Inversion Of Soil Nutrient Content Based On Improved CNN-Stacking Model

Soil nutrient content is an important criterion for judging the level of soil fertility,and soil nutrients are the main means of obtaining the nutrients required by crops between sowing and maturity,and an important component of farmland cultivation.Therefore,the use of convenient and efficient methods for accurate estimation of nutrient content in soil is important for the sustainable development of agriculture and soil resources.The gradual emergence of remote sensing technology has opened up new ideas for research.Remote sensing satellites have short operation cycles,capture large image areas,and are used in various fields.At present,many scholars have used remote sensing spectral information for inversion analysis of soil nutrient content,which has gradually become an important research direction.However,there are still many shortcomings in this method.On the one hand,the current research data sources are mostly hyperspectral images,however,due to the narrower channels and richer details,hyperspectral images are generally more expensive to obtain;on the other hand,because the multispectral images have fewer wavelength bands and contain little information,the methods used in the research based on multispectral data are mostly multiple linear regression or simple machine learning models,resulting in The generalization ability of the models is limited and overfitting is easily produced,which greatly limits the research on soil nutrient content inversion based on multispectral images.To solve the mentioned problems and limitations,this paper is based on the project of the National Natural Science Foundation of China(NSFC)"Research on the intelligent decision method of maize precision operation based on multi-source heterogeneous big data" and the scientific and technological research project of Jilin Provincial Education Department "Research on the remote sensing inversion method of soil organic matter in black soil area based on machine learning"."The research work is carried out to improve the prediction ability of soil nutrients and enhance the generalization ability of the inversion model.The research area of the experiment is Nongan County,Changchun City,Jilin Province,and the base model of the experiment is chosen as a convolutional neural network and Stacking integrated learning method in machine learning.The main contents of the study are as follows:(1)Construction of soil nutrient inversion dataset.Data such as geographic location and chemical detection results of soil sample sampling points in the study area were collected and organized,while Sentinel-2A remote sensing images of the studied area were downloaded from the ESA Copernicus Data Center.The collected data are also pre-processed for the preparation of subsequent experiments.(2)A remote sensing inversion method of soil organic matter with an improved CNN network is proposed.Given the problems that the processed data features contain noise and the onedimensional data contains less information than the two-dimensional data,the model is improved by combining the feature selection method and the optimized model structure,and then the optimized CNN network and other inversion models are compared and analyzed to obtain a more suitable soil organic matter inversion method.(3)Construction of the soil nutrient content inversion model based on optimized CNNStacking.The improved CNN model was used to address the problem of overfitting when the inversion of soil nutrients other than organic matter was performed.The model is further improved by adding the idea of split convolution,which reduces the number of layers of the model and reduces the risk of model overfitting while retaining the predictive capability of the model.The Stacking integrated learning model is also added before the model result output to further improve the model prediction accuracy through secondary feature extraction of the model and model prediction of the meta-trainer.The idea of combining the CNN network with the Stacking integrated learning model is determined according to the characteristics of the data set: firstly,the optimized CNN network is used for feature extraction of the input features,and the feature vector output from the last fully connected layer of the model is intercepted and fed into the Stacking integrated learning model together with the original target value as the input data.The model is trained independently on the input data by the base trainers such as SVR,RF,and k NN,and the output data is input into the meta-trainer by stacking,and the soil nutrient content is predicted by the meta-trainer.Based on the actual experiments,the prediction results of the model and the influencing factors are discussed and analyzed,and it is proved that the CNN-Stacking based soil nutrient content inversion model proposed in this paper can predict the content values of different soil nutrients,which provides a reliable theotical support for soil resource improvement and utilization,and has an important guiding role in realizing the sustainable development planning of land resources.