Satellite data are widely used within remote sensing to respond to the growing need for a deeper understanding of the Earth’s bio- and geophysical parameters. Applications, such as land cover classification has for long been an important task within the field. Optical satellite data have proven to be efficient tools, however, they are unavailable in some conditions, such as cloudy weather. This deficit can be addressed with synthetic aperture radars (SAR), and recently, improvements have been made in their spatial and temporal coverage. Furthermore, a fusion of these data takes advantage of their different characteristics and can lead to even improved outcomes. The aim of this study was to develop and implement an effective land cover classification approach for the boreal forest zone by using multi-temporal SAR and optical data. Optical and SAR satellite data were collected from the area around Hyytiälä, Finland. One Landsat 8 scene and a time series of Sentinel-1 data spanning over a year were used. Co- and cross-polarized data were available. A very high resolution (VHR) reference image was manually interpreted to form training and test data. Features were extracted from both data sets and those from the SAR data were reduced using feature selection. A land cover classification was then performed separately on each data set and with a fused data set. Different features were tested to find an optimal combination. The classifications were performed with the nearest neighbor rule and the maximum likelihood classifier. This resulted in several classification maps which were validated with the test plots.The results showed that the single-sensor classifications were noisy. Classifications with only optical imagery performed better. Additionally, removing some of the original data from the calculations, which can speed up the process, led to worse results. The multi-sensor classifications with the fused data improved the results significantly. Much of the noise was no longer present. The best classification was reached with a fused data set of four SAR features from VH polarized data and four optical features, which gained a final accuracy of 89.8 %. This classification was done with the maximum likelihood classifier. Accuracies up to 97.3 % were also reached but this result had clear flaws in the visual interpretation. It was concluded that fusing optical and SAR data for land cover classification in the boreal zone is a very promising strategy and should be investigated further to reach even better results.
展开▼
