Use this link to cite:
https://hdl.handle.net/2183/49394 Exploración y análisis de embeddings geoespaciales de Google Earth Engine
Loading...
Identifiers
Publication date
Authors
Fernández Rubal, Pablo
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen] El objetivo de este trabajo de fin de grado es explorar y evaluar el potencial de los embeddings de AlphaEarth Foundations (AEF) para la clasificación del uso del suelo a escala municipal, tomando el ayuntamiento de Burela (Lugo) como territorio de estudio. Los embeddings de AEF son representaciones vectoriales de 64 dimensiones generadas por un modelo fundacional preentrenado de forma autosupervisada sobre más de 6 PiB de datos de observación de la Tierra sin etiquetar, disponibles públicamente en Google Earth Engine a resolución de 10 metros. Para alcanzar este objetivo se muestrearon 60.000 puntos sobre el municipio y se integraron con dos fuentes de referencia oficiales: el Catastro, que distingue entre suelo urbano y rústico, y el SIOSE, que ofrece una clasificación temática detallada de la cobertura real del terreno. Sobre estos datos se aplicaron dos técnicas de clasificación no supervisada, MiniBatchKMeans y GaussianMixture, evaluadas mediante métricas internas (Silhouette, Davies–Bouldin y Calinski–Harabasz) y externas (Purity, Homogeneity, Completeness y V-measure). Se desarrolló también un clasificador supervisado KNN entrenado con únicamente 10 muestras etiquetadas por clase sobre 6 categorías de uso del suelo. Finalmente, se implementó un dashboard web interactivo que permite explorar los resultados de la clasificación sobre el mapa del municipio y comparar los métodos y configuraciones evaluados. El proyecto se desarrolló siguiendo una metodología iterativa e incremental, con reuniones periódicas de seguimiento con el tutor. Las tecnologías principales empleadas fueron Python como lenguaje de desarrollo, con las librerías scikit-learn para los algoritmos de clasificación, GeoPandas y Shapely para el procesamiento de datos geoespaciales, y rasterio para la manipulación de datos ráster. La obtención y el filtrado inicial de los embeddings se realizó mediante la API de Google Earth Engine, y la herramienta de visualización se implementó en JavaScript con React, Recharts y Leaflet. Los resultados muestran que ambas aproximaciones logran un nivel de concordancia notable con el SIOSE: 70,6 % de acuerdo global para la clasificación supervisada y 71,6 % para la no supervisada. Aunque se debe tener en cuenta que SIOSE puede tener errores y no constituye una verdad de terreno perfecta, esto evidencia que los embeddings de AEF capturan estructuras territoriales semánticamente coherentes sin necesidad de etiquetas, y que un número muy reducido de muestras etiquetadas es suficiente para obtener clasificaciones razonables cuando se utilizan como entrada a modelos supervisados sencillos.
[Abstract] The aim of this bachelor’s thesis is to explore and evaluate the potential of AlphaEarth Foundations (AEF) embeddings for land use classification at the municipal scale, using the municipality of Burela (Lugo, Spain) as the study area. AEF embeddings are 64-dimensional vector representations generated by a foundation model pre-trained in a self-supervised manner on more than 6 pebibytes of unlabelled Earth observation data, publicly available in Google Earth Engine at 10-metre resolution. To achieve this objective, 60, 000 points were sampled across the municipality and integrated with two official reference sources: the Catastro, which distinguishes between urban and rural land, and SIOSE, which provides a detailed thematic classification of actual land cover. Two unsupervised clustering methods were applied to these data, MiniBatchKMeans and GaussianMixture, and evaluated using internal metrics (Silhouette, Davies–Bouldin and Calinski–Harabasz) and external metrics (Purity, Homogeneity, Completeness and V-measure). A supervised KNN classifier was also developed, trained with only 10 labelled samples per class across 6 land use categories. Finally, an interactive web dashboard was implemented to explore the classification results on a map of the municipality and compare the methods and configurations evaluated. The project followed an iterative and incremental development methodology, with periodic supervisory meetings. The main technologies used were Python as the development language, with scikit-learn for the classification algorithms, GeoPandas and Shapely for geospatial data processing, and rasterio for raster data manipulation. The initial retrieval and filtering of the embeddings was carried out using the Google Earth Engine API, and the visualisation tool was implemented in JavaScript using React, Recharts and Leaflet. Results show that both approaches achieve a notable level of agreement with SIOSE: 70.6 % overall agreement for the supervised classification and 71.6 % for the unsupervised one. Despite the fact that SIOSE might have update mistakes, this demonstrates that AEF embeddings capture semantically coherent territorial structures without requiring any labels, and that a very small number of labelled samples is sufficient to obtain reasonable classifications when used as input to simple supervised models.
[Abstract] The aim of this bachelor’s thesis is to explore and evaluate the potential of AlphaEarth Foundations (AEF) embeddings for land use classification at the municipal scale, using the municipality of Burela (Lugo, Spain) as the study area. AEF embeddings are 64-dimensional vector representations generated by a foundation model pre-trained in a self-supervised manner on more than 6 pebibytes of unlabelled Earth observation data, publicly available in Google Earth Engine at 10-metre resolution. To achieve this objective, 60, 000 points were sampled across the municipality and integrated with two official reference sources: the Catastro, which distinguishes between urban and rural land, and SIOSE, which provides a detailed thematic classification of actual land cover. Two unsupervised clustering methods were applied to these data, MiniBatchKMeans and GaussianMixture, and evaluated using internal metrics (Silhouette, Davies–Bouldin and Calinski–Harabasz) and external metrics (Purity, Homogeneity, Completeness and V-measure). A supervised KNN classifier was also developed, trained with only 10 labelled samples per class across 6 land use categories. Finally, an interactive web dashboard was implemented to explore the classification results on a map of the municipality and compare the methods and configurations evaluated. The project followed an iterative and incremental development methodology, with periodic supervisory meetings. The main technologies used were Python as the development language, with scikit-learn for the classification algorithms, GeoPandas and Shapely for geospatial data processing, and rasterio for raster data manipulation. The initial retrieval and filtering of the embeddings was carried out using the Google Earth Engine API, and the visualisation tool was implemented in JavaScript using React, Recharts and Leaflet. Results show that both approaches achieve a notable level of agreement with SIOSE: 70.6 % overall agreement for the supervised classification and 71.6 % for the unsupervised one. Despite the fact that SIOSE might have update mistakes, this demonstrates that AEF embeddings capture semantically coherent territorial structures without requiring any labels, and that a very small number of labelled samples is sufficient to obtain reasonable classifications when used as input to simple supervised models.
Description
Editor version
Rights
Attribution-NonCommercial-ShareAlike 4.0 International







