基于深度学习的薇甘菊图像识别模型构建与智能监测

Mikania micrantha Kunth is one of the most harmful invasive plants in China. It climbs on the phorophyte and hinders the growth of the phorophyte, and even lead the phorophyte to death. At present, the monitoring of Mikania micrantha Kunth is mainly depends on the manual inspection. It is not only difficult to accurately monitor the distribution of Mikania micrantha Kunth but also have low efficiency. In satellite remote sensing monitoring, the small and medium-sized Mikania micrantha Kunth as the error dot is overlooked. To precisely and rapidly monitor the Mikania micrantha Kunth in real-time, this project intends to use Unmanned Aerial Vehicle (UAV) remote sensing to acquire the ground hyperspectral images and geographical information of the monitoring area. The spectral analysis technology will be used to study hyperspectral specificity of Mikania micrantha Kunth and prove Mikania micrantha Kunth sensitive imaging spectral bands, which will solve the problem of the similarity between Mikania micrantha Kunth and phorophyte in visually. Meanwhile, for resisting the interference of the environmental factors and unknown background, the convolutional neural networks based on deep learning will be studied to precisely identify Mikania micrantha Kunth, phorophyte and other background. That will quickly and automatically identify the Mikania micrantha Kunth. At last, the identification results will be transformed to ecological landscape parameters. A quantitatively monitoring model will be studied, with the ecological landscape parameters as input, to evaluate the hazard ranking of Mikania micrantha Kunth. In addition, the results of the model and the geographical information will be used to generate the distribution heat maps of Mikania micrantha Kunth by GIS technique. It’s realized that Mikania micrantha Kunth will be precisely monitored in real-time. Our results will provide some theoretical and scientific basis for the development of intelligent monitoring and early warning system of Mikania micrantha Kunth and other invasive species. Besides, it will be beneficial to the intelligent monitoring technology for invasive plant.

针对重大入侵植物薇甘菊（Mikania micrantha Kunth）缺乏有效监测的实际需求，围绕“薇甘菊智能监测受相似颜色附主植物和未知环境因素干扰难以快速准确识别”这一关键难题，以野外薇甘菊无人机高光谱图像和对应地理信息为数据基础，重点研究：（1）薇甘菊高光谱特异性，探明视觉上最能区分薇甘菊与附主植物的成像敏感光谱波段；（2）针对敏感光谱波段下图像特征，探明环境参数和未知背景对薇甘菊图像识别率的干扰机制，基于深度学习构建抗干扰的图像识别模型，探明能快速准确识别薇甘菊的深度卷积神经网络模型结构和各功能函数；（3）构建薇甘菊生态景观参数与危害等级的定量模型，结合地理信息和GIS技术探明薇甘菊扩张趋势，并生成薇甘菊分布可视化热图。研究结果将实现复杂环境下薇甘菊实时精准监测及早期预警，丰富入侵植物智能监测理论，为利用无人机高光谱图像智能监测薇甘菊及其他外来入侵植物提供理论依据和技术支撑。

薇甘菊在我国南方各省广泛分布，严重破坏森林和农田的生态环境，并造成严重经济损失。对薇甘菊自动监测，进而防范薇甘菊进一步扩散，保护生物多样性迫在眉睫。薇甘菊与附主植物可见光下视觉辨识度低，分布无规律，野外环境复杂多变，一般检测方法很难实现对薇甘菊实时精准监测。本项目分别以薇甘菊的高光谱图像和高清彩色图像为对象，利用计算机视觉技术对野外薇甘菊的识别模型和监测方法进行了研究，具体研究内容如下：.（1）基于高光谱图像的薇甘菊光谱特异性研究：本项目提出了5种预处理方法，主成分分析方法和3个分类器，以不同的组合方式处理从野外采集到的138维薇甘菊高光谱数据。其中SG平滑、主成分分析和随机森林的组合具有88.71%的准确率、88.68%的平均准确率、0.7740的Kappa值和9.647 ms的总耗时。相比之下，SG平滑、主成分分析和支持向量机的组合具有稍低的准确率（84.68%）、平均准确率（84.66%）、Kappa值（0.6934）和稍短的总耗时（1.318 ms）。根据具体的识别精度要求和时间成本，SG-PCA-RF和SG-PCA-SVM是两种可能的野外识别薇甘菊的方法。.（2）基于高清彩色图像的薇甘菊识别方法研究：为满足野外薇甘菊及外来入侵植物大范围高精度监测的需求，采用无人机携带高清相机采集野外薇甘菊及外来入侵植物图像，构建了基于深度学习的卷积神经网络模型MmNet和IAPsNet分别对图像中的薇甘菊及外来入侵植物目标进行识别。经训练、验证和测试，MmNet对薇甘菊测试集400个样本的识别准确率高达94.50%，IAPsNet对外来入侵植物测试集893个样本的识别准确率93.39%。.（3）薇甘菊预警评估方法研究：基于识别结果应用模糊层次分析法以及盖度公式，构建薇甘菊预警评估方法，划分了五个薇甘菊入侵危害等级，可根据所需监测精度的不同，设置不同尺寸的网格和辐射半径R，绘制出薇甘菊入侵的精准分布热力图，能够清晰准确的体现不同区域的入侵薇甘菊的危害程度。.以上研究为薇甘菊及其他入侵植物自动监测技术的开发提供理论依据。