Land Cover Continuous Change Detection And Classification Based On Multi-Temporal Remote Sensing Imagery

Land Cover Change plays a crucial role in addressing global environmental change and sustainable development for governments and researchers worldwidely.Land cover change can be classified into two types based on the nature of surface landscape before and after the change:Land Cover Conversion(LCC)and Land cover Modification(LCM),withing different ecological implications.Remote sensing technology,with advantages such as large-scale coverage,low cost,and various spectral and temporal resolutions,has become a powerful tool for monitoring land cover change.In recent years,the development of computer technology and time-series change detection algorithms has greatly enhanced the detection capabilities of LCC and LCM.However,most studies have focused on LCC,and few have simultaneously detected and classified both types of change,leading to a lack of quantitative analysis and in-depth understanding of their differences.This study focuses on the LCC and LCM change detection based on Continuous Change Detection and Classification(CCDC)algorithm and Bayesian Estimator of Abrupt change,Seasonal change,and Trend(BEAST)algorithm,which takes into account the sudden changes,seasonal changes,and trends using DeepLabV3+semantic segmentation algorithm in Jinchang,Gansu Province.The proposed approach can accurately classify the LCC and detect the comprehensive changes in LCM.The results of this research have important significant for the management of land and resources management and regional sustainable development.The main contributions of this study are summarized as follows:(1)A coupled continuous change detection method for LCM and LCC was developed using the CCDC and BEAST algorithms.The CCDC algorithm,which has excellent noise filtering and multi-band change detection capabilities for time series,was applied to detect LCC based on long-term Landsat images.Meanwhile,the BEAST algorithm,which has non-linear detection,sub-pixel detection,trend detection,and small-scale change detection capabilities,was to detection of sub-time series segmented by the CCDC change points.This method addresses the drawbacks of the fixed parameters of the CCDC algorithm and its difficulty in detecting non-linear and trend change points,while also resolving the issue of the BEAST algorithm’s inability to directly classify detection results.The combination of the two algorithms achieved an overall accuracy of 81.5%in detecting change sample points,with overall accuracies of 71.2%and 81.5%for LCC and LCM,respectively.The overall accuracy of distinguishing between the two types of change points reached 88.4%.The detected LCC changes had a large magnitude and were closely related to human activities,while the LCM had a smaller magnitude and were more closely related to climate conditions and surface phenology changes.These two types of changes have different ecological indicator meanings.(2)A Semantic Segmentation Method Based on DeepLabV3+for CCDC LCC Results.In this study,a dataset consisting of the multi-dimensional harmonic model coefficients generated by CCDC was used as the input image,and a land cover map based on land use surveys and high-resolution imagery correction is employed as the ground truth label.Spatial generalization evaluation of the land cover classification results showed that the average F1 score and average Intersection over Union(IoU)for each class were 80%and 69.9%,respectively.The overall performance of the model was evaluated using the metrics of Frequency Weighted IoU(FWIoU),overall accuracy,and Kappa coefficient,which were found to be 87.9%,93.6%,and 0.84,respectively.The temporal generalization performance of the model was found to be better than its spatial generalization performance,with an average F1 score and average IoU of 84.1%and 75.7%,respectively,which were higher than the corresponding spatial evaluation results.The overall metrics for temporal generalization were found to be 89.5%,94.3%,and 0.89 for FWIoU,overall accuracy,and Kappa coefficient,respectively,which were slightly higher than the corresponding spatial generalization metrics.Overall,the DeepLabV3+based land cover classification exhibited good spatiotemporal generalization.Compared to the random forest algorithm,the DeepLabV3+model achieved higher overall accuracy and Kappa coefficient of 95%and 0.93,respectively,compared to 81%and 0.71 for the random forest algorithm,and did not exhibit "salt and pepper" noise.Compared to the GlobeLand30 and CLCD products,our classification showed consistency and better spatiotemporal classification accuracy.(3)Analysis of Land Cover Classification and Change in Jinchang from 2000 to 2020.In this study,we utilized the optimally trained DeepLabV3+model to perform annual land cover classification and change analysis for Jinchang from 2000 to 2020.Our findings revealed that the dominant land cover types in Jinchang were grassland,cropland,and bare rock.Grassland exhibited the largest change in area,with an average annual decrease of 18.5 km2,making it the land cover type with the fastest rate of decline.In contrast,built-up land exhibited the fastest rate of increase,with an average annual increase of 8.9 km2,and the period from 2010 to 2015 exhibited the highest degree of dynamic change across most land cover types.The main distribution of land cover transition occurred in the grassland,which is a widespread and climate-sensitive land cover type,as well as in the cropland,which is influenced by intercropping and crop rotation strategies.The main distribution of LCC occurred in newly-cultivated and reclaimed cropland,as well as in the expansion and development of built-up land.