【目的】 森林冠层密度与林分年龄、植被生长状况有关,在区域森林净初级生产力遥感估测中,结合森林冠层密度以期提高估测精度。【方法】 以广东省韶关市为研究对象,选用2017年Landsat-8 OLI影像、2017年357块森林资源连续清查固定样地数据为主要信息源,分别采用随机森林、多元线性回归、人工神经网络和K最近邻分类法等4种模型,结合森林冠层密度制图器(FCD)进行区域森林净初级生产力特征变量的选取、参数建模、模型精度评价和森林净初级生产力空间制图。【结果】 特征变量中,红光波段(B4)、归一化植被指数(NDVI)、比值植被指数(RVI)、叶面积指数(LAI)、缨帽变换土壤植被因子、纹理特征和地形特征在森林净初级生产力反演中有重要作用。将森林冠层密度因子加入反演模型后,4种遥感估测模型精度均有大幅度提高。对4种遥感估测模型进行性能比较,随机森林模型精度最高,其次是多元线性回归模型、人工神经网络模型,K-最近邻分类模型精度最低。研究区内森林净初级生产力平均值为10.689 t/(hm²·a),高森林净初级生产力 [≥18 t/(hm²·a)]林分面积仅占研究区的19.61%,主要分布在海拔较高的西北部。【结论】 结合冠层密度进行森林净初级生产力的建模,可有效提高模型估测精度。

【Objective】 Forest canopy density is related to stand age and vegetation growth status. In the remote sensing estimation of regional forest net primary productivity, combining forest canopy density may substantially improve the estimation accuracy. 【Method】 In this study, Shaoguan City in Guangdong Province was used as the case study area; the Landsat-8 OLI images and the fixed sample plot data of 357 forest resources continuous inventory in 2017 were collected as the main information sources. Using four models of random forest the multiple linear regression, artificial neural network and K-nearest neighbor classification methods, and forest canopy density, the selection of modeling variables, the evaluation of model accuracy, and the creation of a spatial map of net primary productivity of regional forest vegetation were performed. 【Result】 The results showed that among the characteristic variables, red band (B4), normalized vegetation index (NDVI), ratio vegetation index (RVI), leaf area index (LAI), soil vegetation factor of tassel cap transformation, texture characteristics, and terrain characteristics played important roles in the prediction of vegetation net primary productivity. After adding the forest canopy density factor into the inversion model, the accuracy of the four remote sensing estimation models was substantially improved. The performance comparison of the four remote sensing-based models indicated that the random forest model had the highest accuracy, followed by the multiple linear regression model and artificial neural network model; the K-nearest neighbor classification model had the lowest accuracy. The average net primary productivity of forests in the study area was 10.689 t/(hm²·a), and the area of high forest net primary productivity (≥18 t/(hm²·a)) accounted for 19.61% of the total area in the study area, which was mainly distributed in the northwest region at higher altitudes. 【Conclusion】 In this study, at the regional scale, the model prediction accuracy of forest net primary productivity can be effectively improved by combining canopy density.