Land Use/land Cover Change Detection In The Huangshui River Basin Based On Random Forest

Land use/land cover change has become one of the hotspots in current global environmental change research.Medium-resolution US terrestrial satellite data is an important source of remote sensing data for land use/land cover at the global and regional scales due to its continuous archival data supply.In complex terrain areas with diverse land use types,obvious vertical differences,and high spatial heterogeneity,it is difficult to obtain high classification accuracy of land use.It is of great significance to study the adaptability of random forest method to complex terrain,improve the accuracy of remote sensing classification under complex terrain and explore the temporal and spatial variation of land use/land cover in the complex terrain.This study is carried out in the area with high altitude and complex broken terrain.Based on the 1999 Landsat7ETM,the 2011 Landsat5TM and the 2017Landsat8OLI imagery,combined spectral,texture,and terrain information,the random forest method was applied to extract and evaluate the land use/land cover information of the three remote sensing images of the Huangshui basin.Finally,the change detection method after classification was selected to dynamically analyze the land use/land cover change in the Huangshui River Basin in the past 18 years.The main research conclusions are as follows:(1)Land use/land cover information was extracted from the mid-high mountainous area,loess hilly area and valley plain area based on the 1999Landsat7ETM+,2011 Landsat5TM,and 2017 Landsat8OLI three-phase remote sensing image of the remote sensing image by using the random forest algorithm.The results show that the overall accuracy of the above three geographical divisions reached 88.53%,87.05%,and 84.70%in 1999,and the kappa coefficients were 0.85,0.84,and 0.82,respectively.In 2011,the overall accuracy of the above three geographical divisions reached 88.50%,87.54%,85.04%,Kappa coefficients are 0.85,0.84,0.82,respectively;In 2017,the overall classification accuracy of the above three geographical divisions reached 89.17%,87.42%,and 85.43%,respectively,and Kappa was 0.86,0.84,and 0.83,respectively.Through the above research,the author found that the random forest method can obtain better classification accuracy when classifying complex terrain areas,and then proves the applicability of random forests to complex terrain classification.(2)Comparing of the classification accuracy with fused and unfused images in the middle and high mountainous areas,loess hilly areas,and valley plains in 1999and 2017 under the random forest algorithm.In 1999,the overall accuracy of the middle and high areas was 88.83%,the overall accuracy of the loess hilly area was87.45%,and the overall accuracy of the valley plain area was 85.01%.The accuracy of image classification after fusion in 1999 was 0.3%,0.4%,0.31%higher than that of unfused images;In 2017,the overall accuracy of the middle and high mountain areas was 90.01%,the overall accuracy of the loess hilly area was 87.88%,and the overall accuracy of the valley plain area was 85.80%.The classification accuracy of the fused image is 0.84%,0.46%,and 0.37%higher than that of the unfused image,respectively.It indicates that the images with higher resolution have more obvious spectral information and texture information,which improves the classification accuracy.It also proves that the images with higher resolution can better extract the land use/land cover information of complex terrain.(3)This paper constructed a random forest model adapted to three geographic regions by quantitatively analyzing the number and characteristic variables of decision trees in each geographic region.Using OOB accuracy to evaluate the effectiveness of random forest models in each geographic region,the author found that the random forest model OOB with 500 decision trees constructed the best precisionin three geographic regions.(4)The land use/land cover change detection in each geographical division of the Huangshui River Basin shows that during the nearly 18 years from 1999 to 2017,the area of grassland and water area in the middle and high mountain areas decreased by 88.86km~2 and 0.99km~2 respectively.The area of forest land and unused land increased by 15.92km~2and 73.93km~2respectively,and the annual change rates of land use type in grassland and water area were 0.11%and 0.43%,respectively.Among the land use types in the loess hilly region,the area change of grassland and urban built-up land is more obvious,the grassland area is reduced by 94.93km~2,the urban built-up land area is increased by 115.17km~2,and the annual change rate is10.28%.The area decreased by 53.59km~2,while the area of forest land increased by33.26km~2.Among the land use types in the valley plain area,the area of cultivated land decreased more obviously,which was reduced by 528.3 km~2,and the area of grassland and urban built-up land increased by 288.67 km~2 and 237.32 km~2respectively.The area of water and unused land was relatively stable and decreased3.41km~2and 2.6km~2respectively from 1999 to 2017,the annual change rates of urban built-up land and waters are 7.49%and 1.25%respectively.