Red Soil Organic Matter Content Prediction Model Basedon Dilated Convolutional Neural Network

Soil Organic Matter (SOM) content is one of the important indicators used to measure soil fertility, and it is of great significance in accurately predicting SOM content from hyperspectral remote sensing images. Traditional machine learning methods require complex feature engineering. Still, they are not highly accurate, while deep learning methods represented by Convolutional Neural Networks (CNNs) are less studied in soil hyperspectral, and the modeling accuracy of small sample data is poor. The spatial feature extraction of spectral data is insufficient. This paper proposes a one-dimensional convolutional network model using a channel attention mechanism (SE Dilated Convolutional Neural Network, SE-DCNN). Taking 207 soil samples collected from Guangxi State-owned Huangmian Forest Farm and State-owned Yachang Forest Farm as research objects, this paper compares and analyzes the modeling effects of 3 machine learning and 4 deep learning methods under different spectral preprocessing. The results show that the SE-DCNN model, because of the use of dilated convolution and channel attention mechanism, expands the receptive field, extracts multi-scale features, and has good modeling accuracy and generalization fitting ability. The best prediction model in this paper is the SE-DCNN model established based on the spectral preprocessing method of Savitaky-Golay denoising (SGD) and first-order derivative (DR), the determination coefficient (R) of the validation set is 0.971. the root mean square crror (RMSE) is 2.042 gkg, and the relative analysis error (RPD) is 5. 273, Therefore, SE- DCNN can accurately predict the organic matter content of red soil in Guangxi forest land