融合多光谱成像与深度学习的作物植株叶绿素检测系统研究

为了满足田间作物长势快速检测与指导变量管理的需求，以玉米为例设计了基于多光谱成像的田间作物植株叶绿素检测系统，包括可见光(RGB)和近红外(Near-infrared, NIR)图像采集模块、主控处理器模块、模型加速模块、显示及电源模块，用于实现玉米植株智能识别与叶绿素指标一体化检测。首先，采集玉米苗期和拔节期冠层图像数据集，比较了植株冠层实例分割与株心目标检测两种深度学习模型，构建了基于MobileDet+SSDLite(Single shot multibox detector lite)轻量化网络的玉米植株定位检测模型，实现玉米植株识别。其次，提取被识别的植株株心RGB-NIR图像，开展RGB和NIR图像匹配与分割，提取R、G、B和NIR灰度值计算植被指数，使用SPXY算法(Sample set portioning based on joint X-Y distances)和连续投影算法(Successive projections algorithm, SPA)分别对数据集进行样本划分及特征变量筛选，选择高斯过程回归(Gaussian process regression, ...

Fusing Multispectral Imaging and Deep Learning in Plant Chlorophyll Index Detection System

In order to meet the needs of rapid detection of field crop growth and guiding variable management, a field crop chlorophyll intelligent detection system based on multi-spectral imaging was designed and developed with maize as an example. It included visible light (RGB) and near-infrared (NIR) image acquisition module, main control processor module, model acceleration module, display and power module, which was used to realize intelligent identification of corn plants and integrated detection of chlorophyll index. Firstly, the canopy image data set of maize seedling stage and jointing stage were collected, and two deep learning models of plant canopy instance segmentation and plant center target detection were compared. A corn plant location detection model based on MobileDet+SSDLite (single shot multibox detector lite) lightweight network was constructed to realize corn plant identification. Secondly, the identified plant heart RGB-NIR images were extracted, the matching and segmentation of RGB and NIR images were carried out, and the gray values of R, G, B and NIR were extracted to calculate the vegetation index. SPXY algorithm (sample set portioning based on joint X-Y distances) and SPA (successive projections algorithm) were used. The samples of the dataset were divided and the characteristic variables were screened, and then GPR (Gaussian process regression) algorithm was selected to establish the chlorophyll index detection model. The results showed that the recognition rate of the model reached 88.7% in the complex environment of occlusion overlap, and the recognition accuracy reached more than 90% in the non-overlapping environment. The model determination coefficient R2 of the modeling set of the chlorophyll content index detection model was 0.62, and the model determination coefficient R2 of the test set was 0.61. Field tests on the developed system showed that the detection rate of the system can reach 14.6 frames per second, and the average accuracy was 92.9%. The research results can effectively solve the problem of corn nutritional status detection in field environment, meeting the real-time detection requirements of field environment, and providing solutions and technical support for intelligent perception of crop production.