基于图像和光谱特征的双分支卷积神经网络土壤有机质含量预测

摘要：土壤有机质是土壤养分的主要来源。土壤有机质含量的快速测定对指导田间管理具有重要意义。SOM含量的变化 会改变土壤的物理性质，如颜色和孔隙度，进而导致土壤在可见光和近红外范围内反射率的明显差异。在华北地区 5 个 试验田采集了 417 个表层土壤（5-15 cm）样品，获得了 350- 1700 nm 波段的光谱数据和采样点的图像数据。在二分支 CNN预测SOM含量的基础上，建立多CNN网络实现图像和光谱的特征融合。首先，构建了基于改进CNN的光谱特征 提取模块，利用Auto Encoder对768 ~ 140波段的光谱信息进行压缩，并利用CNN进行特征提取。 通过ResNet 50网络 实现土壤图像的特征提取，最后通过全连通层的双线性池化融合模块融合图像和光谱的特征信息，完成 SOM 含量预测 并输出结果。经过实验验证，BLMulti CNN在有机物模型预测中的准确率最高，其中预测集R2为0.910，比基于concat的 Multi-CNN高0.017，比AE-CNN的预测方法高0.075，比基于ResNet 50的图像预测模型高0.10。BLMulti-CNN有效提 高了模型的预测精度，为田间土壤有机质含量的快速检测提供了技术指导。

关键词：土壤有机质；高光谱；数据融合；双分支CNN。

期刊：Computers and Electronics in Agriculture

DOI：https://doi.org/10.1016/j.compag.2023.108561

设备：GaiaSorter

Soil organic matter content prediction based on two-branch convolutional neural network combining image and spectral features

Abstract：Soil organic matter (SOM) is the main source of soil nutrients. Rapid determination of SOM content is of great significance for guiding field management. The change of SOM content will change the physical properties of soil, such as color and porosity, and then lead to obvious differences in soil reflectance in the visible and near infrared range. In this study, 417 topsoil (5–15 cm) samples were collected from five experimental fields in North China, and spectral data in the range of 350–1700nm and image data of sampling points were obtained. Multi- CNN network was established to realize the feature fusion of image and spectrum based on two-branch CNN to predict SOM content. Firstly, the spectral feature extraction module based on the improved CNN was con structed, and the spectral information from 768 to 140 bands was compressed by using Auto Encoder, and the feature was extracted by CNN. The feature extraction of the soil image is realized by the ResNet50 network, and f inally the feature information of the image and the spectrum is fused by the bilinear pooling fusion module in the fully connected layer to complete the SOM content prediction and output the results. After experimental verification, BLMulti CNN had the highest accuracy in predicting organic matter models, in which the prediction set R2 is 0.910, 0.017 higher than that of the Multi-CNN based on concat, 0.075 higher than that of the prediction method of AE-CNN and 0.10 higher than that of the image prediction model based on ResNet50. The BLMulti- CNN effectively improves the prediction accuracy of the model and provides technical guidance for the rapid detection of SOM content in the field.

Keywords：Soil organic matter, Hyperspectral, Data fusion, Two-branch CNN.