基于遗传算法优化BP神经网络的数字图像土壤含水量反演研究

Research on Soil Water Content Inversion of Digital Images Based on BP Neural Network Optimized by Genetic Algorithm

快速准确地测量土壤含水量在农业、水文、生态等领域的应用至关重要,数字图像技术测量土壤含水量因其具有廉价、快速和不破坏土体的优势成为研究的热点。基于手机相机获取的数字图像,提取了R、G、B、H、S、V和DN 7种图像特征参数,并利用与土壤含水量相关性较大的图像特征参数R、V和DN构建了BP神经网络和遗传算法优化的BP神经网络土壤含水量反演模型,来获取高精度的土壤含水量数值。结果显示:将BP神经网络和遗传算法用于数字图像技术,均可提高数字图像技术测量土壤含水量的精度,其中BP神经网络土壤含水量反演模型的决定系数(R^(2))可达到0.940~0.972,均方根误差(RMSE)为0.936%~1.694%;遗传算法优化的BP神经网络土壤含水量反演模型的R^(2)可达到0.976~0.993,RMSE为0.559%~0.878%,遗传算法优化的BP神经网络模型用于数字图像技术反演土壤含水量的R^(2)更接近1,RMSE更小,精度更高。同类型研究中提出的多元线性模型的R^(2)介于0.60和0.96,RMSE介于1.11%和7.00%,与其相比,研究模型的预测精度和稳定性更高,这些结果展示了BP神经网络和遗传算法在数字图像技术测土壤含水量的应用优势。此外,研究表明手机相机获取的数字图像经过处理后可用于室内条件下预测表层土壤含水量,后续研究还需在室外深层土壤中展开,以扩展研究模型的适用性。

Rapid and accurate measurement of soil water content is crucial for applications in agriculture,hydrology,ecology and more.Digital image technology for measuring soil water content has become a hot research topic because of its advantages of being inexpensive,fast and non-destructive soil.In this paper,R,G,B,H,S,V and DN seven image characteristic parameters are extracted based on the digital image obtained by mobile phone camera,and the soil water content inversion model based on BP neural network and BP neural network optimized by genetic algorithm is constructed by using the image characteristic parameters R,V and DN which have great correlation with soil water content to obtain high precision soil water content.The results show that both BP neural network and genetic algorithm can improve the accuracy of soil water content measurement by digital image technology,in which the coefficient of determination(R^(2))of the BP neural network soil water content inversion model can reach 0.940~0.972 and the root mean square error(RMSE)is 0.936%~1.694%;the genetic algorithm optimized BP neural network soil water content inversion model can reach R^(2) of 0.976~0.993 and RMSE of 0.559%~0.878%.The BP neural network model optimized by using genetic algorithm for digital image technique inversion of soil water content has R^(2) closer to 1,smaller RMSE and higher accuracy.The R^(2) of the multiple linear model proposed in the same type of research is between 0.60 and 0.96,and the RMSE is between 1.11%and 7.00%.Compared with the multiple linear model proposed in the same type of research,the prediction accuracy and stability of the model proposed in this study are higher.These results show the advantages of BP neural network and genetic algorithm in the application of digital image technology to measure soil moisture content.In addition,this study shows that the digital image obtained by the mobile phone camera can be used to predict the surface soil water content under indoor conditions.Subsequent studies need to be carried out in outdoor deep soil to expand the applicability of the model.