Forest mapping from multi‐source satellite data using neural network classifiers—an experiment in Portugal

Abstract

The monitoring of forest ecosystems and their biodiversity is a major challenge for satellite remote sensing. The opportunity now exists for operationally using multi‐sensor high resolution satellite data from optical, infrared and microwave parts of the electromagnetic spectrum. An experiment is reported on the integrated use of Landsat Thematic Mapper (TM) imagery combined with ERS‐1 Synthetic Aperture Radar (SAR) imagery for mapping eight identifiable types of forest ecosystem in a test area of approximately 100 x 100 km in western central Portugal. The TM and SAR images were geometrically co‐registered and resampled to a spatial resolution of 25 m. The SAR data were despeckled using a multiple regression method based on the TM scene. The classification of the integrated multi‐sensor dataset was carried out using multi‐layer perceptron neural networks which were trained using ground truth samples taken from test sites within the experimental area. The neural network was used in two separate mapping stages: the first to map nine basic classes of land cover including one of forest, and the second to map eight different types of forest within the generic forest land cover area. The total classification accuracy achieved for the eight forest classes was 77.7% from the integrated TM and SAR dataset, which was 1.7% higher than the accuracy achieved using the TM data alone. Whilst the overall classification accuracy did not improve significantly through the use of SAR data in addition to TM, there were some significant gains for certain individual classes. In particular, the addition of the SAR channel gave a 13.1% improvement in the discrimination of a mixed conifer/broad‐leaf class largely by giving better separation from a pure conifer class.