Abstract: [Objective] Remote sensing technique provides a highly effective means for extracting tree species (group) spatial distribution information. The objective of this paper is to develop a method for estimating the provincial spatial distribution information of forest tree species (group) composition using multi-sources data. Thus it could indicate the spatial distribution information of the main tree species (group) and provide a new method for extracting vegetation information in large area. [Method] The experiments were carried out over the test site of the whole Jilin Province. The time series MODIS NDVI product of 250 m pixel size and 8 days cloudy free composite and the permanent forest plot data collected by the national forest inventory (NFI) were used as the key data sources. The weather observation data and topography data were also integrated into the data sources. We developed a gradient nearest neighbor (GNN) based approach for estimating provincial forest tree species (group) composition distribution information. Firstly, the method of canonical correspondence analysis (CCA) was implemented to extract effective composited features from the original dataset. Secondly, the k-nearest neighbors (k-NN) method was applied on the extracted feature space to estimate forest tree species (group) composition number using one two-layer stratification scheme. As the value of k needs to be determined, the changing trend of k-NN estimation accuracy with the k values was analyzed. Finally, the estimation accuracy for each tree species (group) of the developed method was validated using the forest plot data of 9 counties collected by the forest resources inventory in second level and the forest plot data collected by the NFI as reference. [Result] 7 tree species (group) composition numbers including Quercus mongolica,Betula platyphylla,Tilia amurensis,Ulmus davidiana,Populus,Juglans mandshurica and Larix olgensis were extracted and the corresponding distribution maps were produced. The results showed a good consistency with the fixed plots in field. Taking county as statistic unit, the following quantitative technical targets have been achieved:the coefficient of determination (R²) was 0.83, and the RMSE was 0.34. Specifically, the accuracy has been further validated by dividing the whole coverage of Jilin Province into grids of 20 km×20 km,30 km×30 km,40 km×40 km and 50 km×50 km, taking the forest plot data collected by the NFI as reference and the grid as statistic unit. Better results could be achieved at the scale of 40 km×40 km and 50 km×50 km. The RMSE of Ulmus davidiana composition number was 0.35 and the RMSE of Quercus mongolica composition number was 0.65. The optimal k-value could be determined for the phenomenon that the RMSE firstly reduced and then tended steady with the rising k-value. In addition, the estimation accuracy of the two-layer stratification estimation method was higher than that of the direct estimation method. The results showed that:the average RMSE of estimating tree species (group) composition using two-layer stratification estimation method was 0.1 less than that using direct estimation method.[Conclusion] The proposed method for estimating the provincial spatial distribution information of forest tree species (group) composition using multi-sources data has proved to be an effective method to estimate forest parameters. Based on this method, the distribution map of forest tree species (group) composition numbers was successfully produced with high accuracy. The results indicated that the value of k needs to be optimized in order to obtain a better result, which varies depending on the experimental area and the selected data. In addition, the estimation accuracy could be improved effectively using two-layer stratification estimation method.

Key words: multi-data sources, GNN, CCA, k-NN, MODIS NDVI, tree species composition, mapping