Multiscale remote sensing for assessment of environmental change in the rural-urban fringe.

Abstract

The objective of this study was to investigate the application of multiscale satellite remote sensing data for assessment of land cover change in the rural-urban fringe. Inherent in this assessment process was the interpretation of multispectral data collected by several medium resolution satellite systems and evaluation of the quality of the resulting change information. Each dataset was acquired for a single date and classified at two levels of detail using standard classification algorithms. The optimum classification approach for each date was identified and the changes in land cover evaluated in several ways. The contribution of spatial and thematic errors and their propagation through the analysis process was investigated.Data for this research were acquired over an area approximately 4.5 km square located in the southern metropolitan area of Perth, Western Australia. At the time of the initial data acquisition in 1972 the area was predominantly rural and comprised mostly dense pine plantations, however by the final stages of data acquisition in 1991, the area was almost completely given over to urban residential land use. Changes were interpreted from classified Landsat Multispectral Scanner (MSS), Landsat Thematic Mapper (TM) and SPOT (System Pour l'Observation de la Terre) High Resolution Visible (HRV) multispectral data, and were compared to reference maps compiled from medium scale aerial photographs. The geometric properties of high resolution panchromatic IRS1-D data were also evaluated to test the geometric potential of high resolution satellite data.Supervised and unsupervised classification algorithms were used for derivation of land cover maps from each multispectral dataset at two levels of detail. Data were classified onto four general levels at the broadest (Level I) classification, and into nine levels at the finest (Level II) classification. The Kappa statistic and its variance were used to determine the optimum classification approach for each dataset and at each level of detail. No significant differences were observed between classification techniques at Level I, however at Level II the supervised classification approach produced significantly better results for the Landsat TM and SPOT HRV data. Classification at the more general Level I did not produce substantially higher classification rates compared to the same data at Level II. Additionally, higher spatial resolution data did not provide increased accuracy, however this was due mainly to a much greater complexity of land covers present at the time the higher resolution Landsat TM and SPOT HRV data were recorded.Land cover changes were assessed separately at Level I for all datasets, and also between Landsat TM and SPOT HRV data at Level II. Integrated multiscale assessment of land cover change was undertaken using classified Landsat MSS data at Level I and Landsat TM data at Level 11. This enabled the continuity of change to be established across classification levels and sensor systems, even though there were variations in the level of detail extracted from each image.The sources of spatial and thematic errors in the data were investigated and their effects on change assessment analysed. The evaluation of high resolution panchromatic satellite data emphasised the contribution to the analysis of spatial errors contained within the reference data. The multiscale data also indicated that combined propagation of spatial and thematic errors requires investigation using appropriate simulation modelling to establish the influence of data uncertainty on classification and change assessment results.This research provides useful results for demonstrating a process for the integration of information derived from remotely sensed data at different measurement scales. Availability of data from an increasing range of remote sensing platforms and uncertainty of long term data availability emphasises the need to develop flexible interpretation and analysis approaches. This research adds value to the existing data archive by demonstrating how historical data may be integrated regardless of the spectral and spatial characteristics of the sensors.