Soil organic matters (SOM), specifically carbon and nitrogen, bring numerous benefits to soil’s physical and chemical properties. In this paper, we employ spectral data obtained by Fourier transform near-infrared (FT-NIR) spectroscopy to predict the content of organic carbon (OC) and total nitrogen (TN) in mineral soils. To address the limitation generated by massive hyperparameters on convolution neural network (CNN), we substitute using a technique named SVD concatenation to learn features. The proposed model combines the layers of fully connected and regression to complete the prediction task. We abbreviate it as SVD-CNN, which is capable provide a multi-tasks output simultaneously. In experiments, we study the prediction performances of SVD-CNN on two datasets of FT-NIR and LUCAS 2009 topsoil. Based on different situations, the highest performance of R² achieves 0.8891 for OC and 0.9048 for TN on the FT-NIR dataset. Similarly, the most prominent results on the LUCAS 2009 topsoil dataset are R² = 0.9304, RMSE = 3.6014 for OC and R² = 0.9319, RMSE = 0.2733 for TN. Furthermore, we also evaluate the results obtained by solely using SVD concatenation, which reveals SVD-CNN performs a better generalization ability.

土壤有机质 (SOM)，特别是碳和氮，为土壤的物理和化学特性带来诸多好处。在本文中，我们利用傅里叶变换近红外 (FT-NIR) 光谱获得的光谱数据来预测矿质土壤中有机碳 (OC) 和总氮 (TN) 的含量。为了解决卷积神经网络 (CNN) 上由大量超参数产生的限制，我们使用一种名为 SVD 串联的技术来替代学习特征。所提出的模型结合了全连接层和回归层来完成预测任务。我们将其缩写为 SVD-CNN，它能够同时提供多任务输出。在实验中，我们研究了 SVD-CNN 在 FT-NIR 和 LUCAS 2009 表土两个数据集上的预测性能。根据不同情况，在 FT-NIR 数据集上，R ²对于 OC 达到 0.8891，对于 TN 达到 0.9048。同样，在 LUCAS 2009 表土数据集上最突出的结果是R ² = 0.9304，对于 OC ，RMSE = 3.6014，对于 TN ，R ² = 0.9319，RMSE = 0.2733。此外，我们还评估了仅使用 SVD 连接获得的结果，这表明 SVD-CNN 具有更好的泛化能力。