关键词: 碳储量估计, 兴安落叶松, 图像理解, 图像分类, 哑变量

Abstract:

Objective: Traditional forest carbon stock estimation is considered as a heavy workload and high cost in labor and financial resources. In this study, a simple and accurate algorithm was proposed for forest carbon stock estimation. Method: The vertical images of the stand contain two kinds of information(density and height), which is related to the forest carbon storage directly. Firstly, a tree identification algorithm was proposed. Based on the global threshold, the classification of pixels was determined by combining the feature of neighborhood pixels which function is to weaken the influences of uneven light on the gray level of the image. Then, a parameter which is closely related to the forest carbon stock and calculation method was proposed. Finally, the parameter was used as input in a prediction model to estimate carbon stock. Result: The results showed that if there were more than 6 similar pixels in the neighborhood of 3×3 centered at the central pixel, the pixel was also classified as the same category as the similar pixels. The method which used the relationships with adjacent pixels to determine the classification of the current pixel has the characteristics of expansion and corrosion. Therefore, this algorithm can increase the accuracy of forest image classification, which was better than the simple global threshold method. As for the prediction model of carbon stock, the accuracy of the two-parameter linear equation estimation was close to the logistic model of three-parameters and the Logistic model. If the virtual variable which represents the altitude was added into the ordinary model, the prediction accuracy of the carbon reserves can be improved greatly. Taking Larix gmelinii as an example, the hypothesis that carbon stock per unit area can be accurately estimated based on the forest vertical image is verified. Conclusion: This study improved the classification accuracy of the symmetric cross-entropy algorithm by taking into account the category of neighbor pixels. This improved algorithm could achieve higher classification accuracy by showing robust performances to the changes of illumination. In terms of the carbon storage prediction model based on image data, the logistic curve might show a good adaptability. Considering the altitude factor could reduce the estimation error, it is necessary to estimate the carbon storage at different altitudes in practical application.

Key words: carbon storage estimation, Larix gmelinii, image understanding, image classification, dummy variable