for understanding the evolution of natural landscapes. Current developments in permanent laser scanning (PLS) technology and 4D change detection methods have made it possible to extract spatiotemporal change objects from near-continuous 3D observations, e.g., 4D objects-by-change. However, the automatic characterization and identification of these processes remain challenging due to the complex spatiotemporal data and unpredictable types of topographic processes in natural environments. In this paper, we present a time series-based unsupervised deep clustering framework for identifying topographic processes without manual feature engineering and annotations. By leveraging the representation learning capability of autoencoders, especially using convolutional neural networks (CNNs) as feature extractors, our approach implements the dimensionality reduction of the original inputs to uniform low-dimensional vectors in latent space. Subsequently, after jointly optimizing the reconstruction and clustering loss, our model generates unique clusters with high intra-cluster similarity and inter-cluster variability. We validated the proposed method on a six-month 4D dataset, acquired at Kijkduin sandy beach (The Netherlands), yielding distinctive clusters that correspond to sediment change phenomena. Our results demonstrate that the deep learning-based method successfully identifies topographic processes, providing an efficient and scalable alternative to traditional feature engineering-based approaches. This work highlights the potential for automating topographic process identification and supporting long-term environmental monitoring.