Research On Soil Classification Method Based On Deep Learning And Near Infrared Spectroscop

In order to improve the recognition rate of the soil rapid classification models,this paper takes the soil near-infrared spectral data released by Eurostat as the research object.At the same time,this study focuses on the following two works in the classification of soil cover and soil texture based on near-infrared spectroscopy using deep learning technology from two different soil classification levels through two different physical and chemical value data.(1)A near-infrared spectral soil covered classification model fused with different convolution scales is established.In this study,a Convolutional Neural Network(CNN)with a single size convolution kernel and a convolution fusion network soil covered classification model fused with multiple size convolution kernels were established based on the soil near-infrared spectral big data and the Short-Time Fourier Transform(STFT)preprocessing method.The rapid distinction between cultivated land,forest land and grassland is realized.In order to meet the requirements of two-dimensional convolution,the one-dimensional spectrum is subjected to Short-Time Fourier Transform to extract the spectral information of the spectral data.In order to avoid the vanishing gradient,the network adopts Re LU function and Batch Normalization(BN).In order to prevent the model from overfitting,Dropout combined with Early Stopping method is used to train the network.First,the study discusses the influence of different STFT window lengths and different convolution kernel sizes on the model classification effect.The experimental results show that when the STFT window length is 100 and the window overlap length is 50%,the overall classification accuracy of the model is the highest.The CNN model with smaller size convolution kernel has a higher classification accuracy,and the overall classification accuracy of the 3×3 size model reaches 78.76%.Models with different sizes of convolution kernels are all good at classifying a certain soil type.The 3×3 size model has the best effect on the classification of cultivated land;The 5×5 size model has the best effect on the forest land classification;The 7×7size model has the best effect on the grassland classification.Secondly,a Fusion-CNN model based on the fusion of convolution kernels of various sizes is designed.The model combines the classification advantages of different size convolution kernels,and the classification accuracy of the three types of soil has been improved to varying degrees.The overall classification accuracy of the model reached 84.39%.The Fusion-CNN model overcomes the shortcomings of the CNN model with a single-size convolution kernel,which has a long selection cycle for the appropriate convolution kernel size and cumbersome parameter adjustment steps,and can simplify and speed up the modeling process.Using the Fusion-CNN convolution fusion network can more effectively automatically extract the internal feature information of the soil near-infrared spectrum,so as to obtain a high and stable soil classification accuracy.(2)A soil texture classification model was established.In order to improve the accuracy of the soil texture classification model,this study used Eurostat’s 17,939 soil near-infrared spectral data to train the model.The performance differences of the Convolutional Neural Network and Long-Short-Term Memory(LSTM)network classification models were compared.And the rapid distinction between sand,loam,clay loam and clay were achieved.This study adopts Re LU,Batch Normalization,Dropout,Early Stopping method and other strategies to optimize the model.This study explores the influence of the number of network layers,network types(CNN and LSTM),and attention mechanism on the classification effect.The experimental results show that the overall classification accuracy of the model increases with the increase of the number of network layers.The accuracy of the 4-layer CNN model reaches 76.58%,and the 4-layer LSTM model reaches77.86%.Both types of models can effectively classify soil texture.After integrating the Squeezeand-Excitation Networks(SENet)attention module,the model can redistribute weights to extract spectral features more efficiently,resulting in higher and stable classification accuracy.The accuracy rate of CNN＿4＿Attention model reaches 77.50%,and the accuracy rate of LSTM＿4＿Attention model reaches 78.39%.