Abstract
Temporal sequences of satellite images constitute a highly valuable and abundant resource for analyzing regions of interest. However, the automatic acquisition of knowledge on a large scale is a challenging task due to different factors such as the lack of precise labeled data, the definition and variability of the terrain entities, or the inherent complexity of the images and their fusion. In this context, we present a fully unsupervised and general methodology to conduct spatio-temporal taxonomies of large regions from sequences of satellite images. Our approach relies on a combination of deep embeddings and time series clustering to capture the semantic properties of the ground and its evolution over time, providing a comprehensive understanding of the region of interest. The proposed method is enhanced by a novel procedure specifically devised to refine the embedding and exploit the underlying spatio-temporal patterns. We use this methodology to conduct an in-depth analysis of a 220 km² region in northern Spain in different settings. The results provide a broad and intuitive perspective of the land where large areas are connected in a compact and well-structured manner, mainly based on climatic, phytological, and hydrological factors.
Introduction
Earth monitoring is essential for tracking climate change and land use, but it faces a massive "data labeling" bottleneck. While we have petabytes of satellite imagery, labeling that data for supervised machine learning is prohibitively expensive and slow. Furthermore, traditional pixel-level analysis often fails to capture the "semantic meaning" of a landscape—the high-level context that defines a region, such as a transition zone between a forest and a valley.
Large-scale analysis also struggles with technical artifacts like "border effects," where different satellite passes do not perfectly align. This research addresses these challenges by proposing a fully unsupervised methodology to create "spatio-temporal taxonomies" of vast regions. By combining deep semantic embeddings with time-series clustering, the system automatically identifies complex geographical structures from unlabeled Sentinel-2 image sequences.
Methodology
The researchers developed a multi-stage pipeline designed to move from raw pixels to abstract semantic clusters:
- Tile2Vec Adaptation: The framework utilizes a Convolutional Neural Network (ResNet-18) based on the "Tile2Vec" approach. This method leverages the First Law of Geography: things that are near each other are likely more similar than things far away. The network learns to map image patches (tiles) into a 512-dimensional "latent space."
- Multivariate Time Series (MTS): Rather than analyzing static images, the model represents each location as a sequence of these 512-D vectors over time. This captures the "pulse" of the land across seasons, reflecting phenological and climatic cycles.
- K-Means Clustering: These time-series representations are clustered to find regions with similar seasonal dynamics.
- Iterative Refinement: This is the paper's primary novelty. The clusters identified in the previous step are used as "pseudo-labels" to re-train the embedding network. This process sharpens the model's understanding, pushing tiles from the same cluster closer together in the latent space while pulling different clusters further apart.
Study Area and Data
The study focused on a 220 km² region in Northern Spain (Navarre). This area was selected for its extreme environmental diversity, encompassing the humid Cantabrian coast, the alpine Pyrenees, and the arid Ebro basin.
The dataset utilized Sentinel-2 satellite imagery (RGB and Near-Infrared bands) spanning five years (2017–2021). Images were sampled seasonally (4 images per year), totaling 80 high-resolution images. The region was decomposed into 100x100 pixel tiles, each representing approximately 1 km².
Results and Discussion
The methodology demonstrated significant quantitative and qualitative advantages over traditional methods.
Quantitative Performance
The embedding-based approach vastly outperformed traditional pixel-based clustering. In the Northeast region, the Calinski-Harabasz (C-H) score—a standard measure of cluster quality—was 4463 for the proposed method, compared to only 750 for pixel-based methods. This indicates that the refined embeddings create much more distinct and cohesive groupings.
Geographical Accuracy and Expert Validation
The system successfully identified seven distinct clusters that aligned with real-world climatic and phytological zones. These included the "Oceanic climate of the Cantabrian coast" and the "Alpine climate of the Pyrenees." A professional geographer validated these unsupervised results, confirming that the AI-discovered boundaries were geographically and ecologically meaningful.
Latent Interpolation and Robustness
Because the latent space is continuous, the authors could mathematically interpolate between cluster centers. This allowed for the visualization of "transitional" landscapes, showing how a valley gradually transforms into a mountain range. Furthermore, the model proved naturally resistant to "border effects" and sensor inconsistencies, as it focuses on high-level semantics rather than raw pixel values.
Conclusion
This research provides a robust framework for unsupervised large-scale land analysis. By letting the data "speak for itself," the methodology overcomes the labeling bottleneck and provides intuitive perspectives on land evolution. Future work will explore applying this framework to monitor desertification, conduct river basin analysis, and implement "bi-clustering" algorithms to identify sub-regions with similarities during specific temporal windows.
Summary updated