【摘要】 土壤肥力是衡量土壤质量的重要指标.为了评估土壤质量和提高作物产量，迫切需要找到快速预测土壤成分的途径.首先提出Inception层短时傅里叶变换卷积神经网络(inception short-time Fourier transform convolutional neural network,简称I-STFT-CNN)单一光谱模型，然后提出2个融合光谱模型II-STFT-CNN(indirect inception short-time Fourier transform convolutional neural network)和CI-STFT-CNN(cascade inception short-time Fourier transform convolutional neural network),最后对这些光谱模型的性能参数进行对比.研究结果表明：相对于SVR(support vector regression),PLSR(partial least squares regression)和STFT-CNN(short-time Fourier transform convolutional neural network)模型，该文提出的单一光谱I-STFT-CNN模型具有更高的预测精度；融合光谱模型的预测精度优于单一光谱模型；两个融合模型中，级联融合CI-STFT-CNN模型的性能优于通道融合II-STFT-CNN模型.因此，3种模型中，CI-STFT-CNN模型的预测精度最高.更多还原

【Abstract】 Soil fertility serves as a crucial indicator of soil quality. To enhance crop yield and evaluate soil quality, it is imperative to rapidly forecast soil makeup by discovering new approaches. Firstly, a single spectral model of inception layer short-time Fourier transform convolutional neural network(I-STFT-CNN) was proposed. Then, two fusion spectral models II-STFT-CNN(indirect inception short-time Fourier transform convolutional neural network) and CI-STFT-CNN(cascade inception short-time Fourier transform convolutional neural network) were proposed. Finally, the performance parameters of these spectral models were compared. The results showed that compared with SVR(support vector regression),PLSR(partial least squares regression) and STFT-CNN(short-time Fourier transform convolutional neural network) model, the single spectral I-STFT-CNN model proposed in this paper had higher prediction accuracy. The prediction accurary of fusion spectral model was better than that of single spectral model. Among the two fusion models, cascade fusion CI-STFT-CNN model had better performance than channel fusion II-STFT-CNN model. Therefore, among the three models, the CI-STFT-CNN model had the highest prediction accuracy.更多还原