Classification Of Forest Types Based On Multi-dimensional Features Using Multi-seasonal Landsat-8 OLI Remote Sensing Images

Rapid and accurate remote sensing forest mapping provides technical support for real-time monitoring forest resources change,spatial allocation of forest resources,forest resources optimization adjustment and auxiliary decision-making.Many studies have found that the difference between different vegetation spectral characteristic is small.The characteristics from single temporal image are difficult to effectively improve the vegetation classification precision.Compared to the information from space and spectral dimension or radiation dimension,the temporal information can provide much more land cover and status information,so phenological information of vegetation types can be effectively used to distinguish vegetation types and improve the classification accuracy of vegetation.In this paper,the time series Landsat-8 OLI remote sensing images in one year were used for vegetation classification.Firstly,the temporal sequence data were precisely preprocessed to obtain the reflectance data,and then the original spectral,vegetation index and texture characteristics of different vegetation types were extracted.By the feature analysis of random forest(RF),the importance scores for all features were calculated for mono-temporal image and multi-temporal image combinations.According to the sorting results of the features,the improved transferred divergence(TD)method was adopted to determine the optimal feature number and characteristics variables of the vegetation classification for mono-temporal image or multi-temporal image combinations.The hierarchical classification method was adopted to classify.Firstly,the study area was divided into vegetation types and non-vegetation types using spectral features and NDVI.Second,the vegetation types were classified using different classification method including decision tree,support vector machine and random forest.And the non-vegetation types are classified by threshold.This paper draws the following conclusions:(1)The time series images were precisely topographic corrected by the corresponding correction models.The calibration images were used to extract spectrum and texture features including GLCM texture and LBP texture.Through the analysis of the mean spectral value of vegetation types and the boxplot of the spectra of different vegetation types,the OLI-133(May 13,2015)image was found to have more abundant vegetation differentiation information.And the vegetation classes are easier to distinguish between the NIR band(band5)and the SWIR(band6 and band7),especially for short-wave infrared band.But the spectra of the vegetation categories are overlapped.(2)To run a RF classifier,two parameters have been set:the number of classification trees(ntree)and the number of input variables(mtry)used at each node.Based on sample prediction error and precision analysis,we chose the 1000 trees and 7 variables at each node.The parameter of decreased accuracy from RF model was used to measure the importance of the features.The importance of the original band spectrum,vegetation indices and texture features were separately analyzed for each images and multi-image combinations.The analysis of band spectrum shows that near-infrared band(band5)or short-wave infrared band(band6)for each image has rich information for distinguishing vegetation types.For analysis of multi-image combinations,the spectral features in the OLI-133(May 13,2015)image and OLI-191(July 10,2013)image have more abundant vegetation information.When the time series was longer than three,the top ten features remained stable which is almost band spectra features of OLI-133(May 13,2015)and OLI-191.Especially,the score for near infrared band(band5)of OLI-191(July 10,2013)image is significantly higher than that of any other images band spectral characteristics.The sorting result analysis of vegetation index for each image found that the normalized difference vegetation index has relatively higher importance score and their mathematical model almost contain near infrared(band5)or infrared wavelengths(band6 and band7);for multi-image combinations,the vegetation index in the day 133 and 191 images is the most important feature.(3)The sorting results of different feature based on RF model were analyzed by improved transferred divergence(TD)method,and then the number of features and the feature variables of the participating classification for single image and multi-image combinations were determined.For single temporal image,the vegetation indices and texture features were selected among sorted features form RF.The separability analysis for each pair of vegetation classes shows that the vegetation indices and texture features selected maximize the separability of vegetation classes that compared to original spectral features,and so feature extraction is important for improving the vegetation classification.For multi-image combinations(OLI-034-133-191-290-191:February-May-July-October-December for OLI image),we choose top twelve vegetation indices and ten texture features from the sorted features based on time series images.The separable values of interclass can be less than 1 by TD analysis of the mono-temporal features,and the separable values of the interclass can be more than 1.5 for multi-temperal features.So,the result from the TD analysis between each pair of vegetation classes found that the use of the temporal characteristics can effectively improved vegetation class separability.The overall accuracy from multi-temperal features was 84.7%.Compared with the result from mono-temporal feature overall classification precision was improved by 15.4%.Especially,the classification accuracy of pure stand of Chinese fir,masson pine and bamboo were improved greatly,with the increase of 25.44%and 15.29%19.54%,respectively.So,the addition of time phase features can improve the recognition accuracy of tree species.In the paper,the improved transformed divergence(TD)algorithm considering samples prior probability of different classes is proposed to identify the optimal number of features and features variables.The best feature number and characteristics variables participating classification were determined fastly from mass characteristics by the improved TD analysis and the decreased accuracy of RF model.It is a feasible scheme for fast and accurate forest mapping in the study area by comparing the classification result.