| Forest phenology refers to the recurring annual growth and development rhythms of forest plants driven by climate and other related factors,including the life processes of germination,branching,leaf development,flowering,fruiting,leaf falling and dormancy.Changes in forest phenology not only reflect the dynamic response of biosphere to climate change,but also affect energy exchange,hydrological cycle,primary productivity,carbon storage and vegetation distribution of ecosystem.Therefore,monitoring forest phenological change has important theoretical significance and practical guidance value.The traditional methods of phenological observation are based on phenological observation networks.In general,results from these networks are difficult to implement in wide forested areas.The development of remote sensing technology enables the monitoring of forest phenology over a long time span and wide region.Existing phenological researches mainly use those sensors with high time frequency and coarse spatial resolution to monitor phenology.However,such sensors induce serious errors in phenology monitoring due to the existence of a large number of mixed pixels in small regional scale and high heterogeneity landscape.Sensors with medium spatial resolution,such as Landsat,face huge challenges in annual phenological monitoring due to their relatively long revisit periods.In addition,these phenology research areas are usually concentrated in the middle and high latitude vegetation zones,and the phenology researches over the subtropical plateau regions remain still insufficient.In addition,the application value of phenology in forest management still needs to be further explored.In this study,the eastern part of Dali Bai Autonomous Prefecture was used as the research area,combined with multi-source data,to get the annual phenological parameters of the forest at a spatial scale of 30 m.Besides,this study realized spatial mapping of forest tree species groups and estimated forest aboveground biomass based on phenological parameters.Therefore,the overall goal of this study is to(1)improve the integration method among multi-source remote sensing data to increase the observation frequency,and then develop an improved image synthesis model aimed at generating daily remote sensing images with a spatial resolution of 30 m;(2)to propose an adaptive forest annual phenological monitoring framework based on daily composite images;and(3)to finally build a forest tree species group classification and biomass mapping model in consideration of phenological parameters,model coefficients and radar signals.The main contents and results of this study were summarized as follows:(1)Improved the integration method of Landsat and Sentinel-2.Modifying the integration coefficients of Landsat and Sentinel-2 Surface Reflectance(SR)was done to improve the band-specific matching degree among different sensors.The band-specific regression models between Landsat and Sentinel-2 showed that the blue bands of Landsat 8,Sentinel-2 and Landsat 7 images had the lowest R~2 and the highest R~2 was observed for SWIR2.After a series of integrations,the SR RMSEs between Landsat 7 and Landsat 8 after adjustment were slightly lower than those before integration,while the SR RMSEs between Sentinel-2 and Landsat 8 after the integration were obviously smaller than those before the integration,indicating that the spectral differences between these two sensors reduced.The integrated SR of Landsat and Sentinel-2 took on more regular and periodic changes over time,making the integration method more conducive to seasonal change analysis.(2)Modified a daily synthetic image construction method by using the integrated SR from long time series of multi-source remote sensing.Modified the continuous change detection and classification(MCCDC)model was conducted to balance model accuracy and computational efficiency and it can be used to synthesize daily Landsat clear images in consideration of land cover change.Calculating the R~2 between the synthesized and the real SR images of Landsat 5,7 and 8in four seasons was implemented to evaluate the composite accuracy,and the results showed that Landsat 8’s synthesis accuracy was the best among the three sensors,which may be because the integration of different sensors took Landsat 8 as reference data.Generally speaking,the R~2value was higher in winter and lower in summer.This may be attributed to the fact that the number of images acquired in winter was larger than that in summer,leading to a more robust fitting in winter.The R~2 of visible bands was generally lower than that of the remaining three infrared bands,especially the R~2 for Blue band was the lowest.The reason was that the blue band was more susceptible to atmospheric influence,resulting in unstable SR values in time series.Visual comparison of the composite image with the real Landsat SR image indicated that the composite image successfully depicted many types of land features.The proposed MCCDC model provided daily synthetic images,enabling the monitoring of annual phenology at a resolution of 30 m.