基于多模态信息融合特征识别的宁夏地区秸秆焚烧监测算法研究

It has aroused wide concern of that the behavior of straw burning during agricultural production has become an important source of environmental pollution. The traditional monitoring methods based on remote sensing are limited by climatic conditions, burning scale, duration and other restrictions. It is unable to carry out real-time monitoring of straw burning, which has become the bottleneck of technology development. Therefore, in this project the UVA altitude aerial video and multi-sensor information were collected as analysis objects. The key technology which affecting the final recognition results were be studied, and a series of algorithms were be explored, include: algorithm of video image key frame extraction and enhancement based on sparse principal component analysis and clustering, fusion algorithm of video and multi-sensor characteristic value matching based on image segmentation and locating, feature recognition algorithm of multiplex mode information based on support vector and probabilistic neural network modal. In order to achieve a reliable, convenient and accurate straw.burning aerial multimodal information fusion feature recognition of straw burning monitoring system, the relevant algorithm hardware design implementation was based on field programmable logic devices.The research results will provides the theoretical and experimental basis to straw burning remote monitoring based on multi-sensor fusion algorithm, improve the level of regional agricultural production and technology of electronic information industry and accelerate region modern agricultural development. It has great significance to the development of.local science, technology, economic and social.

秸秆焚烧行为已成为环境污染的一个重要来源，引起了广泛关注。基于卫星多光谱遥感监测技术受气候条件、焚烧规模、持续时间等条件限制，难以准确开展实时监测，成为制约秸秆焚烧监测技术发展的瓶颈性技术难题。为此，本项目将以无人机高空航拍视频及多传感器采集信息为对象，探索基于稀疏主成分分析聚类的关键帧提取和增强技术、基于优化粒子群图像分割定位的不同维度和分辨率的多传感器信息自适应聚类配准融合算法、基于支持向量机及概率神经网络的多模态融合特征识别算法，以及算法基于FPGA的硬件平台实现。上述研究将弥补秸秆焚烧监控体系上的短板，拓展无人机技术的应用，解决基于高空航拍的多传感器秸秆焚烧监测的共性难题，实现基于无人机的大区域复杂环境下多模态信息融合特征识别的秸秆焚烧监测系统。项目研究成果将进一步提高宁夏现代农业及信息产业技术水平、对加速地区农业发展提供理论和实践基础，对宁夏地区可持续发展具有重要意义。

针对基于卫星多光谱遥感监测秸秆焚烧技术存在的技术难题，以无人机高空航拍视频及多传感器采集信息为对象，重点研究了：基于视觉的秸秆焚烧图像分割、基于K-Means聚类及样本熵的秸秆焚烧检测算法以及基于轻量化YOLO及模型压缩的秸秆焚烧检测算法，构建了服务器端和机载嵌入式端的轻量化部署和基于FPGA端的硬件实现，实现了每秒25帧以上的实时监测，监测的准确度在99.35%；研究了数据传输和无人机路程规划，改进的势场蚁群算法。引入人工势场法影响下的信息素,利用势场启发信息加快蚁群算法的收敛速度,利用势场中净力的衰减系数求解局部最优问题，提高了收敛搜索时间44.83%；针对可见光和红外秸秆焚烧图像融合时图像难以完全匹配的问题，提出一种端到端的非匹配红外/可见光秸秆焚烧图像融合的生成对抗网络模型。秸秆焚烧融合模型生成的红外图像比预训练模型的CC提高了2.7%，EN、SD、SF分别提高了4.8%、2.5%、2.8%，运行效率提高了53.3%，同时实现了良好的主观图像效果；针对现有远程采集系统存在的无线传输容易受噪声干扰问题，提出了一种无线的智能去噪系统。实验结果表明：该系统能够实现信号的无线远程采集，消除信号中的环境噪声，具有较好的采集效果。整个研究形成了软硬件系统，探索了相关领域的关键技术，形成了理论研究基础，搭建了实践研究平台，能够提高秸秆焚烧的智能化监测水平，具有较好的应用前景。.项目团队包括韦海成、张白、肖明霞以及研究生团队。项目经费严格按照要求执行，无经费外拨情况。在基金项目的支持下，完善移动嵌入式硬件的深度学习平台1个，实现基于FPGA的深度学习硬件平台1个，发表中文核心期刊以上论文17篇（全部基金标注），其中SCI检索10篇，中文核心及CSCD检索3篇，ISTP检索4篇；参加会议3次，主场报告1次，分组报告4次，发表会议论文4篇；授权发明专利2项；在项目执行期间以项目团队骨干为核心，获批自治区科技创新团队1个，培养硕士毕业生5人。对照项目任务书要求，本项目已经顺利完成了相关研究工作。

内容获取失败，请点击重试