叶面积指数(Leaf Area Index, LAI)是表征植被冠层结构的基本参数之一，是作物病虫灾害及长势监测、产量预测、气候和生态模型等的重要指标，因此地表植被LAI的估算对农业生产、全球气候变化和生态平衡等方面的研究具有十分重要的意义。大面积、快速有效的LAI估算主要依靠遥感技术手段，植被指数作为地表植被状况的一种简单度量方式，被广泛应用于各类植被LAI的遥感估算研究。支持向量机回归(Support Vector Regression, SVR)方法是叶面积指数估算的一种有效手段，在LAI估算中具有一定的应用价值和前景，但SVR算法中惩罚系数C、核函数宽度参数g、不敏感损失函数参数ε的取值对模型精度有显著的影响，如果不根据实际问题设置合适参数，将显著影响SVR模型精度。智能算法如遗传算法(Genetic algorithm, GA)、粒子群算法(Particle Swarm Optimization, PSO)、人工蜂群算法(Artificial Bee Colony, ABC)等在参数优化方面具有较强的并行处理能力，且寻优速度快，具有全局寻优等特点，是当前热门的参数优化算法。因此，为了测试智能算法优化SVR关键参数提高LAI估算精度的能力，本文基于中国河北省张北县和美国爱荷华两个研究区2002年7月至8月LAI地面实测数据和同期Landsat-7 ETM+地表反射率数据，提取NDVI、RVI、DVI、PVI、SAVI、MSAVI 6类植被指数作为自变量，LAI地面实测值为因变量构建6个样本集，并随机抽取每个样本集的70%作为训练集，剩下的30%作为测试集，然后基于训练集建立3类LAI估算模型：（1）传统回归模型：指数模型、线性模型、对数模型；（2）默认参数的SVR模型：C=1、g=1/k，k为输入自变量个数、ε=0.1；（3）智能算法（GA、PSO、ABC）优化参数的SVR模型：GA-SVR、PSO-SVR、ABC-SVR。最后以R²、RMSE作为模型回归拟合精度指标，比较上述三类模型优劣，并进一步从多角度（运行效率、收敛速度和优化参数后SVR模型精度的差异显著性）对比不同智能算法对SVR模型估算LAI的优化能力。研究结果表明：

（1）默认参数的SVR模型较传统回归拟合模型具有更佳模型预测效果。对于河北省张北县其默认参数SVR模型的RMSE低于最优传统回归模型15%-25%，对于美国爱荷华默认参数SVR模型RMSE低于传统回归模型50%-60%。且对于不同地表植被覆盖信息，默认参数SVR模型都优于传统回归模型，说明SVR方法具有较好的鲁棒性。

（2）智能算法同时优化三个参数(C、g、ε)的GA-SVR、PSO-SVR、ABC-SVR模型较默认参数和仅优化两个参数(C、g)的SVR模型具有更佳模型性能。在95%的置信区间下，GA-SVR、PSO-SVR、ABC-SVR模型与默认参数SVR模型之间精度差异具有显著性(P<<0.001)，智能算法极显著地改善SVR估算模型精度。两个研究区优化三个参数(C、g、ε)的SVR模型较优化两个参数(C、g)的SVR模型具有更高决定系数和更低RMSE，且其回归直线斜率k更接近于1，张北县和爱荷华分别提高5%-10%和2%-10%，偏差p更接近于0，张北县和爱荷华(除了PSO外)减少5%-20%和20%-35%。

（3）3种智能算法中GA算法具有较高寻优效率和收敛速度，是两个研究区最理想的参数优化算法。在95%置信区间下，PSO-SVR、GA-SVR、ABC-SVR模型精度无显著性差异，但GA算法的运行时间和收敛次数与PSO、ABC算法具有显著性差异，表明GA、ABC、PSO优化算法对SVR模型精度改善没有明显差别，但GA算法寻优效率和收敛速度显著高于PSO和ABC算法。

（4）不同研究区最适于LAI估算的植被指数不同。对于河北省张北县研究区，植被指数MSAVI是LAI估算建模的最优指数。而对于美国爱荷华研究区，植被指数DVI是LAI估算建模的最优指数。

