Land Use/Cover Classification Of ALOS Imagery Assisted With Texture Information

The emergence of earth observation system, as intergrated with remote sensing , geographic information systems and global positioning systems, provides completely new scientific tools for land and resources surveys. Spectral characteristics and texture feature both can be used for land use/cover classification. Elements exhibit considerable spectral diversity at fine spatial scales with the presence of numerous spectrally unique materials as well as spectrally ambiguous materials. Traditional pixel-based classification methods, relying solely on the spectral information, are limited to deal with such problems and can not meet the growing need for remote sensing applications. As an important indicator of spatial structure information in remote sensed images, texture plays major role in accurate land cover mapping. Decision tree shows great advantage in land use/cover mapping, for sake of its flexibility, clear physical meaning, stability and high efficiency. It has been widely used in remote sensing classification.With a case study in Anji County, Zhejiang Province, this work tried to map land use/cover using ALOS (Advanced Land Observing Satellite) /AVNIR-2 (Advanced Visible and Near Infrared Radiometer type 2) images. GLCM (Gray Level Co-occurrence Matrix) was employed to extract the texture information from the first principal component and C5.0 decision tree algorithm, using both texture and spectral information, was used to map land use/cover. The final results were compared to that of maximum likelihood classification. The contributions of different indicators to classification accuracy were then analyzed.The main results were as follows:(1) By using DEM, slope map and Modified Soil adjusted Vegetation Index(MSAVI), and studying the land use/cover characteristics and topography, the study area was divided into hilly and mountainous region, and valley plains region. Also, two indices were derived: Normalized Difference Water Index (NDWI) and B23 (B23=(band2-band3) / (band2 + band3)). By integrating these two indices, improved accuracy for decision tree classification and maximum likelihood classification were obtained. MSAVI helped distinguish forest and water, but failed to play its due role in mapping agricultural land for sake of the winter season when the imagery was obtained. B23 increased the probability to distinguish construction sites from bare soils, and NDWI contributed to the increased accuracy in water mapping.(2) The use of GLCM texture extracted the following five texture characteristics: entropy, variance, mean, contrast and dissimilarity. Based on the texture features of typical land use/covers,a new parameter Mean/Contrast was introduced to enhance the texture features differentiation. Results showed that, all texture metrics except variance improved the classification accuracy, amorg which entropy and dissimilarity showed the most improvement. In maximum likelihood classification, entropy and dissimilarity raised the overall accuracy by 6.0% and 8.2%; for decision tree classification, the two metrics increased by 7.7% and 4.5%, respectively. The performance of Mean/Contrast was slightly better than the average. These results indicated that a reasonable combination of texture metrics could improve the classification accuracy.(3) Decision tree was superior to maximum likelihood method, in term of overall classification accuracy. Using entropy, the overall accuracy of decision tree classification was 84.5%, higher than maximum likelihood (77.3%); using MSAVI, (NDWI and B23), the overall accuracy of decision tree classification was 5.0% and 6.4% higher than the maximum likelihood. Compared to maximum likelihood classification,decision tree classification uses multi-source information as a decision-making knowledge, which helps enhance the efficiency of auxiliary information.