基于多视角图像形态颜色纹理特征融合的生物量获取

Plants Biomass Acquisition Based on Morphological, Color and Texture Features of Multi-view Visible Images

摘要:

可见光成像以其快速、经济和非破坏性等优势，正成为高通量植物表型和遗传研究的有效工具，但仍有待解决基于可见光图像评估肉眼不可见的产量表型特性。本文针对植物叶片遮挡重叠及变量尺度单一导致图像数据精度受限的问题，提出了一种利用多视角图像融合多类别特征评估高粱地上生物量的技术方法。对15个种质基因的300株高粱进行了双因素（水分和养分）双水平（高和低）试验。基于旋转平台，利用可见光相机对每株高粱等角度间隔自动采集10幅侧视图像和1幅俯视图像，通过植物掩膜图像提取每株高粱形态特征（俯视、侧视投影面积）、颜色特征（RGB像素值）与纹理特征（均值、协方差、同质性等），将多个视角下的信息平均化处理，并基于图像R、G、B像素值构建16个颜色植被指数。结果表明，相对于考虑单一类型变量和单视角下的图像信息，基于多视角平均化图像信息融合形态、纹理、颜色特征能显著增加对高粱地上生物量表型的获取能力。利用SVR、RF、BPNN算法融合21组优化图像数据变量构建高粱地上生物量回归模型，精度最高的RF算法模型测试集决定系数R2为0.881，均方根误差(RMSE)为60.714 g/m2，平均绝对误差(MAE)为42.364 g/m2。为进一步优化RF算法模型的参数，选取GA、GS、SSA对RF算法模型进行超参数寻优。结果表明，SSA-RF优化模型测试集R2提升至0.902，RMSE为48.706 g/m2，MAE为39.877 g/m2。基于多视角图像形态-颜色-纹理特征融合能从有限的信息中衍生得到更多有效信息用于估测高粱地上生物量，从而为高粱生长监控、胁迫检测、水肥精确施用和良种快速筛选提供理论依据和技术支持。

Abstract:

Visible light imaging is becoming an effective tool for high-throughput plant phenotyping and genetic research due to its advantages of rapidity, economy and non-destructiveness. However, the evaluation of yield phenotypic characteristics that are invisible to the naked eye based on visible light images remains to be solved. A technical method for evaluating sorghum aboveground biomass by fusing multi-class features with multi-view images was proposed to address the problem of limited image data accuracy due to overlapping plant leaf occlusion and single variable scale. A two-factor (water and nutrient) and two-level (high and low) experiment was conducted on 300 sorghum plants of 15 germplasm genes. Based on a rotating platform, totally ten side-view images and one top-view image were automatically collected at equal angles for each sorghum plant by using a visible light camera. The morphological characteristics (top-view and side-view projection area), color characteristics (RGB pixel values) and texture characteristics (mean, covariance, homogeneity, etc.) of each sorghum plant were extracted through plant mask images. The information from multiple perspectives was averaged, and 16 color vegetation indices were constructed based on the image R, G, and B pixel values. The results showed that compared with considering image information of a single type of variable and a single perspective, the fusion of morphological, texture and color features based on multi-perspective average image information can significantly increase the ability to obtain the aboveground biomass phenotype of sorghum. The SVR, RF and BPNN algorithms were used to fuse 21 sets of optimized image data variables to construct a regression model for aboveground biomass of sorghum. The RF algorithm model with the highest accuracy had a test set determination coefficient (R2) of 0.881, a root mean square error (RMSE) of 60.714 g/m2, and a mean absolute error (MAE) of 42.364 g/m2. In order to further optimize the parameters of the RF algorithm model, GA, GS and SSA were selected to optimize the hyperparameters of the RF algorithm model. The results showed that the test set R2 of the SSA-RF optimization model was increased to 0.902, the RMSE was 48.706 g/m2, and the MAE was 39.877 g/m2. Based on the fusion of multi-view image morphology, color and texture features, more effective information can be derived from limited information for estimating the aboveground biomass of sorghum, thereby providing a theoretical basis and technical support for sorghum growth monitoring, stress detection, precise application of water and fertilizer, and rapid screening of improved varieties.