Mapping Forest Stock Volume Of Planted Forest Using Multi-source And Multi-temporal Remote Sensing Data With Isolated Random Forests

Forest stock volume(FSV)is one of the most important indicators to evaluate the quality of forest resources,and the richness of the forest ecosystem in a country or region.Traditional field survey methods cannot meet the needs of modern resource monitoring.In recent years,the rapid development of remote sensing and artificial intelligence technology has made it possible to accurately estimate FSV based on multi-source and multi-temporal remote sensing data.Currently,FSV estimation using remote sensed images mainly focuses on feature extraction,feature selection and estimation models.However,due to the influence of factors such as sensors and atmospheric conditions,the spectral features of multi-temporal remote sensing data from the same growing season show significant spatial and temporal differences.This not only weakens the ability of remote sensing features to accurately reflect forest growth characteristics,but also limits the generalizability of remote sensing estimation models and restricts the accuracy and reliability of FSV remote sensing estimation.This study addresses the problem of spatial and temporal variation in multi-temporal remote sensing data within a short period of time.Using the Wangye Dian forest field in Karachin Banner,Inner Mongolia as the study area,the study proposes a spatial and temporal integrated multi-temporal remote sensing feature extraction method based on isolated random forests,and constructs a feature selection framework based on forward feature selection to take into account the autocorrelation and combination effects of remote sensing features.On this basis,the adaptive Stacking integration algorithm is proposed by combining the optimal feature set and integration algorithm,and a model and method for estimating the amount of plantation forest stock applicable to multi-source and multi-temporal remote sensing data is constructed,thus improving the accuracy and stability of the estimation of the amount of plantation forest stock.The main research findings are as follows:(1)The Multi-temporal Feature Extraction based on Isolated Random Forest(FEIRF)improves the sensitivity of remote sensing features to FSV.For the same place,among the features extracted from multi-temporal remote sensing data,those based on FEIRF method have a smaller standard deviation than those based on Window Averaging Method(WAM).In addition,there was a higher degree of response between the FEIRF-extracted multi-temporal features and the FSV.Among the multi-temporal Sentinel-2 image features extracted using the FEIRF method,the absolute mean value of Pearson coefficients of the top 10 remote sensing features with the highest correlation to FSV increased by 10.50%.The features extracted based on the FEIRF method can obtain better FSV estimation results.Among them,the model accuracy of the multi-temporal Sentinel-1 and Sentinel-2 features extracted based on the FEIRF method was improved by 3.85%～5.56%and 2.14%～5.26%respectively.(2)A feature selection framework that takes into account feature autocorrelation and combination effects(Feature Selection Considering Autocorrelation and Combination Effect,FSACE)can reduce the covariance between features to a greater extent,and more accurately select the optimal combination of features with a higher degree of response to FSV.Compared with methods that only consider the response of features to FSV or the training accuracy of the model,FSACE can better exploit the combination effect between features and significantly improve the estimation performance of the model.In FSV estimation based on single-source remote sensing data or on three remote sensing data simultaneously,including Sentinel-1,Sentinel-2 and Landsat-8,the accuracy of model estimation based on the features selected by FSACE was better than that of feature selection methods that focused only on feature and FSV responsiveness or model training accuracy.Based on single-source remote sensing data,the mean rRMSE values were reduced by 2.35%-4.39%and 4.36%-5.29%,respectively.The mean rRMSE values were reduced by 3.00%and 4.36%respectively based on 3 remote sensing data simultaneously.Among them,FSACE obtained the best model estimation accuracy based on three kinds of remote sensing data,and the rRMSE(Relative Root Mean Squared Error)and R 2(Coefficient of Determination)of the optimal model were 24.35%and 0.53 respectively.(3)Based on the complementary effects of the models,the Adaptive Stacked Generalization(ASG)is proposed to improve the accuracy and robustness of the models by making full use of the characteristics and advantages of each base learner.Integration methods based on Bagging,AdaBoost and Stacking can all improve the estimation performance of a single model to a certain extent.Among them,the ASG has the best estimation accuracy(rRMSE:20.97%,R~2:0.65),and the rRMSE decreases by 3.00%～10.28%compared with the Bagging or AdaBoost integrated models.In addition,the stability of the integrated model is better than that of the single model,and the FSV distribution maps obtained from the model inversions show that the FSV inversion maps of the integrated model have smaller coefficients of variation for the small group of image elements.Among them,the ASG has the best stability.(4)Remote sensing features with better stability can be extracted based on multi-source multi-temporal data,thus improving the accuracy and reliability of FSV estimation.Through the feature selection based on FSACE,the advantages of multi-source remote sensing features are fully utilized,and the estimation performance of the model is improved.The mean value of rRMSE decreases by 2.59%～6.04%in the model based on multi-source multi-temporal remote sensing data for FSV estimation.In addition,the adaptive Stacking integrated model constructed based on the model complementary effects breaks through the limitations of a single model and further exploits the complementary advantages of multiple models.The experimental results show that by combining FEIRF,FSACE and ASG,the advantages of multi-source remote sensing data can be better utilized and the estimation model can be more accurate and stable.