Application Of Spatio-temporal Sampling Strategy Based On Supervised Learning In Land Cover Classification

Long time series of land cover changes are critical in the analysis of long-term climate,environmental,and ecological changes.Although several moderate to fine resolution global land cover datasets have been publicly released and they show strong consistency at global scale,they have large deviations at regional scal e.Furthermore,the high-quality land cover datasets before 2000 are not available and the classification consistency among different datasets is low.Thus,long time series of land cover datasets with high quality and consistency are in great demanding.The Landsat series of satellite imagery composed of 8 successive satellites can be traced back to 1972.In addition,the newly available satellite data have the capability to construct time series satellite images and the time series analysis method can subsequently be employed for making high-quality land cover datasets.Therefore,a new method based on machine learning for temporal and spatial expansion of land cover mapping was proposed by using the advantages of two types of satellite data.1.In terms of temporal expansion,an algorithm for automatic sampling with high precision based on existing data sets is proposed and a new strategy for sample migration for earlier years is proposed.Because large amount of accurate samples are reqired for high precision land cover mapping based on machine learning methods,which is time consuming and labor intensity,the high-quality land cover datasets at Heihe river basin from 2011?2015,which were retrieved using existing method,are used for automatically,quickly and accurately collecting training samples.In order to map earlier land cover with high accuracy,temporal expanding experiments through model migration and sample migration are carried out respectively;the strategy of sample migration has been proved better for temporal expansion and it is consequently adopted to map the Heihe River Basin long-term land cover dataset,which covers the land cover of 1986,1990,1995,2000,2005,2010,2011,2013,2014,2015.The confusion matrix was used for verification;the average accuracy reaches ?90% and the average Kappa coefficient is ?0.88.It is the longest time-series land cover dataset with high accuracy and 30 m resolution at Heihe river basin.The experimental results show that the automatic sampling algorithm proposed in the study and the strategy of stable sample migration for earlier years can effectively improve the accuracy of historical land cover mapping and the continuity over several decads.2.A spatail expansion method for supervised automatic labeling samples based on maximum feature spectral similarity is proposed.It aims to the problem of insufficient training samples for supporting supervised classification.The Gini index of the classifier is used to select important spectral indicators to calculate the maximum similarity between the data.The method was used to produce the land cover mapping in the entire basin of Qilian Mountain for verifying the feasibility.The overall accuracy of the mapping is 91.18% and the Kappa coefficient is 0.89.The results show that the sample automatic labeling method proposed in the study can quickly and accurately label samples,and can improve the problem of insufficient training samples in large-scale land cover mapping to a certain extent,which makes contribution to improve the application of machine learning in the field of land cover mapping.