Data and Code for "Discovering landscape transformation regimes using unsupervised trajectory clustering: a hemeroby-based approach for Belgium"

Data and Code for "Discovering landscape transformation regimes using unsupervised trajectory clustering: a hemeroby-based approach for Belgium" This repository contains the datasets, analysis code, derived data products, and the interactive figure associated with the manuscript "Discovering landscape transformation regimes using unsupervised trajectory clustering: a hemeroby-based approach for Belgium". The study converts CORINE Land Cover (CLC) data for Belgium into hemeroby indices (1 to 7, after Walz & Stein, 2014) and applies an unsupervised pipeline (Dynamic Time Warping distance, UMAP, HDBSCAN) to identify landscape transformation regimes among transitioning pixels over the period 1990 to 2018. Note on file naming Files and scripts carry a filtered_ (or filtered-) prefix. The analysis separates stable pixels, whose hemeroby grade does not change across the five dates and which make up the large majority of the study area, from transitioning pixels, whose grade does change. An earlier, unfiltered pipeline embedded and clustered all pixels together, where the dominant stable mass left only coarse structure. This filtered pipeline sets the stable pixels aside before the UMAP embedding and HDBSCAN clustering, so the transition typologies resolve; the stable pixels are re-attached, by hemeroby level, in the combined GeoTIFF. The prefix distinguishes the outputs of this filtered pipeline from that full-pixel version. This repository contains the filtered pipeline, which produces the transition clusters reported in the paper. Data files CLC1990.tif, CLC2000.tif, CLC2006.tif, CLC2012.tif, CLC2018.tif: CORINE Land Cover rasters clipped to Belgium for the five reference years. Raster values are CLC index codes; the index-to-code mapping is given at the top of filtered-umap-dimred.py. Source: Copernicus Land Monitoring Service / EEA (see Data sources). stacked_datasets.npz: the five CLC rasters stacked into a single array (height x width x 5 timesteps), used as the input to filtered-umap-dimred.py. filtered_embedding_n15_min0.02.npz: the reference UMAP embedding and all metadata needed to reproduce the clustering and the GeoTIFF. Arrays: embedding (2-D UMAP coordinates per transitioning pixel), unique_trajectories and inverse_indices (the unique hemeroby trajectories and their mapping back to pixels), transition_trajectories and all_trajectories (per-pixel hemeroby sequences), valid_mask / stable_mask / transition_mask and original_shape (to rebuild the raster grid), stable_classes (hemeroby level of each stable pixel), and umap_params. UMAP is configured with n_neighbors = 15 and min_dist = 0.02. Code Python 3.x. The required libraries are listed at the top of each script. Run the scripts in the numbered order; the three visualization scripts can be run in any order once the embedding exists. filtered-umap-dimred.py: converts CLC codes to hemeroby grades, separates stable from transitioning pixels, computes DTW distances on the unique transition trajectories, and embeds them with UMAP. Produces the filtered_embedding_*.npz files. filtered-hdbscan-param-explore.py: sweeps HDBSCAN parameters (min_cluster_size, min_samples, selection method) on a single embedding and reports cluster count, noise fraction, and silhouette score, to justify the chosen settings. filtered-clustering-stability.py: runs HDBSCAN across all embedding configurations and computes the pairwise Adjusted Rand Index (ARI) matrix used to confirm the robustness of the cluster solution. filtered-export-clusters.py: clusters the reference embedding (HDBSCAN with cluster_selection_method = leaf, min_cluster_size = 20, min_samples = 15) and writes the combined GeoTIFF and a per-class summary CSV. Visualization: filtered-plot-umap-clusters.py: static publication scatter of the embedding coloured by cluster. filtered-trajectory-vis.py: per-cluster hemeroby trajectory profiles (median, IQR, and 10th to 90th percentile) and stacked-composition figures. filtered-umap-interactive.py: builds the interactive HTML embedding (umap_embedding_clusters_n15_min0.02_interactive.html); each point is a unique hemeroby trajectory, coloured by cluster, with the full hemeroby sequence shown on hover. Derived products and figures filtered_clusters_n15_min0.02.tif: combined GeoTIFF. Value scheme: -1 = nodata; 0 = noise (unclustered transitions); 1 to 6 = transition clusters TC0 to TC5; 101 to 107 = stable pixels by hemeroby level (1 to 7). The raster inherits the CRS of the source CORINE rasters, and the colour table is embedded in the file. filtered_clusters_n15_min0.02.qml: QGIS layer style (symbology and class labels) matching the GeoTIFF value scheme above. umap_embedding_clusters_n15_min0.02_interactive.html: the interactive embedding figure; self-contained and opens in any modern browser. filtered_cluster_summary_n15_min0.02.csv: per-class pixel counts and median and interquartile hemeroby per timestep, behind the cluster table in the paper. Workflow workflow.png: the workflow, from CLC input through hemeroby conversion, DTW and UMAP embedding, HDBSCAN clustering, and expert-guided spatial validation, summarising the steps needed to reproduce the analysis. Data sources This publication has been prepared using European Union's Copernicus Land Monitoring Service information; https://doi.org/10.2909/c89324ef-7729-4477-9f1b-623f5f88eaa1, https://doi.org/10.2909/ddacbd5e-068f-4e52-a596-d606e8de7f40, https://doi.org/10.2909/08560441-2fd5-4eb9-bf4c-9ef16725726a, https://doi.org/10.2909/a84ae124-c5c5-4577-8e10-511bfe55cc0d, https://doi.org/10.2909/960998c1-1870-4e82-8051-6485205ebbac, more specificallyCORINE Land Cover (1990, 2000, 2006, 2012, 2018), Copernicus Land Monitoring Service, European Environment Agency, provided under the Copernicus open data policy. Please credit Copernicus / EEA when reusing the rasters. License Creative Commons Attribution 4.0 International (CC BY 4.0). The associated journal article will be linked as a related identifier once published.