基于MA-ARIMA-GASVR的食用菌温室温度预测

食用菌温室温度具有时变、非线性、多耦合特性,准确预测对稳定食用菌生产具有重要意义。本研究从挖掘温室历史温度数据时序信息角度出发,提出一种MA-ARIMA-GASVR组合方法建立温度预测模型,利用移动平均方法将历史温度序列分解成线性序列和残差序列,然后采用移动平均差分自回归模预测线性序列的趋势,再将移动平均差分自回归预测值、历史残差数据、历史温度数据作为支持向量回归模型的输入,并结合遗传算法优化支持向量回归模型参数改善其性能,从而获得更符合实际情况的温度预测值。最后选取实测温度数据作为训练集,对未来2d的温度进行预测验证。结果显示,MA-ARIMA-GASVR组合方法能更好地拟合原始温度数据,间隔1h的均方误差、平均绝对误差和平均绝对百分误差分别为0.18、0.36和1.34,均显示本研究方法预测精度优于支持向量回归、遗传算法优化的支持向量回归单一模型,也优于未经移动平均以及未经遗传算法优化的组合模型;此外,间隔6h的均方误差、平均绝对误差和平均绝对百分误差为0.29、0.52和1.95,说明本研究方法还能满足6 h以内的多步预测,为食用菌生产者预留更多调整时间。     

基于机器学习算法的蒸发量模型评估

[目的]探讨合理的气候因子个数,建立蒸发量模型,提出基于特征选择算法筛选最优特征集。[方法]以陕西榆林、泾河和汉中3个气象站16年(2005-01至2021-03)的逐小时观测资料为研究对象,利用特征选择函数和遍历循环方法对模型参数、特征变量个数进行优化。基于最佳参数结合随机森林模型和多元线性回归模型两种机器学习算法建立榆林、汉中和泾河地区蒸发量模型,采用平均绝对误差、均方根误差和平方相关系数三项指标评估模型的预测精度。[结果]特征变量和随机森林模型中的决策树个数分别是8,61时,模型预测效果最佳。采用优化的随机森林模型、多元线性回归模型评估3个地区的平均绝对误差均为0,均方根误差除泾河地区相等外,榆林、汉中地区的均方根误差均小于优化的多元线性回归模型。优化的随机森林模型预测榆林、泾河和汉中地区蒸发量拟合效果分别为0.85,0.90,0.86,优化的多元线性回归模型的拟合效果分别为0.77,0.83,0.79。[结论]整体而言,优化的两种模型都具有良好的预测效果且随机森林模型的预测效果优于多元线性回归模型。     

基于多源环境变量和随机森林的橡胶园土壤全氮含量预测

土壤全氮与土壤肥力和土壤氮循环紧密相关。掌握土壤全氮详细的空间分布信息对提高土壤肥力管理效率和更好地了解土壤氮循环至关重要。该文以儋州国营橡胶园为研究区域,采集2511个土壤样品,利用随机森林(random forest,RF)、逐步线性回归(stepwise linear regression,SLR)、广义加性混合模型(generalized additive mixed model,GAMM)以及分类回归树(classification and regression tree,CART)结合多源环境变量(成土母质、平均降雨量、平均气温和归一化植被指数)对研究区橡胶园土壤全氮含量进行空间预测,并通过754个独立验证点比较了4种模型的预测精度。结果表明RF对土壤全氮的预测值和实测值的相关系数(0.82)明显高于SLR(0.68)、GAMM(0.70)和CART(0.69),而RF的预测平均绝对误差(0.08836 g/kg)和均方根误差(0.13090 g/kg)均低于SLR、GAMM和CART。此外,RF模型预测结果能反映更为详细的局部土壤全氮含量空间变化信息,与实际情况更为接近。可见,RF模型可作为橡胶园土壤全氮含量空间分布预测的高效方法,为其他土壤属性的空间分布预测提供了一种新的方法。     

