Land cover classification in Portugal with intra-annual time series of MERIS images
Hugo Carrão(1), António Araújo(1) and Mário Caetano(1)
(1) Portuguese Geographic Institute, Rua Artilharia Um, 107, 1099-052 Lisboa, Portugal
Remote sensing data have been extensively used to generate land cover and land use maps for a variety of purposes through a wide range of classification approaches. Land cover products became very popular in forestry, geology, landscape analysis and management, or simply in biodiversity assessment and monitoring, giving real time information about the spatial distribution of natural features and their changing patterns. At regional level, a number of remotely sensed data sources, including Landsat Thematic Mapper (TM), Satellite Probatoire d’Observation de la Terre (SPOT), and Advanced Very High Resolution Radiometer (AVHRR), have been commonly applied to derive comparable land cover products. However, the AVHRR instrument possesses only two broad spectral bands for land observation that are sometimes insufficient to distinguish subtle differences in vegetation types with similar annual phenologies. On the other hand, sensors with higher spatial and spectral resolutions, such as Landsat TM and SPOT, have incomplete spatial coverage, infrequent temporal coverage with inevitable atmospheric contamination, and the associated large data volumes not practicable in operational program contexts, thus making difficult the regular production of comparable regional land cover products.
Most recently launched medium spatial resolution Earth Observation (EO) sensors, such as the MEdium Resolution Imaging Spectrometer (MERIS) and the Moderate Resolution Imaging Spectroradiometer (MODIS), exhibit enhanced spatial, temporal and spectral resolutions, thus providing a wide range of new possibilities for the regular production of land use-cover maps at regional scale. However, remote sensing data classification remains a challenge because of many factors, such as the complexity of the landscape in the study area, data source, image processing and classification approaches. In this paper we present the methodological approach developed by the Remote Sensing Unit (RSU) of Portuguese Geographic Institute (IGP) for the regular/operational production and updating of land cover maps derived from medium spatial resolution satellite images in Portugal. The focus of this paper is to explain the adopted land cover nomenclature, the satellite images selection process, the classifier selection, the training and testing samples collection processes and final map accuracy assessment design.
We start by presenting the proposed land cover nomenclature for our maps, which was defined through the Land Cover Classification System (LCCS) from Food and Agriculture Organization (FAO). The rationale behind the development of this 16 land cover classes nomenclature was three-fold: (1) a nomenclature that is well adapted to the type of landscapes existent in regions with characteristics similar to the Portuguese mainland, (2) a nomenclature that is compatible with established ones (e.g., CORINE Land cover, Global Land cover and the International Geosphere-Biosphere Programme nomenclatures) in order to turn possible the comparison between our maps and others using different nomenclatures, and (3) a nomenclature that matches the spatial resolution of used satellite imagery.
Then, the most adequate satellite images and acquisition dates for land cover discrimination in Portugal are depicted accordingly. Previous studies have shown that high dimensional spectral data, acquired within consecutive periods of the year, provide adequate information in quantifying biophysical characteristics of vegetation. Thus, in this context, we explore classification techniques that take advantage of satellite images acquired by sensors with high temporal resolutions, and justify image selection and compositing on that basis. An intra-annual set of MERIS Full Resolution images was considered as the most adequate for that task. In the sequence we discuss about one of the most important processes of map production, i.e. classifier choice. Classifiers applied to different problems and trained by different sets perform differently. Thus, we performed an experimental comparison of several classifiers on our dataset. Maximum Likelihood (ML), Support Vector Machines (SVM), Linear Discriminant Classifier (LDC), minimum distance, K-Nearest Neighbours and Self-Organizing Maps (SOM) were investigated. SVM appeared to be the most adequate for this specific undertaking. However, SVM is not yet implemented in most image processing softwares and thus the results we present at this moment were derived with LDC.
The assessment of final land cover map involved the estimation of overall accuracy, and user’s and producer’s accuracies for particular land cover classes. Accordingly, the use of a stratified random sampling design seemed to be the most adequate to ensure that all classes hold an adequate number of sample observations for accuracy measures estimation at a certain confidence level.
The final land cover map produced for Portugal within our approach shows an overall accuracy of almost 80%, which is a promissory basis for our survey. These preliminary results allow for the categorization of a set of future actions, namely land cover nomenclature redefinition to better accommodate the linkage between images’ spatial resolution and landscape characteristics; the development of a toolbox for image classification including innovative classifiers; and a method for automatic update of training and testing samples for forthcoming years.
Keywords: ESA European
Space Agency - Agence spatiale europeenne,
observation de la terre, earth observation,
satellite remote sensing,
teledetection, geophysique, altimetrie, radar,
chimique atmospherique, geophysics, altimetry, radar,