基于光谱和改进极限学习机的土壤含盐量分析

研究盐碱地的性质、组成，对于生态环境具有重要意义。传统的含盐量测定方法大多基于化学分析，因其成本高、效率低的缺点使得应用于大面积土地的可行性很低。极限学习机（ELM）作为一种基于前馈神经网络构建的机器学习系统，在许多研究中作为一种光谱处理方法被成功运用。为了改进传统的盐碱地含盐量检测方法，采用光谱学结合改进的极限学习机（ELM）模型的方法对盐碱地进行研究。根据镇赉县采集得到的62个土壤表层样本得到对应的光谱反射率和含盐量数据，提出了基于随机值改进粒子群优化算法（RVIPSO-MELM）优化的多层极限学习机模型。首先使用主成分分析提取光谱数据特征，并使用ELM算法对光谱数据建立分类模型，引入改进的粒子群优化算法以提高精度和速度。该模型结合了具有随机值的多层ELM（RV-MELM）和改进PSO算法的多层ELM（IPSO-MELM）二者的优点，在运用启发式算法搜索最优值的同时还具有随机性，提高了模型优化速度，同时以提高模型性能为目的对隐含层之间参数进行优化和选择。并且该模型可以推广到多层，对隐含层之间的参数的两种选择方法，根据经验公式计算和使用改进的启发式算法搜索，进行了提高模型性能和优化时间的讨论，实践结果表明，选择第一层参数使用改进粒子群优化算法，确定随后的隐含层之间参数选择，使用经验公式进行计算得到一种更具现实意义的方法模型。模型在进行启发式搜索最优值之前，利用蒙特卡罗方法确定一个较好的初值，使得模型能保持较高准确率的条件下，优化速度进一步提高。相比于传统方法，这种光谱分析结合ELM的模型节省时间和经济成本，有一定推广意义。

Analysis of Soil Salinity Based on Spectrum and RVIPSO-MELM

Studying, the nature and composition of saline soil, are significant to the ecological environment. Most traditional methods for determining salt content are based on chemical analysis. Due to their high cost and low efficiency, the feasibility of applying them to large areas of land is greatly reduced. An extreme learning machine (ELM), as a machine learning system based on a feedforward neural network, has been successfully used as a spectral processing method in many studies. In order to improve the traditional salt content detection methods of saline-alkali soils, this paper uses spectroscopy combined with an improved extreme learning machine (ELM) model to study saline-alkali soils, further expanding the application scenarios of spectroscopy analysis methods. We obtain the corresponding spectral reflectance and salt content data according to the 62 surface samples collected in Zhenlai County and then propose the multi-layer extreme learning machine model optimized by improved particle swarm optimization (PSO) algorithm with improved particle swarm optimization (PSO) algorithm with random values(RVIPSO-MELM) model. Firstly, we use the principal component analysis(PCA) to extract the characteristics of the spectral data and then adopt the ELM algorithm to establish a classification model for the spectral data. Finally, to improve the accuracy and speed, an improved particle swarm optimization algorithm is applied. This model combines the advantages of both multi-layer ELM with random values (RV-MELM) and the multi-layer ELM model optimized by an improved PSO algorithm (IPSO-MELM), using the heuristic algorithm to search for the optimal value and also having randomness, which improves the speed of model optimization. The parameters are optimized and selected to improve the performance of the model. Moreover, the model can be extended to multiple layers, and the two methods of selecting parameters between hidden layers, calculated by empirical formulas or improved heuristic algorithm, are discussed about the model’s performance and optimize the time. The practical results show that it is a more realistic method to select the parameters of the first layer to use the improved particle swarm optimization algorithm and determine the parameters of the subsequent hidden layers by using the empirical formula calculate. Before the heuristic search for the optimal value, the Monte Carlo method is applied to determine a better initial value, enabling the model to maintain a high accuracy rate and further improving the optimization speed. Compared with traditional methods, this spectral analysis combined with the ELM model saves time and economic costs, giving it a certain promotion significance.