(3)Mapped the spatio-temporal dynamics of annual forest phenology and analyzed their responses to climate change at the scale of 30 m resolution.Two long time series classification algorithms were used:(i)MCCDC coefficients coupling support vector machine(SVM)and forest field data,and(ii)vegetation change tracker(VCT).The two algorithms were evaluated in terms of their ability in annual forest cover mapping from 1999 to 2019.The results showed that the forest cover mapping accuracy in consideration of MCCDC coefficients was higher than that of VCT,and VCT algorithm was not suitable for applying in cloudy and rainy areas.Subsequently,based on three different indices derived from daily synthetic images,the logistic regression equation was applied to extract annual forest phenology parameters including the start of season(SOS),the end of season(EOS)and the length of season(LOS),followed by a responsive analysis of forest phenology to climate change.Assessing the ability of different indices in extracting phenology parameters,it was found that the forest phenology accuracy extracted from the land surface water index(LSWI)was the highest.This may be attributed to the LSWI used NIR and SWIR1 bands,which had the higher synthetic accuracy.The major SOS were distributed in 110-135th day,average EOS were mainly distributed in over 330th day,and average LOS were distributed in 200-225 days.The change trend of forest phenology was further analyzed.The SOS and EOS in most forest areas showed a prolonged trend,with an average delay of 0-0.5 days per year.The LOS of most forests shows a shortening trend,with an average reduction of 0-0.5 days per year.Partial correlation coefficients were used to analyze the effects of mean temperature and total precipitation on phenology in dry and wet seasons.The results showed that SOS/EOS was positively correlated with temperature in wet and dry seasons,and SOS/EOS was negatively correlated with total precipitation in wet and dry seasons.The correlations between SOS and meteorological factors in dry season were stronger than those in wet season,while the correlations between EOS and meteorological factors in wet season were stronger than those in dry season.These results indicated that SOS delay was mainly caused by temperature increase and total precipitation decrease in dry season,while EOS delay was driven by temperature increase in wet season and SOS delay.(4)Developed a phenology-based forest tree species group classification and optical-radar-phenology integration AGB mapping framework.At a time interval of 3-4 years,the effects of different input variables and machine learning algorithms on the classification of forest tree species groups were compared.The results showed that the classification algorithm using MCCDC coefficients and phenological parameters was better than the traditional spectral inputs.After determining the MCCDC coefficients and phenological parameters as input variables,two machine learning algorithms including SVM and Random Forest(RF)were compared in terms of classification performance,the study found that SVM was slightly better than RF.Finally,the SVM algorithm coupled with MCCDC coefficients and phenological parameters was used to map forest tree species group in the entire study area,and the results showed that the area of pines>oaks>walnuts>other forest tree species.The impacts of forest tree species groups,combinations of different variables and different machine learning algorithms on forest AGB mapping were further analyzed.The results showed the optical-phenology-radar model constructed by forest tree group classification can improve the accuracy of AGB model,and the RF accuracy was always greater than those of Stochastic Gradient Boosting(SGB)and SVM.This was because the important variables were obtained from the feature importance command of RF algorithm.In addition,SGB involved more parameter settings than RF,and the performance of the model may be different due to the selection of these parameters.Thus,RF algorithm coupled with the spectral variables,MCCDC coefficients,phenological parameters and radar variables under the forest species group classification produced the best AGB prediction.The AGB in most study areas was within 30-55t/ha,indicating that the AGB level was relatively low,but the average value,standard deviation and total amount of AGB in the study area increased,indicating that the level of AGB was increasing.With the delay of SOS and EOS and the shortening of LOS,the AGB of pines and oaks increased,while the AGB of walnuts decreased.The framework proposed in this study provides new insights for phenological monitoring in subtropical plateau areas with frequent cloudy weather events,and successfully extracts the spatiotemporal pattern of forest phenology in this area and realizes the phenology-based forest species group mapping and AGB mapping.These products reconstructed the change trend of local forest distribution and phenological change characteristics under the scenario of global warming,which can help forest researchers and managers better understand the evolution process and health status of forests,and lay a solid technical foundation for formulating reasonable forest management strategies in the near future. |
PDF Full Text Request