Annual Land Use/Land Cover (LULC) Dataset for M'Sila, Algeria (1985–2025)
Description This dataset provides a temporally consistent annual Land Use/Land Cover (LULC) dataset for the Municipality of M'Sila, Algeria, covering the period from 1985 to 2025. The dataset was generated from the Landsat archive using a deep learning framework developed in Google Earth Engine and described in the associated peer-reviewed publication. The annual maps were produced using Landsat 5, Landsat 7, and Landsat 8 imagery through a Deep Neural Network (DNN)-based classification framework. To improve the spectral consistency of the training samples, the workflow incorporates an additional refinement step based on the Spectral Angle Mapper (SAM). The complete methodology, validation procedure, and accuracy assessment are fully described in the associated publication. Dataset Contents The archive (.zip) is organized as follows: README.txt — Documentation filelulc_symbology.qml — QGIS style file (colors + class labels)lulc_symbology.lyrx — ArcGIS Pro style file (colors + class labels)rasters/ — Folder containing 41 annual GeoTIFF files File naming convention (inside rasters/): lulc_1985.tif, lulc_1986.tif, ..., lulc_2025.tif Each raster represents the annual land use/land cover map for the corresponding year. Data Specifications Study area: Municipality of M'Sila, AlgeriaTemporal coverage: 1985–2025Temporal resolution: AnnualRaster format: GeoTIFF (single-band)Coordinate Reference System (CRS): WGS 84 (EPSG:4326)Nominal spatial resolution: 30 m (derived from Landsat imagery)NoData value: 0 Each raster encodes the following six LULC classes as pixel values: 1 – Impervious surfaces2 – Trees & dense vegetation3 – Scattered/low-density vegetation4 – Water bodies5 – Bare land6 – Agricultural land without vegetation cover Symbology Files Two ready-to-use symbology files are included to visualize the LULC classes with the correct colors and labels: lulc_symbology.qml — for QGIS (Layer Properties → Symbology → Load Style)lulc_symbology.lyrx — for ArcGIS Pro (Symbology → Import from layer file) Without these files, rasters will display as a continuous surface by default, since no raster attribute table (VAT) is embedded in the GeoTIFF files. Methodology These data were generated using the deep learning methodology described in the following publication:Tahmi, S., Rui, X., Guechi, I., Hannache, F., Wang, X., & Chebika, H. (2026). Mapping four decades of change: Deep learning and Google Earth Engine for annual LULC classification in semi-arid regions. Environmental Modelling & Software. https://doi.org/10.1016/j.envsoft.2026.106940 Users interested in reproducing the workflow or understanding the processing chain should refer to the publication above, where the complete methodology, training strategy, preprocessing steps, validation procedures, and performance evaluation are presented. Processing Notes The classified rasters represent the direct, unfiltered output of the pixel-wise classification model. No post-classification smoothing, majority filtering, or sieve operations were applied to remove isolated or scattered misclassified pixels. Users requiring a smoother or generalized output for cartographic purposes may wish to apply their own post-processing filters (e.g., majority filter) as needed. Citation If you use this dataset in your research, please cite both this dataset and the associated publication. Notes This dataset is intended for research, environmental monitoring, land change analysis, and related scientific applications. Users are kindly requested to cite both this dataset and the associated publication when using these data in scientific outputs. This dataset is distributed under the CC-BY 4.0 license, permitting reuse, redistribution, and adaptation, including for commercial purposes, provided appropriate credit is given. Note on 2025 data: the 2025 layer is based on a partial-year (Jan–Aug 2025) composite due to imagery availability at the time of processing.