基于遥感特征参量和卷积神经网络迁移学习的茶园场景识别方法研究

In order to coordinate economic development of tea and ecological environment, optimize tea distribution and prohibit illegal reclamation, the area and spatial distribution of tea plantations need to be well known. However, the remote sensing survey of tea plantations is limited by spectral confusion, and most of them are judged by field investigation with visual interpretation, which is inefficient. This study proposes a method for tea plantations detection based on the spectral and texture features, according to the scene-oriented method. Firstly, the seasonal regularity of the tea canopy spectrum and the texture scale of the tea garden will be studied, from which the suitable temporal (or multi-temporal) and the scale of remote sensing images, the spectral and texture features with distinguishing ability will be determined. On this basis, instead of the three-color channels of the image, the artificial designed features will be used to train the convolution neural network. Finally, combined with support vector machine, the two-class detection of tea garden scenes will be completed.

为了更好地协调茶产业发展和生态环境，优化茶叶生产布局，严控茶园违法开垦，迫切需要准确掌握茶园的面积和空间分布。当前茶园调查受限于光谱混淆，多采用遥感影像目视解译与外业调查相结合的方法，效率低且时效性不高。本研究从实际需求和当前学科前沿出发，依据面向场景的图像分析新思维，探索遥感茶园场景识别的基本流程和关键技术。首先，开展茶树冠层光谱季相变化规律和纹理特征尺度优选研究，明确识别茶园适宜的遥感（多）时相（或茶树生长发育期）和影像空间尺度，构建具有区分能力的光谱和纹理特征参量，为茶园遥感识别在数据源和特征选择上提供参考依据；在此基础上，利用具有茶园识别能力的光谱和纹理特征参量替代传统图片三色通道，基于卷积神经网络迁移的非监督场景特征学习，结合支持向量机，构建多尺度茶园场景识别模型，有望有效提高茶园识别精度，同时为作物遥感识别构建了一种新的研究思路。

本项目针对国内茶园遥感识别机理研究缺乏，茶园分类难度较大这一问题开展茶园提取研究。项目执行过程中，完成了多时相茶园的高分辨率无人机图像采集，开展茶树冠层光谱季相变化规律和茶园纹理特征尺度优选研究，明确识别茶园适宜的遥感时相和空间尺度，构建具有区分能力的光谱和纹理特征参量，为茶园遥感识别在数据源和特征选择上提供参考依据。利用经人工设计的具有茶园识别能力的光谱和纹理特征参量替代传统的图像三色通道，建立茶园遥感识别样本库，基于FCN和LSTM深度学习框架搭建茶园识别模型，结果表明加入遥感特征能够有效提高茶园识别精度，但是这一精度距离应用预期仍然具有一定的差距，需要在样本库建设、遥感数据质量等方面做进一步加强。