Leaf area index (LAI) is one of the basic parameters characterizing the vegetation canopy structure, and it is an important index for crop diseases and insect pests monitoring, growth monitoring, yield prediction, climate and ecological model, etc. Therefore, the LAI estimation of surface vegetation is of great significance for agricultural production, global climate change and ecological balance. Fast and effective LAI estimation over large area mainly depends on remote sensing technology. And as a simple measure of vegetation condition, vegetation index is widely used in estimation of LAI with remote sensing for various vegetation. Support vector regression (SVR) as an effective method to estimate LAI, which has great application value and prospect. However, the value of penalty coefficient C, width parameter g of kernel function and insensitive loss function parameter ε in the SVR algorithm have a significant impact on regression accuracy of the model. If the parameters are not set according to the actual problems, the accuracy of SVR model will be significantly affected. Artificial intelligence algorithms have the characteristics of strong parallel processing ability in parameter optimization, fast speed, optimization and so on, such as genetic algorithm (GA), particle swarm optimization (PSO) and artificial bee colony (ABC), which are the popular parameter optimization algorithms. Therefore, in order to test the ability of AI algorithm to optimize the key parameters of SVR and improve the accuracy of LAI estimation this paper based on the LAI measurements and Landsat-7 ETM + surface reflectance data from July to August 2002 in Zhangbei County, Hebei Province, China and Iowa, USA, extracted six vegetation indices: NDVI, RVI, DVI, PVI, SAVI,MSAVI as independent variables, and LAI measurements as dependent variable, constructed six sample sets, the training set was 70% of the sample set and the test set was 30% of the sample set, then based on the training set, three types of LAI estimation models are established: (1)Traditional regression models: exponential model, linear model and logarithmic model;(2)SVR model with default parameters: C = 1, g = 1 / k, k is the number of input independent variables, and ε = 0.1;(3)SVR model with artificial intelligence algorithm(GA, PSO, ABC)optimizing three parameters:GA-SVR, PSO-SVR , ABC-SVR. Finally, R² and RMSE were used as the model fitting accuracy indicators to compare the advantages and disadvantages of the three models. Furthermore, the optimization ability of different intelligent algorithms to SVR model for estimating LAI is compared from multiple perspectives (operation efficiency, convergence speed and the difference of accuracy of SVR model after optimizing parameters).The results suggest that:

(1) The SVR models with default parameters have better prediction effect than the traditional regression fitting models. For Zhangbei County of Hebei Province, the RMSE of the SVR models with default parameters is 15% - 25% lower than that the optimal traditional regression models. And for Iowa, the RMSE of SVR models with default parameters is 50% - 60% lower than that of the traditional regression model. It is worth noting that the SVR models with default parameters is better than the traditional regression model for different vegetation cover information, which shows that the SVR method has good robustness.

(2) GA-SVR, PSO-SVR and ABC-SVR models with three parameters (C, g andε) optimized by intelligent algorithm simultaneously have better model performance than the  SVR models with default parameters. And at the confidence interval of 95%, There were significant differences in accuracy between GA-SVR, PSO-SVR, ABC-SVR models and SVR models with default parameters(P<<0.001), it indicated that intelligent algorithm improved the accuracy of SVR estimation model significantly. And compared with the SVR model optimized two parameters (C, g), the SVR model optimized three parameters (C, g,ε) in two study areas has higher coefficient of determination and lower RMSE, in which the regression linear slope k is closer to 1, Zhangbei County and Iowa are increased by 5% - 10% and 2% - 10% respectively, the deviation p is closer to 0, and Zhangbei County and Iowa (except PSO) are reduced by 5% - 20% and 20% - 35%.

(3) GA algorithm has higher optimization efficiency and convergence speed among the three intelligent optimization algorithms, which is the most suitable algorithm for the SVR parameters optimization in the two research areas. Under the 95% confidence interval, there is no significant difference in the accuracy of PSO-SVR, GA-SVR and ABC-SVR models, but the running time and convergence times of GA algorithm are significantly different from PSO and ABC algorithms, which indicates that GA, ABC and PSO algorithms have no significant difference in improving the accuracy of SVR models, but the optimization efficiency and convergence speed of GA algorithm are significantly higher than those of PSO and ABC algorithm.

(4) In different study areas, the most suitable vegetation indexes for LAI estimation is totally different. For Zhangbei County of Hebei Province, the vegetation index MSAVI is the best index for LAI estimation and modeling. Similarly, for Iowa, USA. The DVI is the best index for LAI estimation and modeling.