Abstract:

Under the background of carbon neutrality, accurate estimation of carbon storage in terrestrial ecosystems is of great significance for assessing the current carbon sink capacity and future potential of China's terrestrial ecosystems. Remote sensing imagery offers the advantage of large-area continuous sampling, and quantitative statistics based on remote sensing technology for forest, grassland, agricultural, wetland, and unused land carbon storage are fundamental for accurately accounting for terrestrial ecosystem carbon storage. A single remote sensing method cannot simultaneously acquire both horizontal and vertical structural information of surface cover. The application of optical and synthetic aperture radar (SAR) data fusion for estimating carbon storage has developed rapidly. The high-resolution characteristics of optical imagery make the scale differences among surface features such as forests, grasslands, farmland, and traffic systems more significant, and wavelet packet analysis is an effective way to improve signal-to-noise ratio and extract multi-scale features. SAR imagery has the advantage of strong penetrability but suffers from noise interference. GF-1 multispectral imagery has a wide coverage and high spatial and temporal resolution. After wavelet packet transform, the features at various levels of wavelet packet decomposition have good transferability, allowing for the extraction of structured multi-direction and multi-scale texture features of GF-1, which is conducive to fine differentiation of ground cover categories. Sentinel-1 satellite’s dual-polarization information (VH, VV) has penetrability and can obtain vertical structural information, unaffected by clouds and fog. Therefore, this study chooses to integrate the advantages of GF-1 wavelet packet texture, spectrum, and Sentinel-1 structural features, combined with the Random Forest and InVEST model for land use classification and carbon storage estimation, aiming to explore a better method for fine-scale regional carbon storage estimation. The first step is to select high-quality GF-1 and Sentinel-1 remote sensing images for preprocessing, registering the two preprocessed images using the Forstner algorithm, constructing feature descriptors for optical and SAR images based on phase consistency, and considering the rotation and scale invariance of the features. The second step involves extracting the Normalized Difference Vegetation Index (NDVI) from GF-1 imagery. Simultaneously, principal component analysis is performed on GF-1 to remove redundant information, and the principal component imagery is selected for wavelet packet transform. The optimal wavelet packet tree is selected based on a cost function, considering the correlation and transferability of information at various scales of the wavelet packet transform to ensure the acquisition of more detailed anisotropic features of ground cover. The extracted multi-scale texture features include the mean (M), variance (V), and energy (E) of high-frequency coefficient statistics in horizontal (H), vertical (V), and diagonal (D) directions, while multi-scale spectral features(S) are derived from low-frequency coefficients. The third step selects Sentinel-1's backscatter coefficients VH, VV, (VH-VV)/(VH+VV), and VV/VH as remote sensing feature factors, combined with GF-1 spectral features, wavelet packet texture features, and NDVI features to construct classification vectors. The fourth step involves constructing a random forest classifier, selecting different samples for multiple experiments based on a method of random diffusion from scattered points, ranking features based on their importance, selecting the most influential features to construct the final land use classification vector, and determining the number of decision trees based on classification accuracy to complete the final land use classification. The fifth step involves using the InVEST model's carbon storage module to estimate regional ecosystem carbon storage based on the fine ground classification. The carbon densities required by the InVEST model for aboveground biomass carbon, belowground biomass carbon, soil carbon, and dead organic carbon were obtained from relevant literature in the study area (such as the Forestry Science Data Center, National Ecosystem Science Data Center, and other research literature). Additionally, carbon densities are adjusted based on research results from other regions within the same climatic zone. The sixth step involves assessing the error of the estimated carbon storage in this study using the 2020 land use type survey results and the carbon density obtained from field surveys from the study area as references. The results show that: 1) After integrating GF-1 spectral and wavelet packet texture with Sentinel-1 structural features, the overall accuracy of each land use classification reaches 92.46%, with a Kappa coefficient of 0.91, and the estimated carbon storage value is 31.27 Tg, with an overall error of 0.81% compared to the measured carbon storage value. 2) After classification, the proportions of forest land, residential areas, water bodies, arable land, roads, gardens, and other unused land are 83.29%, 2.32%, 1.38%, 4.44%, 1.29%, 6.42%, and 0.86%, respectively, and their corresponding contributions to carbon storage are 91.53%, 0.03%, 0.08%, 2.77%, 0.01%, 5.50%, and 0.08%, respectively. 3) The error of this study's method for estimating carbon storage is reduced by 3.09% compared to using only GF-1 spectral features. 4) Compared to other carbon storage estimation models such as the Random Forest model and the K-Nearest Neighbors Regression model, this study's method reduces carbon storage estimation errors by 2.03% and 2.51%, respectively. Additional findings of this study include: To improve the accuracy of multi-source remote sensing data fusion, attention can be given to the interaction of multi-modal features of multi-source images based on wavelet transform. Higher-resolution GF-1 imagery can be used to extract spectral, texture, and biophysical information to construct biomass estimation models, and multi-dimensional SAR can be utilized to estimate forest structural parameters such as canopy height, tree crown volume, stand basal area, canopy closure, and leaf area index for fine-grained tree species identification. The combination of GF-1 and SAR enables precise estimation and stratification of carbon storage. Furthermore, regional ecological data such as regional climate, forest growth and population changes from mechanistic ecological models can be integrated to construct dynamic and spatially continuous carbon storage estimation models. Overall, compared to previously proposed remote sensing carbon storage estimation methods, this study's method takes digital image processing techniques, multi-scale analysis methods, and Random Forest classification as the main lines, combined with the InVEST model to achieve intelligent carbon storage estimation. It can rapidly calculate large-area regional biomass and facilitate carbon storage statistics in deeply cut mountain gorge areas difficult to access by technicians, providing a reference for subsequent carbon sequestration and sink enhancement efforts.

Key words: land use classification, carbon storage estimation, wavelet packet transform, integration of GF-1, Sentinel-1

摘要：

碳中和背景下陆地生态系统碳储量的精细化估测具有重要意义。为更精确地利用多源遥感影像调查陆地生态系统碳储量情况，该文提出融合高分1号（GF-1）多尺度特征及哨兵1号（Sentinel-1）结构特征的土地利用分类及碳储量估测方法。对GF-1经特征向量主成分分析的主分量影像进行小波包变换，依据代价函数选出最优小波包树，提取多尺度特征，再结合植被指数特征、Sentinel-1结构特征构造分类向量，利用随机森林（RF）筛选特征完成土地利用分类，最后基于InVEST模型估测碳储量。结果表明：1）GF-1光谱及小波包纹理融合Sentinel-1结构特征后，各土地利用分类总体精度可达92.46%，Kappa为0.91，估测的碳储量值为31.27 Tg，与实测碳储量值相比总体误差为0.81%；2）分类后各土地利用类型中，林地、住宅、水域、耕地、道路、园地、其他未利用土地占比分别为83.29%、2.32%、1.38%、4.44%、1.29%、6.42%、0.86%，对应的碳储量贡献占比分别为91.53%、0.03%、0.08%、2.77%、0.01%、5.50%、0.08%；3）该文方法较仅用GF-1光谱特征估测碳储量的误差降低了3.09%。可见，光学影像经最优小波包变换提取多尺度特征，再结合Sentinel-1的垂直特征，细化了地物分类，提高了碳储量的估测精度。该方法可为碳储量的遥感估测提供借鉴。

关键词: 土地利用分类, 碳储量估测, 小波包变换, 融合高分1号（GF-1）, 哨兵1号（Sentinel-1）