Desarrollo de metodologías para la comparación, validación e integración de información procedente de productos de cobertura del suelo
- Autores: Ana Perez Hoyos
- Directores de la Tesis: Francisco Javier García-Haro (dir. tes.)
- Lectura: En la Universitat de València ( España ) en 2011
- Idioma: español
- Tribunal Calificador de la Tesis: Joaquín Meliá Miralles (presid.), Gustavo Camps Valls (secret.), Eva Hugh Douglas (voc.), Eva María Rubio Caballero (voc.), José Antonio Martínez Casasnovas (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: TESEO
- Resumen
Land cover is an Essential Climate Variable (ECV) that describes the flow of carbon, water and energy through biosphere. Detailed and accurate land cover data provides key environmental information required for science and policy application, such as climate change, biodiversity conservatio, ecosistem assessment, hydrology and environmental modeling.
Actually, there is a large number of global and regional datasets.Four major land cover datasets have been selected in this assessment (CORINE, GLC2000, MODIS and GlobCover). Existing differences in temporal scales, data sources, classification systems and methodologies produce important discrepancies among these land cover datasets. Evaluation of these products to determine which is best for a particular purpose is the first step to reconcile these differences.
A major drawback of classification systems is that they often employ descriptive class names without explicity mentioning the criteria used to define these classes. In the recent years, different approaches based on fuzzy set theory have been proposed for the comparison and validation of land cover maps; these tke into account the semantic uncertainty on the basis of information provided by Land Cover Classification System (LCCS).
Therfore, the objective of this thesis is developing methodologies for comparison, validation and integration of land cover datasets. The methodology proposed for comparing datasets allows for differences in legend definition and positional errors among products. The classical Boolean approach on a per-pixel basis have been extended to accommodate both partial overlap between classes and geolocation errors. In addition, fuzzy land cover comparisons were implemented to evaluate similarites and discrepancies in a more robust manner. In the proposed fuzzy methodology the LCCS acted as general bridging system, taking into account the semantic uncertainty among nomenclatures. Fuzzy agreement have been thus established in terms of a set of independent attributes, offering new insights to understand the causes of discrepancies found among the different datasets.
The thesis also porposes a methodology for perfoming accuracy assessment of large scale land cover products. Different validation methods, including Boolean and fuzzy, further extended to accommodate positional and thematic errors, are proposed to perform the validation analysis of the considered datasets in a more robust and unbiased manner. The fuzzy validation introduces two desirable properties: (1) the inclusion of sub-dominant fractions, which enables reducing the bias due to the existence of more than one land cover class in a validation sample, and (2) the use of a framework for computing thematic affinity or overlap between legend classes. The proposed approach offers also estimates for the classification confidence associated with the pixel label, which is required boy potential users to increase the overal confiden in the final map.
Finally, a general framework was developed for combining different land cover products in order to produce a synergetic or hybrid land cover map. Although in essesnce the methodology relies on a voting scheme, this novel technique introduces two desirable properties: (1) the use of the accuracy of each classification on each individual class, which enables reducing the effect of major errors in individual classifications, and (2) the use of LCCS acting as a general bridging system for the translation among land cover products, which allows for reconciling existing ambiguities in legend definitions. In addition to the hybrid map, an estimate of the classification uncertainty assigned to each pixel is provided. This information is required to increase the usr's overall confidence in the final map.
