ABSTRACT

        In the last years, the needs of obtaining reliable and updated information of land use/land cover have increased. Several local, national and international projects are carried out to obtain and update land use/land cover geospatial databases. Nowadays, the existing methodologies require a high amount of human intervention, because they are mainly based on image interpretation or on the comparison of an image with other image or with a database. 
        Recent advances in quality and quantity of airborne and satellite sensors have entailed an important increase in the availability of high resolution images. Besides, Spanish public administrations are working jointly in the National Plan for Territory Observation (PNOT) to acquire and to distribute satellite and aerial imagery. At the same time, new methodologies are being developed to analyze these data. Tests and developments included in this Thesis aims for contributing to the progressive automation of land use information extraction by combining image analysis of data provided by PNOT with information contained in databases.
        The aim of this Thesis is the study and evaluation of methodologies for cartographic updating of land use/land cover databases in agricultural Mediterranean landscapes combining information usually available in Spain for updating cartography projects: images of the update date, the cartographic database to be updated and ancillary information. Different data sources like LIDAR, multi-temporal images, hyper-spectral images, microwaves images, etc. haven’t been considered because their cost would limit their use in the present Spanish situation and probably, they will not be available periodically and with coverage for the whole country. 
        The update is performed through integration and analysis of cartographic information in different formats: vector cartography, alphanumeric information contained in the database to be updated and high resolution aerial images. Specifically, an ortoimage mosaic with 0.5m/pixel has been used, generated combining the panchromatic band and the multi-spectral bands acquired with the DMC photogrammetric camera system. The cartographic database used in tests is the cadastral cartography with its alphanumeric information in eleven cadastral polygons located in the municipal district of Benicarló (Castellón). Ancillary information considered is: altitude, slope and coast distance that have been extracted from Digital Terrain Models with a spatial resolution of 1m. The data integration is made by means of feature extraction and object-based image classification. The cartography provides the boundaries to define the objects, which are the cadastral sub-parcels in our work. The land use of each object is determined with the analysis of its previous use, its shape and an exhaustive set of features extracted from the high resolution image.
        Quantitative description of each object is carried out by means of a feature set specifically designed to describe agricultural parcels. Information provided by these features covers several aspects: spectral response, texture, planting pattern, objects contained in the parcel, shape of the parcel, etc. Development and application of these descriptive features are the core of this Thesis.
        In several tests, the performance of these features to describe the characteristics of the uses in the parcels has been evaluated. For example, position of the first maximum point in the omnidirectional semivariogram calculated over the infrared band is a precise indicator of diameter of crown trees planted with regular planting patterns. Besides, the Hough transform applied over a binary image of tree locations allows determining the size of planting patterns in tree crops. An automatic method to segment and to describe trees in the parcels has been designed. This method gives information regarding tree area in the parcel and tree characteristics to determine the class of tree crops.
        It has been proved, that the previous use of a parcel employed as an additional feature increases the accuracy of the classifications. This accuracy increase is related with the updating degree of the database and its value was around 4% in the tests done.
        The multi-classifier boosting was employed over a set of decision trees, created with C5.0 algorithm from a set of training samples, to assign the class to each object. In addition to the assigned class, a confidence value is got that indicates the reliability of the outcome classification being correct.
        The proposed parcel-based classification considered 10 classes, and was evaluated with a set of parcels in the working area. The overall accuracy, when the previous use was considered, was 78.4% and it was 74.7% without considering it.
        To conclude, this methodology was tested in the context of a real updating problem. Comparing the classes assigned to the parcels with the information contained in the database to update, detected discrepancies are due to changes in the land use or to classification errors and must be checked by photo-interpreters in order to determine the actual class.
        The described methodology has been successfully applied in some production works: detecting changes in the citrus inventory of Castellón, updating the crop inventory in the region of Murcia and mapping the land use for the Land Bank of Galicia. Overall accuracies obtained in these works, with different data, legends and landscapes, ranged between 78% and 93%. In all cases, the application of this methodology involves a substantial decrease in the amount of parcels to be reviewed by photo-interpreters which is reflected in a great reduction in the necessary time for updating databases.