Extraction of tea plantation with high resolution Gaofen-2 image

Abstract

Tea is the most popular drink in China. The spatial distribution information of tea plantation is useful for local government management. Lantian Country, with an area of 99.77km2, located in the midwest of Anxi County, which is famous for Oolong Tea, was chosen as study area, and image from Chinese high resolution satellite Gaofen-2 acquired on Jan 22, 2015 was used to study the method of tea plantations extraction. In order to construct best features for classification, optimum index factor (OIF) were firstly calculated on different original spectral bands combinations and the one with max OIF was chosen. Secondly, spectral enhancement was carried on multi-spectral bands.Difference between two vegetation indexes, namely, normalized difference vegetation index and modified normalized difference vegetation index was calculated and named as DNDVI. In DNDVI image, the brightness difference between tea plantation and background was improved and shadowed area in either index image was reduced. Thirdly, gray level co-occurrence matrix (GLCM), Gabor filter, local binary patterns (LBP) extraction, and method combined LBP and Gabor was carried on pan image to construct texture features. Among eight common features based GLCM, contrast, dissimilarity, entropy, variance, tea plantation area was darker. In homogeneity and angular second moment, this phenomena is just the opposite. In mean and correlation, there was no obvious difference between target tea plantation and background. So the gray level co-occurrence texture (GLCT) subtract sum of second two features from sum of the first four feature was used as final GLCM feature, and window size for GLCM set to be 15 was preferred. Multi-scale and multidirectional Gabor texture with max frequency set to be 1HZ was derived. For LBP, the operator LBP16, 2 with rotation invariance was tested to be the best. Finally, five schemes combine these spectral and textural features as inputs of classifier were evaluated in term of classification accuracy. Six categories including tea plantation, forest, roads, water, build-up, bare soil, shadows were classified by support vector machine. The result showed that overall accuracy range from 75.55% to 89.11%, Kappa coefficient range from 0.613 to 0.843, for plantation, user accuracy range from 84.95% to 100%, producer accuracy range from 53.29% to 91.53%. Gaofen-2 show its capacity to map the tea plantation area accurately. Schemes utilized spectral and textural features together perform much better than that utilized spectral only. The scheme combination of band1, band 3, ban4, DNDVI, LBP_Gabor outperformed other Scheme, with the highest overall accuracy and Kappa coefficient. The textures feature of high resolution image helps to improve the accuracy, and the way to construct suitable texture feature and merge different texture feature deserved study more. The proposed method to extract tea plantation is applicable at administrative level of country.