Research On Forest Classification Based On Super Pixel Segmentation And Its Application In Carbon Storage Estimation

Forest is the largest carbon pool on land in several ecosystems on the earth's surface.As the largest main body on land,it absorbs about one twelfth of CO₂ in the earth's atmosphere.Therefore,forest has become a key participant in the global carbon cycle.With the rapid development of remote sensing technology,the application of high resolution remote sensing image in forestry is more and more extensive.These images provide rich and complex information on the earth's surface,including not only abundant spatial information,but also clear geometric and texture information.Estimation of forest carbon reserves is to establish an estimation model based on the survey data of the fixed sample plots in the monitoring area,and to estimate the carbon reserves of the forest areas in the land cover type to obtain the carbon storage distribution map.Firstly,forest areas are classified and distinguished by high resolution remote sensing images of monitoring areas,and then the carbon reserves of forests are estimated.In this paper,the above-ground forest carbon storage in Nanan District of Chongqing City was studied from three aspects.Firstly,the image is segmented based on simple linear iterative clustering?SLIC super-pixel segmentation?,and the best segmentation object is obtained.Secondly,the land cover types are extracted by decision tree classification.Finally,combining with historical field survey data,the estimation model of extreme learning machine is established to estimate forest carbon reserves in the study area.?1?The experimental data in this research area are 8m image data with high score No.1.The image is segmented by using SLIC super-pixel segmentation algorithm to obtain the best scale segmentation parameters which overlap with the image edge.The segmentation results are merged by region adjacency graph to form multi-level images.?2?In view of the shortcomings of traditional remote sensing image classification methods based on spectral information,the feature utilization of high-resolution remote sensing images is insufficient.In this paper,expert knowledge CART decision tree method is used to classify land cover in high-resolution remote sensing images.The distribution information of forest vegetation in Nanan District of Chongqing City is extracted by utilizing a variety of features?such as texture,vegetation index,water body index,etc.?.Compared with the classification results of maximum likelihood method based on spectral information in ENVI,the results show that the CART expert knowledge decision tree classification method with multi-image features has higher accuracy.?3?Carbon storage estimation models of forest?broad-leaved forest and coniferous forest?in sample plots were constructed by combining classification results with historical survey data.Extreme learning machine with non-linear fitting characteristics was used in the model.In view of the differences between coniferous forest and broad-leaved forest in remote sensing image characteristics,this paper selects vegetation-related vegetation index,terrain and image texture features as independent variables,and constructs carbon storage estimation model with carbon storage as dependent variable,which further improves the estimation performance of the model.The main conclusions are as follows:?1?Super-pixel segmentation based on SLIC is adopted in this paper.The boundary of segmentation and the type of object are maintained well,and the advantages of contour preservation are obvious.At the same time,it combines with the regional adjacency map,combines the results of over-segmentation,forms a multi-level classification system,and classifies the land cover in the study area by decision tree.The results show that the total classification accuracy is 84.7578%,and the overall Kappa coefficient is 0.8071,which is higher than the maximum likelihood method based on pixels,and the classification effect is better.?2?According to the forest distribution in the classification results,the extreme learning machine was selected to construct Carbon Storage Estimation Models for broad-leaved forest and coniferous forest respectively.The results show that the accuracy of the model is higher than that of the traditional carbon storage model based on BP neural network,and the spatial distribution of forest vegetation carbon storage in the south coast area is basically the same as that of land cover.The higher the vegetation coverage,the greater the carbon storage.