Remote Sensing Recognition Of Plastic-film-mulched Farmlands On The Loess Plateau Based On The Google Earth Engine

Plastic film mulching has made an outstanding contribution to agricultural production and food security in China,but also caused many serious environmental problems.It is of great importance to quickly and accurately obtain the spatial distributions of plastic-film-mulched farmlands.In order to establish a framework for remote sensing recognition of plastic-film-mulched farmlands,this study took the Tuanjie Town of Dingxi City in Gansu Province and the Changwu County of Xianyang City in Shaanxi Province as the research areas,which are typical dry farming areas with heavy plastic film application on the Loess Plateau.Based on the Google Earth Engine and Landsat-8 surface reflectance data,the methods of feature importance analysis and best classification accuracy were used to select the optimal textural features.Then,the random forest(RF)algorithm with optimized parameters was used to extract the plastic-film-mulched farmlands and to select the best feature combination.Finally,based on the best feature combination,the classification performance of four different algorithms was evaluated through the comparisons among the classification results based on the algorithms of random forest(RF),support vector machines(SVM),decision tree(DT),and minimum distance classifier(MDC)and through the Mc Nemar’s test.This study realized accurate identification of plastic-film-mulched farmlands in areas with complex terrain features in China.The results can provide theoretic and technical supports for the studies related to spatial variations and production sustainability with plastic film mulching in the near future.Some main conclusions were drawn as follows.(1)The Landsat-8 satellite could provide an effective data source for the recognition of plastic-film-mulched farmlands on the Loess Plateau.The Landsat-8 surface reflectance data set provided by the GEE platform was used to extract the plastic-film-mulched farmlands.The results showed that the overall accuracy of classification under different feature combinations all exceeded 80%.The farmlands mulched with plastic film in Tuanjie Town were mainly concentrated in the north and southwest,and scattered in the middle and east parts.Plastic-film-mulched farmlands in Changwu County were concentrated in the northwest and central,sparsely distributed in the eastern gully areas,and rarely distributed in the southern mountainous areas.The area of plastic-film-mulched farmlands in Tuanjie Town was about 18.05 km~2,accounting for 13.42%of the total area,of which the areas mulched by white and black plastic film were about 7.79 km~2and 10.26 km~2,respectively.The area of farmlands mulched with plastic film in Changwu County was about 3.48 km~2,or about 0.61%of the total area.(2)Optimization of the key parameters of the random forest(RF)algorithm could greatly improve the classification accuracy of remote sensing images.When the non-critical parameters remained default,the optimal parameter combination of the number of trees(T)and the number of variables per split(M)could minimize the out-of-bag error,but with acceptable computing efficiency.Therefore,before using the random forest algorithm to conduct image classification,its key parameters should be optimized to obtain more reliable classification results.(3)The classification performance based on the selected optimal texture features was better than that based on all texture features.Compared with other schemes,the recognition result based on the combination of‘spectral+index+optimal textural features’was the best.Based on the combination of‘spectral+index+optimal textural features’,the overall accuracies of Tuanjie Town and Changwu County were 95.05%and 96.77%,which were2.48%-14.55%and 3.22%-12.50%higher than those based on the single feature schemes,respectively.At the same time,the overall accuracies were 1.24%and 2.01%higher than the combination of‘spectral+index features’,and 0.62%and 3.62%higher than the combination of‘spectral+index+all textural features’in Tuanjie Town and Changwu County,respectively.Thus,we recommended the methods of feature optimization and combinations of multiple features to identify plastic-film-mulched farmlands.(4)The random forest(RF)algorithm had obvious advantages over the order three algorithms of support vector machine(SVM),decision tree(DT),and minimum distance classification(MDC).The overall accuracies of the random forest algorithm in the land covers classification of Tuanjie Town were about 3.10%,7.74%,and 50.78%higher than those of the other three algorithms,respectively.The overall accuracies of the random forest algorithm in the land covers classification of Changwu County were 3.22%,9.67%,and16.53%higher than those of the other three algorithms,respectively.Combined with the results of Mc Nemar’s test,the Z values between the random forest algorithm and the other three algorithms were 2.89,5.00,and 12.88 in the land covers classification at Tuanjie Town,respectively.At the same time,the Z values between the random forest algorithm and the other three algorithms were 2.31,4.71,and 6.40 in the land covers classification at Changwu County,respectively.There was a significant difference in the classification accuracy between the random forest algorithm and the other three algorithms.