程清 张航 张承明 殷复伟 王程成

摘要：针对目前利用深度学习技术进行高分光学遥感图像分类方法研究中尚存在的不足，本文提出了一种以多时相遥感数据为数据源，面向农作物种植信息提取的分类算法。该算法首先获取农作物在若干典型生长时期的光学遥感图像并进行配准等预处理，然后建立了一种以像素为单位的数据组织结构，该结构包含不同生长时期的作物信息、纹理信息，能较好地解决现有分类研究中信息不足的问题；接着以前馈神经网络为基础，建立了一种以像素为单位的分类算法，最后以得到的逐像素分类结果为基础进行成图。与同类方法相比，本文提出的算法综合考虑了农作物在不同生长时期的特征，更能发挥深度学习技术的优势，且多时相数据在提高农作物提取信息精度方面具有明显优势。

关键词：遥感分类；多时相数据；信息提取；农作物；神经网络

中图分类号：S127文獻标识号：A文章编号：1001-4942（2018）04-0149-05

Abstract In view of the shortages in researching high resolution optical remote sensing image classification method by the deep learning technology， a classification algorithm for crop information extraction based on the multi-temporal remote sensing data was proposed in this paper. The algorithm firstly obtained the optical remote sensing images of crop at several typical growth stages， and preprocessed these images such as registration. Then it established data organization structure based on pixels to solve the problem of insufficient information in the existing classification researches， which contained crop information and texture information at different growth stages. And it proposed a pixel classification algorithm based on the feedforward neural network. Finally，it mapped images based on pixel by pixel classification results. Comparing to the previous methods， this method comprehensively considered the characteristics of crop at different growth stages，could give full play to the advantages of the deep learning technology and had obvious advantages in improving the precision of crop information extraction.

Keywords Remote sensing classification； Multi-temporal data； Information extraction； Crop； Neural network

获取准确的农作物种植种类、面积、空间分布等信息，对于加强农业生产管理和国家宏观调控、保障农业可持续发展并最终保障国家粮食安全具有重要的意义。

围绕该问题，研究者们已开展了很多研究，并取得一定成果。闫慧敏等[1]利用MODIC/EVI时间序列影像，分析了鄱阳湖农业区多熟种植时空格局特征；郑长春等[2]以黑龙江852农场为研究区域，利用 SPOT 影像基于简单决策树分类器提取了水稻、小麦和玉米三大作物组成的种植结构信息；……