Research On Optimization Of Remote Sensing Classification Methods Based On Random Forest Algorithm

Ningxia Qingtongxia Irrigation District is one of the four ancient irrigation in China,it is one of China’s grain production base.It is easier to obtain high quality optical images due to drought and less rain,and the variety of crops and planting structure,The acquisition of land use,arable land and crop distribution maps in this area has important guiding significance for the formulation of food security policy,sustainable agricultural development,land readjustment and development planning.At the same time,it also has a good research and promotion significance..In this paper,Landsat8 data and Proba_v data are uesed to study the Qingtongxia Irrigation District.Based on the random forest algorithm combined with the crop growth cycle and ground sample data,the optimal parameters of the random forest algorithm are obtained by the data,and apply this parameter to the subsequent landsat inage classification.The differences between single-time and time series classification and the influence of additional features on the classification results were compared and analyzed by Landsat data.And analyzing the importance of the classification results based on the time phase.Finally,the optimal land use map,cultivated land map and crop distribution map of the study area in 2017 and 2018 were obtained.The main research work is asfollows:(1)Parameter determination of random forest algorithm.Proba_v image data with 100m resolution in 2018 was adopted,and training samples and verification samples were selected by combining with field sampling data.Input image is divided into three conditions:image the series data acquired in 2018,high-quality image time series data and high-quality image time series data changing the order of feature arrangement.Then the random forest algorithm tree set for 100 and 200 respectively,and the number of characteristics of land use classification for step 5 for the study area land use classification tests.The results show that the classification accuracy increases with the number of trees.When the number of trees is set large enough,the classification accuracy tends to be stable,and the number of trees should be set to 100 or 200 to ensure timeliness.The increase of classification accuracy is not positively related to the number of set feature numbers.When the feature quantity is set to the square root or logarithm of the total feature number,the accuracy is the highest.Poor quality images have a negative impact on classification accuracy,and data should be screened before classification.The input order of image features has a limited effect on classification accuracy,which may be caused by random errors and can be ignored.(2)Comparison of single-phase image and time-series image classification.The random forest algorithm was used to classify the land use classification,cultivated land layer and crop classification based on single-phase and time-series images in the 2017 and 2018.The results show that the classification results in 2017 and 2018 are consistent.The three classification results show that the time series images have obvious advantages over the single-phase images.The time series images can overcome the cloud image and the one-sided problems of classification existing in the single-phase image classification.The accuracy of classification is higher than that of single-temporal images of any period,and it is more consistent with the actual ground objects;the classification of the single-phase image is often affected by the quality ot the influence and the difference between the objects in the period.(3)The influence of additional features on the classification results.The images of 2017 and 2018 are calculated based on the original bands,and seven scenarios are set as characteristic bands and added 1nto time series images to participate in classification based on random forest algorithm.Evaluating the accuracy of the results of the classification and tObtain the land use map,cultivated layer map and crop classification map of the optimal study area.The results show that the classification results in 2017 and 2018 are consistent.The classification results show that the classification accuracy of NDVI time series images cannot be optimized.The classification accuracy is improved with the increase of image feature bands.And the speed of improvement is the classification of land use>crop classification>the extraction of cultivated land.When all the characteristic bands are classified,the precision is the highest.In 2017,the highest overall accuracy of land use is 89.2%,Kappa is 0.88,F1-Score is 83.1;the overall accuracy of cultivated land is 97.2%,Kappa is 0.94,FI-Score is 96.8;the overall accuracy of crop classification is 96.4%,Kappa is 0.95,and F1-Score is 95.2.In 2018,the highest overall accuracy of land use is 92.8%,Kappa is 0.92,F1-Score is 87.1;the overall accuracy of cultivated land is 97.9%,Kappa is 0.95,F1-Score is 97.6;the overall accuracy of crop classification is 97.2%,Kappa is 0.97,and Fl-Score is 95.1.(4)Feature importance analysis.Based on the results of land use classification and crop classification in 2017 and 2018,the feature importance analysis was earried out according to the image phase as a whole,and the differences in time-phase cumulative classification were compared according to the order of feature importance and time.The results show that the order of the importance of the phase is not related to the classification method,but the change of importance will be affected.he feature importance ranking in 2017 is 0517>0805>0704>1125>0906>0922,and the feature importance ranking in 2018 is 0504>0605>0520>0824>0621>0808.The difference between the image quality and the perfonnance of the objects during the image period detenmines the proportion of the image in the time series classification.Images based on the order of feature importance achieve faster classification results than time sequence.When the importance of the feature reaches about 80%of the total time series,the classification accuracy of the two cases tends to be the same.