In recent years, the spatial resolution of satellite data has improved with advancements in satellite technology, and acquisition of even more detailed information on the surface of the earth can be expected in the future. Various complications, however, are still associated with retrieval
and classification of ground surface information from very-high-resolution satellite data. One important issue is the occurrence of isolated pixels with high local spatial heterogeneity between neighbouring pixels in the classification result. The main objective of this research was to evaluate
the effectiveness of two Multiple Classifier System (MCS), which combine the results from different supervised classifiers, as a means for reducing isolated pixels when using high-spatial resolution satellite imagery. The research was conducted in the Tohoku region, where the Great East Japan
Earthquake of 11 March 2011 and subsequent tsunami have greatly changed the regional land use and land cover. Multiple input features (five RapidEye spectral channels, five spectral indices, digital elevation model, and slope) were prepared from satellite data and used for the machine learning
and validation. Ground truth data belonging to six major land cover classes (forest, shrub/grassland, cropland, urban area, waterbody, bare ground) were collected. Six machine learning classifiers; Random Forest (RF), Support Vector Machines (SVM), K-nearest Neighbours (KNN), XGBoost,
Bagging, and Neural Network (NNET), were individually tested for classification accuracy and generation of isolated pixels, and the results were compared with those of the MCS combinations. The XGBoost provided the highest overall accuracy (0.987) and kappa coefficient (0.985) of the individual
classifiers, but suffered from a large number of isolated pixels. The MCS by combination weighing the kappa coefficients (MCS kappa), however, produced equivalent overall accuracy (0.989) and kappa coefficient (0.987), but at the same time was able to significantly reduce the number of isolated
pixels. This reduction amounted to 30% for all land cover classes. In particular, the number of isolated pixels in the forest class was reduced by 50%. These results clearly show that the MCS kappa is capable of reducing isolated pixels in the classification of very-high-resolution land cover
information, without sacrificing classification accuracy. The authors hope that this system will prove effective for monitoring fine-scale changes in land cover, and contribute to conserving landscapes and ecosystem services in the region impacted by the earthquake and tsunami.
No Reference information available - sign in for access.
No Citation information available - sign in for access.
No Supplementary Data.
No Article Media
Document Type: Research Article
Graduate School of Tokyo University of Information Sciences, Wakaba-ku, Chiba, Japan
Department of Informatics, Tokyo University of Information Sciences, Wakaba-ku, Chiba, Japan
Publication date: April 3, 2019
More about this publication?