基于众包数据和遗传规划算法的农作物遥感智能识别方法

Intelligent remote-sensing-based method for crop identification by crowdsource and genetic programming

关键词:  农作物遥感分类  众包  特征提取  遗传规划

Key words:  Crop remote sensing classification  crowdsourcing  feature extraction  genetic planning

摘要:

【目的】 农作物空间分布信息是支撑相关科学研究与政策制定的重要依据。当前农作物空间分布遥感分类在理论和技术方法方面取得了长足的发展，但仍面临一些难题，包括地面样本数据的获取困难、作物特征选择存在主观性和冗余、特征构建过程中缺乏针对性和代表性等，导致农作物空间分布遥感分类的效率与精度不足。【方法】 针对这些问题，文章开展快速、准确、低成本的样本获取、特定作物分类的最优特征构建与优选，并分别选择多个研究区开展实证研究。样本获取方面，开发基于“视田”众包的样本获取平台，通过迭代更新的任务采集和历史样本库的方式高效获取地面样本。作物分类方面，提出遗传规划算法为不同作物提取差异化的特征，通过遗传进化思想实现定制化特征的构建，能够在原始特征的基础上构建高层次特征。【结果】 在北方地区，利用“视田”众包工具，由8名工作人员2天内完成了位于义县、辽中区、新民市及开原市4个区域的水稻、玉米、大豆和花生的样本采集，内业工作人员同步进行分类并迭代样本需求，分类的总体精度均大于90%，kappa系数均高于0.87。在南方地区，位于湖北省枝江市区域的春秋两季作物分类结果的总体精度均大于94%，kappa系数均高于0.86。【结论】 该文提出了一套快速、高效开展农作物遥感分类的技术体系：利用众包采集快速扩大样本数量，同时利用遗传规划算法提高样本训练效率。在不同区域、不同作物类型研究区应用，可实时、准确生产农作物空间分布图，总体效果稳定，在支撑科学研究与政策制定方面具有较强的应用前景。

Abstract:

[Purpose] Crop spatial distribution information is the basic data for many relevant scientific researches and policy formulation. Crop remote sensing mapping has made great development in terms of theory and technical methods. However，there are still some difficulties in crop remote sensing classification，firstly，it is difficult to obtain ground sample data，secondly，there is subjectivity and redundancy in crop feature selection，and the feature construction process lacks relevance and representativeness.[Method] To address this problem，we selected different regions in the north and the south of China as research objects to investigate the rapid，accurate and low-cost sample acquisition，and the optimal feature construction and selection for specific crop classification. In terms of sample acquisition，we have developed a crowdsourcing platform based on “Shitian”，which can efficiently acquire ground samples through iterative update of task collection and historical sample library. For crop classification，the proposed genetic planning algorithm can extract differentiated features for different crops and build customized features through the idea of genetic evolution，which can build high level features based on the original features.[Result] In the northern region，8 staff members completed the collection of samples of rice，corn，soybeans and peanuts in 4 areas of Yixian，Liaozhong District，Xinmin City and Kaiyuan City within 2 days，and the overall precision of crop classification in each region was greater than 90%，with kappa coefficients higher than 0.87. In the southern region，the overall precision of crop classification results for fall 2020 and spring 2021 in the study region located in Hubei was greater than 94%，with kappa coefficients were both higher than 0.86.[Conclusion] This paper proposes a rapid and efficient technical system for crop remote sensing classification： using crowdsourcing to rapidly expand the number of samples，and using genetic programming algorithms to improve the efficiency of sample training. Applied in different regions and research areas of different crop types，the spatial distribution map of crops can be produced in real time and accurately with stable effect. It has strong application prospects in supporting scientific research and policy formulation.