Skip to main content
European Commission logo print header
Content archived on 2023-01-04

Spectral Mixture Analysis in Land Degradation, 1992-1994

Exploitable results

A forest, grassland and natural vegetation mapping and monitoring project has been set up aimed at providing information on environmental themes and biodiversity, and including productivity parameters. A watershed study and 2 grassland mapping study contracts have been defined and approved. Design of a remote sensing based mapping and monitoring system for forest ecosystems has been undertaken. This includes a land degradation aspect. The key thematic points considered were: area; resource (ie volume); damage; ecological value; relation to landscape; stratification and regionalization; nomenclature (to ensure data compatibility). A stratification and regionalization study has been initiated and an expert workshop prepared on nomenclature. Research on providing standardized and operational methods for mapping forest ecosystems on a European scale has focussed on: radiometric correction; image segmentation; spectral phenological profiles; automatic classification (in particular neural networks); spectral mixture modelling. A software package is being developed to provide an integrated preprocessing and processing environment for classifying land vegetation. It is currently being developed within an existing commercial image processing system running in most UNIX environments.
Mapping and continuous monitoring of vegetation and soil conditions has been undertaken in the Mediterranean member states of the European Union, along with identification of the extension and time dynamics of degradation processes, and standardized interpretation of satellite based indices for producing land degradation indicators. The suitability of spectral mixture analysis was investigated as a tool for mapping soil conditions and vegetation abundance under controlled experimental conditions. Based on the excellent results obtained during these experiments it was decided to analyse the possibility of employing this approach as a standardized processing scheme for land degradation monitoring. The Landsat image from 1985 was chosen as reference image, and atmospherically corrected through a radiative transfer model. The approach proved its ability to retrieve ground reflectances with an accuracy of about 10% even under difficult atmospheric conditions. The scene from 1990 was then radiometrically rectified to the 1985 scene by using linear transformations which were obtained through a regression analysis of pixel spectra from nonvegetated surfaces, and both scenes spectrally unmixed. The resulting fraction images were normalised for shade, in order to analyse the changes in green vegetation abundance independently from illumination effects. Soil conditions were also examined. It was found that 40% of the study region had been affected by a considerable reduction of natural vegetation cover. This was confirmed by the fact that the soils within these areas are dominated by degraded categories suggesting that accelerated soil erosion processes are linked to the reduced vegetation cover density. Recent field checks have confirmed that fire impact is an important cause of the degradation processes. This an approach has been developed allowing each image to be viewed as the result of a combined analysis of soil and vegetation conditions. This system needs to be refined and developed but has the advantage that any updating of validation rules is easily incorporated.
Spectral mixture analysis was applied to multi temporal Landsat thematic mapper (TM) and advanced visible infrared imaging spectrometer (AVIRIS) data in order to analyse the suitability of linear spectral unmixing for identifying soil degradation and the erosional state of soils on landscape scale (soil condition mapping). These experiment objectives were successfully achieved, providing both quantified assessments of soil conditions and precise mapping of soil conditions with AVIRIS and Landsat TM data. It was possible to overlay the resulting maps on topographic maps at scale 1 to 50 000, so that they can be considered a valuable alternative to conventional mapping approaches. One of the most important aspects of the work is that the method is reproducible (standardized data processing, including the use of spectral libraries), and that it can be applied to other study sites with equivalent or similar lithological conditions. For the purpose of mapping green vegetation abundance, a strategy was adapted which aimed at minimizing the number of endmember spectra while optimizing the selection of these endmembers for each pixel. Band residuals and, to a lesser extent, the root mean squared unmixing error seemed to provide valid criteria for controlling this selection process. The approach was based on the assumption that the mixed reflectance signature of a pixel is primarily conditioned by 3 components which represent so called foreground and background materials, and shade (in order to account for albedo differences). The method was applied to the high spectral resolution AVIRIS data from the study site. In this case, several 3 endmember configurations were tested for each pixel, and the algorithm finally selects the endmember set which produces the smallest band residuals between measured and modelled spectrum.
