土壤重金属污染是由于人类活动导致重金属物质大量残留在土壤中, 超过土壤环境承载力, 这种现象将造成土壤质量退化、 生态环境恶化。 高光谱遥感可以实现图谱合一, 能有效地识别出土壤中不同元素的异常情况。 为实现农田土壤重金属高效、 准确监测, 提出了一种特征提高型竞争性自适应重加权算法（CARS）选取特征波段的粒子群算法（PSO）优化支持向量机（SVM）农田土壤重金属砷(As)含量高光谱估测分析方法。 利用CARS对暗室实测光谱值进行粗选； 利用一阶导数（FD）、 高斯滤波(GF)、 归一化（N）进行特征提高； 在特征精选阶段利用皮尔逊相关系数（PCC）求取预处理后的光谱指标与土壤重金属As之间的相关系数, 获取相关性大于0.6的波段作为特征波段； 最后利用PSO对SVM所选择的核函数σ和正则化参数γ进行优化, 以均方根误差(RMSE)作为适应度函数, 通过迭代最优适应度得到SVM最优参数值。 选择江汉平原典型区域洪湖市燕窝镇的土壤为研究对象, 预测结果表明基于PSO-SVM算法其验证集的决定系数R2为0.982 3, 均方根误差RMSE为0.521 6, 平均绝对误差MAE为0.416 4。 主要结论如下： PSO算法优化SVM参数, 通过迭代更新个体极值和群体极值, 可以迅速获取全局最优解, 与支持向量机回归（SVMR）和随机森林回归（RFR）相比, 在预测精度有了较大的提高； 特征提高型CARS算法可以有效剔除无关信息, 提高相关性, 且选取波段少, 模型简单, 大大提高了效率； 可以实现土壤污染预警、 满足精准农业需求、 为后期重金属污染土地生态修复提供数据基础。

Heavy metal pollution in soil is caused by human activity factors that bring heavy metals into the soil, resulting in deterioration of soil quality and ecological environment. Heavy metals in the soil tend to accumulate, are difficult to be degraded, are highly concealed for long periods of time, and can be enriched by atmospheric circulation and food chains, ultimately threatening human life and health. Hyperspectral remote sensing technology presents a combination of image and spectrum, and can effectively identify the abnormal conditions of different elements in the soil. At present, traditional soil monitoring techniques mainly rely on laboratory-based chemical detection methods such as photometry, chemical analysis, and atomic fluorescence spectroscopy. This kind of method can test the heavy metal content of farmland soil, but the precision depends on a large amount of manpower, material resources and equipment, and its detection efficiency and promotion are lacking. In order to achieve efficient and accurate monitoring of heavy metals in farmland soils. A method of hyperspectral estimation of heavy metal arsenic (As) content in farmland soils based on particle swarm optimization (PSO) and support vector machine (SVM), which use characteristic-enhanced competitive adaptive reweighted sampling (CARS) was proposed. In the characteristic rough selection stage, the measured spectral values from the darkroom are roughly selected by CARS. In the characteristic improvement stage, First Derivative (FD), Gaussian Filtering (GF), Normalization (N) are used to improve features. In the carefully chosen stage, Pearson Correlation Coefficient (PCC) is used to obtain the correlation coefficient between different pre-treated spectral indices and soil heavy metal As. The band whose correlation coefficient has an absolute value greater than 0.6 is selected as a feature band. Finally, PSO is used to optimize the kernel parameter sigma and the normalization parameter gamma used by the SVM. The root mean square error (RMSE) is used as the fitness function, and the optimal parameters of SVM are obtained by iterating the optimal fitness. The soil of Yanwo Town in Honghu City, a typical area of Jianghan plain, was selected as the research object in this paper. The prediction results showed that the decision coefficient (R2) of the verification sets based on PSO-SVM algorithm is 0.982 3, the root mean square error (RMSE) is 0.521 6, and the mean absolute error (MAE) is 0.416 4. The main conclusions are as follows: the PSO algorithm is used to optimize the SVM parameters, and the global optimal solution can be obtained quickly by iteratively updating the individual extremum and the group extremum. Compared with the support vector machine regression (SVMR) and random forests regression (RFR), the prediction accuracy has been greatly improved; The characteristic enhanced CARS algorithm can effectively eliminate irrelevant information and improve correlation. And it selects fewer bands, simplifies the model so that efficiency is greatly improved; It can realize early warning of soil pollution, meet the needs of precision agriculture and provide data basis for ecological restoration of heavy metal contaminated land in the later period.