AUTOMATED CLASSIFICATION OF NATURAL FORESTS WITH LANDSAT TIME SERIES USING SIMPLIFIED SPECTRAL PATTERNS
- Institute of Geography, Vietnam Academy of Science and Technology, Vietnam
Keywords: Natural Forest, Landsat, Time Series, Simplified Spectral Patterns, Automated Classification
Abstract. Natural forests are a basic component of the earth ecology. It is essential for biodiversity, hydrological cycle regulation and environmental protection. Globally, natural forests are gradually degraded and reduced due to timber logging, conversion to cropland, production forest, commodity trees, and infrastructure development. Decreasing of natural forests results in loss of valuable habitats, land degradation, soil erosion and imbalance of water cycle in regional scale. Thus operational monitoring natural forest cover change, therefore, has been in interest of scientists for long time. Forest cover mapping methods are divided to two groups: field-based survey and remotely sensed image data based techniques. The field-based methods are conventional and they have been used widely in forestry management practice. Satellite-image-based methods were developed since beginning of earth observation. These methods, except visual image interpretation, can be grouped to supervised and unsupervised classification that rely on various algorithm as statistical, clustering or artificial intelligence. However, there is little report about method, which can extract natural forests from generic forest cover. Over the last couple of decades, natural forests have been over-exploited by various reasons. This practice led to urgent need of development of fast, reliable and automated method for mapping natural forests. In this study, a new method for mapping of natural forest by Landsat time series is presented. The new method is fully automated. It uses spectral patterns as principal classifier to recognize land cover classes. The proposed method was applied in study area consisted of Ratanakiri of Cambodia, Attapeu of Laos and Kon Tum of Vietnam. About 2000 Landsat images were used to generate land cover maps of the study area across years from 1989 to 2018.