Methodologies for the development of fully automated techniques for land cover mapping and land cover map updating from satellite imagery have been investigating using: state of the art approaches to automated image analysis; new techniques for spatial generalization of image data to a desired map scale; analysis and comparison of photo interpretation products, automated image products and associated nomenclature issues. Experiments were conducted on integrating data from the optical and near infrared part of the electromagnetic spectrum with radar data. Experiments were carried out on classification of integrated Landsat thematic mapper (TM) imagery and earth remote sensing satellite (ERS)-1 synthetic aperture radar SAR imagery in Portugal. In addition significant improvements were made in techniques for spatial analysis of satellite imagery with the eventual aim of automating the derivation of land cover vector maps from raster based satellite imagery. The generalization technique based on iterative majority filtering and iterative reduced class growing was perfected and demonstrated in a variety of experiments. Also, an improved segmentation technique was developed based on the integration of robust edge detection and edge following methods with a best merge region growth approach. A prototype computer system has been developed which is based on proprietary geographical information system (GIS) software, well adapted to the structure of the database. The system as a whole has functions for satellite image processing, processing of geographic data and on line updating of the existing database. Statistics on land cover changes will also be a feature.
Techniques for exploiting complex remotely sensed imagery for environmental mapping and ecosystem monitoring purposes have been developed and improved. The techniques included: combination and effective use of multisensor multisatellite imagery; development of robust image classification techniques; integration of neural and statistical image analysis methods; effective exploitation of ancillary and contextual information; utilization of methods from machine vision in remote sensing. Experiments were carried out on using integrated data from conventional optical and infrared sensors and radar imagery. These experiments demonstrated that through the use of integrated coincident Landsat thematic mapper (TM) scenes and earth remote sensing satellite (ERS) synthetic aperture radar (SAR) scenes it was possible to obtain considerable improvements in the visual quality and absolute accuracy of derived land cover products. In order to use the integrated datasets, procedures were put in place for operationally coregistering Landsat TM images and ERS-1 SAR images. A practical technique was also implemented for removing speckle noise from the SAR imagery by making use of a multidimensional regression approach based on the coincident Landsat imagery. Experiments were also conducted on integrating neural network and statistical classifiers. It was found that significant gains could be achieved in classification accuracy. This takes advantage of the fact that in a typical classification problem, some classes are well modelled by statistical distributions whereas others are better modelled by a semilinear approach as implemented in multilayer perceptron neural networks. In order to demonstrate the additional discrimination power of combined sensor data, an experimental test was carried out to map the diversity of forest species in Portugal. The experiment showed that integrated TM and SAR data could be used to map 8 separate forest classes with an average of 80% accuracy including both broad leaved and coniferous varieties.
Research has been undertaken to explore how fractal methods can be used to assist in the understanding of satellite imagery. An extensive bibliographic search was carried out on the existing uses of fractals in remote sensing and the geosciences. The following main applications were identified: characterization of terrain angularity or roughness, including mountains, ice and snow surfaces, geomorphological processes, application to digital terrain model quality evaluation and cartographic generalization; characterization of 2-dimensional image textures, including the linking of fractal textures to human visual perception of texture, use of fractal textures to improve land use classification in satellite imagery and use of fractal textures for image segmentation; characterization of climatic and atmospheric phenomena; image compression using fractal models based on iterated function systems (IFS), including use of neural networks for IFS generation; synthetic image generation. Soil erosion prediction was investigated by fractal surface analysis to see if the fractal analysis of pixel brightness in airborne and satellite imagery could be used directly to infer the topographic properties of bare or partially denuded soil surfaces without the need for complex stereo image acquisition and analysis, especially in zones prone to soil erosion. A project has been set up aimed at the development of software tools for measuring fractal dimension in remotely sensed images.

Searching for OpenAIRE data...

There was an error trying to search data from OpenAIRE

No results available