| Longxi Loess Plateau is located in the western part of China's Loess Plateau.The geomorphic type is mainly loess hills and gullies.The terrain is broken and complex with dense hills and gullies.Affected by the natural geographical environment,most areas of the Longxi Loess Plateau are relatively arid,with sparse and uneven rainfall.The vegetation coverage in the region is low,and most of the north is desert grassland.The soil erosion is serious,leading to desertification,and the ecological condition is fragile.Land cover plays an important role in ecological environment research,urban planning and management,soil and water conservation,and its change research provides an important reference for regional sustainable development.Land cover plays a leading role in the ecological monitoring of the Loess Plateau in western Gansu.However,due to the complex distribution of land features,too broken terrain and large seasonal changes of vegetation in this area,the traditional land cover classification methods have some problems,such as"pepper and salt"noise or easy to ignore the information of broken features.How to effectively improve the classification accuracy of broken small target features and complex boundaries in land cover classification has always been a difficult problem to be solved.In the past 5-6years,the rapid development of deep learning semantic segmentation methods has provided a new way to solve this problem.In this paper,based on the Landsat data,the proposed deep learning IA?UNet model provides a reference for the complex land cover classification of the Loess Plateau in western Gansu Province,by using Tasseled-cap transform.The main results and conclusions are as follows:(1)Combined with Landsat data and field survey data,this paper constructs the land cover reference data sets of Huining County,Lintao county and Weiyuan County in the Loess Plateau of western Gansu.The data set focuses on the integrity of small target features and the complex boundary of Longxi Loess Plateau.(2)Build the semantic segmentation model of land cover in the Loess Plateau of western Gansu.Based on Attention U-Net semantic segmentation model,this paper introduces location attention module and channel attention module to enhance the extraction of local and global features.The IA?UNet model proposed in this paper is compared with the segmentation results of Attention U-Net?FCN?Res Net-FCN?U?Net and Deep Lab V3+models.The results show that the total accuracy,FWIo U and Kappa coefficients obtained by IA?UNet model are 0.25-2.09%,0.44-3.23%and0.01-0.04 higher than the other five models respectively.It is found that IA?UNet model can effectively extract small target features and complex feature boundaries of Longxi Loess Plateau.(3)It is verified that the introduction of tassel transform feature component is conducive to the improvement of classification accuracy.Experiments on the combination of model input data show that the combination of Landsat Image band(blue,green,red,near infrared,short wave infrared 1 and short wave infrared 2 band)and tassel transform three components(brightness,greenness and humidity)as input data can effectively distinguish easily confused features.The total accuracy,FWIo U and Kappa coefficients obtained from the combination of input data are 82.76%,71.00%and 0.68 respectively;Compared with the band combination of Landsat Image,the total accuracy,FWIo U and kappa coefficients are increased by 0.73%,1.17%and 0.01 respectively;Compared with the three component combination of tassel transform,the total accuracy,FWIo U and kappa coefficients are increased by1.31%,1.92%and 0.02 respectively.(4)Comparing the results of the proposed deep learning method with the existing pixel based(From-GLC dataset)and object-based methods(Globe Land30),the results show that the proposed method is more complete in the extraction of broken ground objects,and its total accuracy,FWIo U and Kappa coefficients are improved by23.64%,27.8%and 0.41 respectively compared with the traditional pixel based classification methods,and 9.83%and 0.41%respectively compared with the object-based classification methods 14.48%and 0.2%.(5)The IA?UNet model proposed in this paper can be generalized with time.The model trained based on 2017 data is transferred to Landsat images in 2002,2009,2015 and 2021.The overall accuracy,kappa coefficient and FWIo U of land cover classification results are 80-85.74%,0.67-0.79 and 67.03-75.05%respectively.Among them,the overall accuracy,kappa coefficient and FWIo U obtained in 2015 are the highest.The changes of water body,residence,cultivated land,grassland and forest land area in the four land cover classification results of Huining County,Lintao county and Weiyuan County from 2002 to 2021 were analyzed.The results showed that the area of residence,grassland and forest land showed an expansion trend in recent 20 years,and Huining County increased by 32.72 km~2,1147.53 km~2 and 90.37km~2 respectively;Lintao county increased 44.67 km~2,160.11 km~2 and 87.16 km~2respectively;Weiyuan County increased by 26.64 km~2,52.59 km~2 and 146.44 km~2respectively.Cultivated land decreased slowly,with 2640.01 km~2,292.22 km~2 and223.14 km~2 in Huining County,Lintao county and Weiyuan County respectively.The water area has no obvious change trend.In the last 20 years,the residential area and forest land in Huining county have changed greatly,which are 5.3%and 6.19%respectively;Lintao county has the largest change in residential area,which is 6.32%;Weiyuan County is also the place of residence,with the largest change of 4.81%. |
PDF Full Text Request