Tea Garden Scene Detection From High-Resolution Remote Sensing Imagery

Tea is one of the most famous beverages in the world,and China,as the origin of tea,is not only the largest consumer but also the largest producer of tea in the world.Tea cultivation is one of the pillar industries of the Chinese agriculture,and has a great impact on the economic development of rural areas.Thus,it is of great significance to monitor and assess the tea gardens.However,tea garden monitoring is usually achieved by fieldwork,which is labor-and time-intensive.Meanwhile,remote sensing technology,an effective and convenient tool,has been widely employed to detect and map crops,but few studies have been so far carried out for tea garden mapping.This paper proposed a framework to detect tea gardens from high-resolution remote sensing imagery,which provides the foundation for the long time series monitoring and large scale assessment of tea garden.In high-resolution imagery,tea gardens are composed of multiple object types,and the tea bushes are generally planted in a unique way.In this regard,a tea garden can be considered as a semantic scene rather than a single land-cover type.Therefore,the methods in this tea garden detection framework are scene-based,including bag-of-visual-words(BOVW)model,supervised latent Dirichlet allocation(sLDA)model,and unsupervised convolutional neural network(UCNN).These scene-based methods can develop direct and holistic semantic representations for tea garden scenes,thus they are more suitable than the traditional pixel-based or object-based methods,which just focus on the local characteristics of pixels or objects.The task of this study is to effectively apply three scene-based interpretation models to tea garden detection,and carefully analyze their results.The main research content of this paper is as follows:(1)Tea garden detection based on BOVW model and sLDA model.This study employed the spectral features and Gabor textural features as low-level features to extracted BOVW features from scenes,then the tea garden detection was finished by the support vector machine.On the basis of the BOVW representation,sLDA model further extracted the topic features of the scenes,and utilized these topic features to detect tea gardens.(2)Tea garden detection based on UCNN.UCNN is a data-driven feature descriptor,and does not rely on the establishment of empirically designed features,which is the limitation of the two aforementioned models.In this paper,a two-layers UCNN was constructed to automatically mine multi-layer structure features from unlabeled scenes,and these features were also fed into a support vector machine to detect tea gardens.(3)The three tea garden detection methods were tested in four high-resolution remote sensing images.In the experiments,the optimal Kappa values of each dataset exceed 0.88 and are all obtained by the UCNN.These results confirmed the satisfactory performance of the proposed scene-based framework for tea garden detection,and UCNN outperformed other methods.