基于图谱融合的土壤营养成分高光谱检测方法

Hyperspectral Detection Method of Soil Nutrients Based on Map Fusion

摘要:

针对土壤养分含量的准确、快速检测需求，提出了一种基于图谱特征结合的深度学习检测方法。首先，使用高光谱仪拍摄土壤高光谱数据，并人工化学方法测定土壤碱解氮(N)、有效磷(P)、速效钾(K)、氢离子(H?)、有机质(OM)等5种物质含量。然后，采用偏最小二乘法(PLS)和随机森林算法(RF)建立预测模型，基于竞争性自适应重加权采样(CARS)、变量组合群体分析(VCPA)和最小角回归(LARS)3种算法筛选出共计17个特征波段。最后使用DeepLab v3 + 网络分割彩色图像并提取CNN特征，与特征波段光谱反射率的FCNN特征一起，构建了Transformer预测模型。实验结果表明，5种物质含量的检测准确率分别达到91.9%、82.4%、91.3%、97.7%和92.3%。该方法使用少量波段即实现了全面精确的土壤养分含量预测，为开发低成本、高便携的检测装置奠定了理论基础。

Abstract:

A deep learning detection method integrating graphical features is proposed to address the demand for accurate and rapid monitoring of soil nutrient content. Initially, hyperspectral data of soil samples were collected using a hyperspectrometer, followed by laboratory-based chemical measurements of five key soil properties: alkaline hydrolyzable nitrogen (N), available phosphorus (P), available potassium (K), H?, and organic matter (OM). Prediction models were then developed using partial least squares (PLS) and random forest (RF) algorithms. Through feature selection techniques- including competitive adaptive reweighted sampling (CARS), variable combination population analysis (VCPA), and least angle regression (LARS)- a total of 17 characteristic spectral bands were identified. Subsequently, a Transformer-based prediction model was constructed by combining deep visual features and spectral reflectance features. Specifically, color images of soil samples were segmented using the DeepLab v3 + network to extract convolutional neural network (CNN) features, which were then fused with fully connected neural network (FCNN) features derived from the reflectance values of the selected characteristic bands. Experimental results demonstrate that the detection accuracies for the five soil parameters reached 91.9%, 82.4%, 91.3%, 97.7%, and 92.3%, respectively. The proposed method achieves comprehensive and accurate prediction of soil nutrient content with a limited number of spectral bands, thereby providing a theoretical foundation for the future development of low-cost, portable, and efficient soil nutrient detection devices.