Research On The Method Of Near-infrared Hyperspectral Detection And Model Transfer Of Soil Nitrogen

The detection of soil nitrogen content is the key basis for decision-making of nitrogen nutrition management in farmland,and is of great significance to the agricultural ecosystem and even human health.There are more than 100 kinds of compounds in the soil that can provide nitrogen nutrition for plants,which can be classified into three categories roughly,namely nitrate nitrogen（NO₃-N）,ammonium nitrogen（NH₄-N）and amide nitrogen.The traditional detection method is mainly based on chemical analysis,which has the disadvantages of high cost,long period,low safety,and pollution risks.Near-infrared hyperspectral imaging（NIR-HSI）analysis technology provides a new method for the rapid determination of soil nitrogen content due to its advantages of non-destructiveness,high efficiency and sensitivity to nitrogen components.In order to improve the accuracy of rapid determination of soil nitrogen content and meet the needs of soil detection in different regions without pretreatment,this study used soils in different regions and three different types of nitrogen fertilizers,ammonium bicarbonate（NH₄-N）,sodium nitrate（NO₃-N）and urea（urea-N）as the detection objects,the detection method and model transfer method of nitrogen in soils based on NIR spectroscopy were studied in depth.The main research contents in this study were as follows:（1）High-precision quantitative detection of different types of nitrogen fertilizers in different soils is achieved based on NIR-HSI technology combined with machine learning algorithms.For three types of nitrogen,NH₄-N,NO₃-N and urea-N,the collected spectral data were modeled by combining data preprocessing methods,machine learning modeling methods,and characteristic wavelength selection algorithms.The average performance index and result distribution after 100 runs of the model were analyzed,and the results showed that:（i）Using the characteristic wavelength selection algorithm greatly reduced the spectral dimension required for modeling,and the prediction results were improved to a certain extent compared with the full-spectrum model;（ii）The optimal prediction results of the models for the three nitrogen fertilizers were:for NH₄-N,the average R_P² was 0.92,the average RMSE_P was0.77%,and the average RPD was 3.63;for NO₃-N,the average R_P² was 0.92,the average RMSE_P was 0.74%,and the average RPD was 4.17;for urea-N,the average R_P² was 0.96,the average RMSE_P was 0.57%,and the average RPD was 5.24.（2）Using the spectral correction algorithm and the transfer learning algorithm,the transfer of the nitrogen detection model from the tableted soil to the soil without tableting was realized.Based on different numbers of standard samples,JDA,Easy TL,DS-JDA and DS-Easy TL models were constructed and evaluated on the 6 transfer tasks.The results showed that:（i）The DS method was used to linearly transform the spectral features of the target domain,and the accuracy of the transfer learning models were significantly improved compared with the direct transfer models;（ii）The influence of the standard sample numbers on the accuracy of the JDA and Easy TL models in the target domain was different.The JDA model had the best transfer accuracy when the standard sample size was 40（33%of the total dataset）,and the average prediction accuracy was about 0.81;The Easy TL model had the best transfer accuracy when the standard sample size was 30（25%of the total dataset）,with an average prediction accuracy of about 0.63;（iii）The comprehensive performance of the above models was:DS-JDA>DS-Easy TL>JDA>Easy TL.（3）Combined with the transfer learning feature transformation algorithm and traditional machine learning algorithm,the transfer detection of urea nitrogen content between soils in different regions was basically realized.The differences in the external and internal characteristics of the six soils indicated by the NIR reflectance spectra and the t-SNE dimensionality reduction visualization were analyzed,and the prediction results of the PLSR model direct transfer and the TCA-PLSR transfer learning model in 30 transfer tasks were compared.The results showed:（i）The problem of model failure（R_P²<0.5）occurred in most of the transfer tasks when the PLSR model was directly transferred;（ii）The effectiveness of the TCA-PLSR（200-dimensional）model and TCA-PLSR（30-dimensional）on 30 transfer tasks were both 50%,but the TCA-PLSR（30-dimensional）model further improved the prediction accuracy,and the best performance was:R_P²=0.91,RMSEP=0.03%,RPD=3.4;（iii）The comprehensive performance of the above models was:TCA-PLSR（30 dimensions）>TCA-PLSR（200dimensions）>PLSR（200 dimensions）.In this paper,methods to improve the detection accuracy of different types of nitrogen in soil were studied,and transfer methods of nitrogen content detection models between soils with different physical forms and between soils in different regions was proposed.The research results showed the great advantages of NIR-HSI technology in soil nitrogen detection,and provided a theoretical basis and guiding scheme for the development of rapid non-destructive detecting techniques and instruments for nitrogen content in soils on a larger scale in the future.