Study On The Construction And Optimization Of Simulation And Prediction Samples Of County Cultivated Land Quality Level Based On NDVI Machine Learning

Land resources are the material basis for human survival and development,and the most important part of cultivated land resources is an important guarantee for national food security.The cultivated land in our country has the characteristics of decreasing year by year,low quality grade,high fragmentation degree and few reserve resources.Actively carrying out the investigation and grade evaluation of cultivated land quality is conducive to the rational allocation of land resources,effective guarantee of food security,and actively promoting the implementation of cultivated land protection policies.With the progress of science and technology,the evaluation method of cultivated land quality has experienced the transformation from the traditional evaluation method(qualitative and quantitative)to the multi-technical comprehensive evaluation method(mathematical statistics and mathematical model,new method and integrated technology).However,at present,most of the multi-technical comprehensive evaluation methods are still based on the traditional sampling analysis,and there are still problems such as long time consuming and low efficiency in the evaluation process.At present,the pursuit of effectiveness and refinement of cultivated land resource management requires the exploration of faster and more efficient methods to evaluate the quality of cultivated land,such as trying to extract large-scale cultivated land vegetation information through high-resolution remote sensing image in batches,using the corresponding remote sensing information data processing and computer methods to obtain the quality level of cultivated land more quickly.In view of this,based on the NDVI(normalized difference vegetation index)remote sensing data,this paper carries out the comparison of the precision differences of cultivated land quality level simulation prediction and the study of sample data set optimization under different sample construction methods,and explores a faster and more efficient method of sample construction optimization for cultivated land quality level simulation prediction.Dongtai City,a county-level city located in Jiangsu coastal area,is selected as the research area.Based on the results of cultivated land quality grading and bimonthly NDVI remote sensing data,and based on traditional random point samples,cultivated land patch samples,regional and global samples,respectively,the simulation prediction of cultivated land quality grade is carried out.The prediction accuracy and precision F1 values under different samples are compared,and the remote sensing based on NDVI is explored The feasibility of using image and machine learning method to simulate and predict the cultivated land quality level and how to build the optimal sample data set.First of all,based on the selection of traditional random point attribute samples,four mainstream machine learning models are used to simulate the prediction of cultivated land quality level,calculate the prediction accuracy and precision F1 value of different machine learning models,and judge whether the traditional point samples are suitable for the prediction of cultivated land quality level.Secondly,introduce the area and type attributes of cultivated land patches,construct random patches and area series patches samples,and calculate the accuracy and precision F1 value of cultivated land quality level simulation prediction of the above samples,to explore whether constructing patch attribute samples can effectively improve the prediction accuracy and precision F1 value.Finally,according to the natural geographical division of the study area,we construct the regional samples and the global samples respectively,calculate the accuracy and precision F1 value of cultivated land quality level simulation prediction under the two sample construction methods,and explore the impact of the regional samples and the global samples on the accuracy of cultivated land quality simulation prediction results.The conclusions are as follows:(1)Based on the selection of traditional random point attribute samples,the accuracy and accuracy of the prediction of the cultivated land quality level obtained by machine learning model are not high in general,and there is a certain gap in the direct application of the prediction and evaluation of cultivated land quality level.Four mainstream machine learning models are used to simulate and predict the cultivated land quality level for random point attribute samples.The prediction accuracy of the four models is about 60%.Among them,SVM model has the highest prediction accuracy,which is 63.03%,and RF model has the lowest prediction accuracy,which is 57.71%.After comparing the prediction precision F1 values of different machine learning models,it can be seen that the prediction accuracy F1values of the four models are relatively low.Among them,the RF model has the highest prediction precision F1 value,which is 45.7%,and the DT model has the lowest prediction accuracy,which is 42.1%.There is a certain gap between the direct application and the prediction and evaluation of cultivated land quality level.In addition,the prediction of cultivated land quality of different levels is unbalanced,and the accuracy and precision F1 of prediction are quite different.The accuracy and precision F1 of prediction of third-class land is the highest,and the accuracy and precision F1 of prediction of first-class land is the lowest.(2)There is a correlation between the cultivated land patch and its area and the cultivated land quality level.When the area and type of cultivated land patch are selected and included in the evaluation sample attribute,the prediction accuracy and accuracy F1 are greatly improved,which can be better applied to the prediction and evaluation of cultivated land quality level.According to the random patch attribute samples,the cultivated land quality level is simulated and predicted.The accuracy of prediction reaches 63.5%,and the prediction precision F1 value reaches 53.2%.Compared with the prediction accuracy and accuracy F1 value based on the random point attribute samples,it has a certain improvement.According to the area series patch samples,the accuracy and accuracy of the prediction of cultivated land quality level are 86.1% and 85.8% respectively.Compared with the random spot attribute samples and random patch samples,the accuracy and precision of the prediction of F1 are significantly improved.(3)There are spatial differences in cultivated land quality and its associated attributes.The prediction precision F1 of cultivated land quality level based on regional samples is generally better than that based on global samples.The accuracy of cultivated land quality level prediction in Lixiahe plain area,coastal agricultural area and coastal beach area is 80%,68.18% and 76.0%,respectively.The prediction accuracy of cultivated land quality level based on the whole area sample is 73.37%,which is lower than that of Lixiahe plain and coastal beach area,and higher than that of coastal agricultural area.Comparing the prediction accuracy F1 values of cultivated land quality level of the regional samples and the global samples,we can see that the prediction precision F1 values of the regional samples are 80.5%,74.9% and 83.5%respectively,and the prediction precision F1 values of the global samples are 73.4%.The prediction accuracy based on the regional samples is generally higher than that based on the global samples.(4)The evaluation samples constructed by different methods will lead to a great difference in the accuracy and precision F1 of cultivated land quality level simulation prediction.In order to improve the accuracy of simulation prediction results of cultivated land quality level based on NDVI machine learning,when building learning training samples,we should consider the factors of cultivated land management and utilization such as patch area,as well as the spatial differences of cultivated land quality and its related attributes;on the basis of patch attribute samples of area series,we should further consider the construction of samples by zones,and carry out simulation prediction of cultivated land quality Evaluation.