Land Cover Classification And Prediction Based On Multi-classifier Ensemble Learning

Traditional individual classifiers often have their own strengths and weaknesses in land cover classification,and not all of them can always achieve high accuracy.To take advantage of the respective advantages of different base classifiers to improve the accuracy of classification,ensemble learning based on the classification results of different base classifiers is a possible solution.Additionally,simulating and predicting land cover changes can provide a reference for future rational development and utilization of land resources,which is important for the protection and improvement of ecological environment.In this study,based on the 1990,2000,2010 and 2020 Landsat images covering Jiangning District,Nanjing,the land cover types including water bodies,construction land,forest land,grassland,bareland and cropland were first classified by using the base classifiers including the Minimal Distance(Min DC),Mahalanobis Distance Classifier(Mah DC),Maximum Likelihood Classifier(MLC),Neural Network(NN),and Support Vector Machine(SVM)by feeding spectral features,vegetation index,terrain variable and retrieved land surface temperature derived from Landsat observations and DEM dataset.Based on this,three ensemble learning methods(including the voting method,random forest and evidence theory method)were used to achieve a classification combination of multiple classifiers’results to obtain higher accuracy land cover classification maps.Next,based on the best ensemble learning land cover classifications in 2000 and 2010,the 2020 land cover pattern was simulated by using CA-Markov model,PLUS model and ANN-CA model,followed by an accuracy evaluation of the simulated results via calculating a spatial agreement index between the 2020 best ensemble learning classifications and the 2020 simulated land cover pattern.As a result,the optimal simulating model was determined to predict the 2030 land cover distribution pattern.Finally,the land cover change and its driving socio-economic,natural and anthropogenic forces were jointly analyzed.The main research results were as follows:(1)Compared with the traditional base classifiers,the ensemble learning classifications gained higher accuracy when performing land cover classification.For the base classifications,in 1990,SVM algorithm achieved the best classification result,with an overall accuracy and Kappa coefficient at 89.00%and 0.80,respectively;In 2000,SVM algorithm continued to achieve the best classification performance,with an overall accuracy of 88.00%and kappa coefficient of 0.77 In 2010,the neural network method achieved the best classification result,with an overall accuracy of 86.75%and a kappa coefficient of 0.80;In 2020,the maximum likelihood classification method achieved the best classification result,with an overall accuracy of 82.25%and a kappa coefficient of 0.73.For the ensemble learning classification of the four periods,in 1990,random forest method achieved the best classification result,with an overall accuracy of 90.50%and a Kappa coefficient of 0.83;In 2000,random forest method achieved the best classification result,with an overall accuracy of 91.25%and a Kappa coefficient of 0.85;In 2010,evidence theory method achieved the best classification result,with an overall accuracy of 90.50%and a Kappa coefficient of 0.86;In 2020,random forest method achieved the best classification result,with an overall accuracy of 93.75%and a Kappa coefficient of 0.91.After exploring the four-period ensembled classifications,the dominant land cover types in the study area were cropland and construction land,the areal sum of both accounted for more than 80%of the total area of Jiangning District,followed by forest land,accounting for about 13%of the total area.Among them,the area of cropland decreased from 1216 km~2 in 1990 to 801 km~2 in 2020,and the area of construction land increased from about 63 km~2 in 1990 to 464 km~2 in 2020.Overall,forest land,grassland and construction land had an increasing trend,but cropland and water bodies had a decreasing trend,and bareland remainded relatively stable.(2)In terms of the simulation predictions,the CA-Markov model obtained a spatial agreement of 69.83%,the PLUS model at 98.54%and the ANN-CA model at 69.37%.The PLUS model was used to simulate the 2030 land cover distribution pattern in the study area and the result showed that the area of cropland in 2030 will be 709.70 km~2,a decreased by 10.30 km~2,and the area of construction land continuously increased by 15.59 km~2 compared to 2020,reaching about 479 km~2;Water bodies showed a slight decreasing trend,with its area decreasing by 3 km~2 in 2030 compared to 2020;There was no obvious change in forest land,grassland and bareland types.Overall,the 2030 simulated land cover distribution pattern was basically consistent with that of 2020.