Open access publications in the SLU publication database

Abstract

The objective of this study is to apply ENVISAT MERIS data in mapping mountain vegetation in Sweden. The Swedish mountain vegetation is characterized by mosaics of different land cover types; a single MERIS pixel (300 meter IFOV) can consist of several of these different land cover types. “Hard” classifications which produce a single thematic class per pixel often give a low accuracy. While many different unmixing methods are reviewed in the literature, the use of regression trees is reported to be more promising than, for example, Linear Spectral Mixture Analysis. Regression trees handle non-linear data and are non-parametric, and can be well-suited for sub-pixel vegetation fraction estimation. Here, the soft classification methods of regression trees and linear regression are applied using spectral data from a MERIS Level 1B FR image. The image is corrected for atmosphere and illumination, and MTCI and PCA are calculated. Nine-hundred training plots are used for seven major vegetation classes. Preliminary results show that regression trees produce a slightly lower overall RMSE (20.1%) than linear regression (20.6%), although generally slightly higher class-wise biases. Results are promising however, and further improvements will be pursued.