基于HJ-CCD数据和随机森林算法的小麦叶面积指数反演

为给小麦长势的遥感监测提供技术支持,该文运用随机森林回归(RF,random forest)算法建立小麦叶面积指数(LAI)遥感反演模型。首先基于2010-2013年江苏地区小麦环境减灾卫星HJ-CCD的影像数据,提取拔节、孕穗和开花3个生育期的卫星植被指数,进而根据各生育期植被指数和相应实测LAI数据,利用RF算法构建各期小麦LAI反演模型,并以人工神经网络(ANN,artificial neural network)模型为参比模型进行预测精度的比较。结果表明:RF算法模型在3个生育期的预测结果均好于同期的ANN模型。拔节、孕穗和开花3个生育期RF模型预测值与地面实测值的R2分别为0.79,0.67和0.59,对应的RMSE分别为0.57,0.90和0.78;ANN模型的R2分别为0.67,0.31和0.30,对应的RMSE分别为0.82,1.94和1.43。该研究结果为提高大田尺度下的小麦LAI遥感预测精度提供了技术和方法。     

基于随机森林算法的冬小麦生物量遥感估算模型对比

为了寻求高效的冬小麦生物量估算方法,该研究获取了2014年陕西省杨凌区拔节期、抽穗期和灌浆期的冬小麦生物量和对应的RADARSAT-2全极化雷达、GF1-WFV多光谱数据,并利用随机森林算法（random forest,RF）将光谱、雷达后向散射、光学植被指数和雷达植被指数结合进行冬小麦生物量回归建模。将相关系数分析（correlation coefficient, r）、袋外数据（out-of-bag data,OOB）重要性和灰色关联分析(grey relational analysis, GRA)与随机森林算法（RF）进行整合,构建了3种冬小麦生物量估算模型：r-RF、OOB-RF和GRA-RF,并分别利用3种估算模型对冬小麦生物量进行了估算。结果表明：r-RF、OOB-RF和GRA-RF3种模型分别采用3、4、10组数据时,验证决定系数分别为0.70、0.70和0.65,平均绝对误差分别为0.162、0.164和0.172 kg/m2,均方根误差分别为0.218、0.221和0.236 kg/m2,r-RF和OOB-RF比GRA-RF对冬小麦生物量有更好而的预测能力。研究结果证实了随机森林算法对冬小麦生物量进行遥感估算的潜力。     

近红外光谱技术快速检测婴配乳粉GOS和FOS含量

为实现婴幼儿配方乳粉中低聚半乳糖（GOS）和低聚果糖（FOS）的快速检测，本研究分别采用标准正态变换（SNV）、多元散射校正（MSC）、归一化（Nor）和Savitzky-Golay平滑（SG）4种方法对获取的乳粉原始光谱进行预处理，再使用变量空间迭代收缩算法（VISSA）和竞争自适应重加权采样算法（CARS）提取具有代表性的特征波长，并建立线性偏最小二乘回归（PLSR）模型和非线性支持向量回归（SVR）模型对婴幼儿配方乳粉中的低聚半乳糖（GOS）和低聚果糖（FOS）含量进行预测。结果表明，乳粉的原始光谱经SNV预处理后，再使用VISSA算法能够有效提取GOS和FOS特征波长，建立的VISSA-SVR非线性模型能够得到较优的GOS和FOS预测结果，GOS的VISSA-SVR模型校正集相关系数为0.998 1，均方根误差为0.050 5，预测集相关系数为0.985 0，均方根误差为0.219 3;FOS的VISSA-SVR模型校正集相关系数为0.994 3，均方根误差为0.053 3，预测集相关系数为0.948 7，均方根误差为0.135 7。本研究可为实现婴幼儿配方乳粉生产过程营养成分...     

基于局部加权回归的土壤全氮含量可见-近红外光谱反演

全氮是土壤肥力的重要指标,对作物产量具有决定性作用,采用土壤可见-近红外(Vis-NIR)光谱预测技术及时获取土壤全氮含量信息具有重要意义。采用来自5省的450个土壤样本来验证局部加权回归方法(LWR)结合Vis-NIR光谱技术预测大面积土壤全氮含量的适用性。结果表明,LWR模型的预测效果优于偏最小二乘回归(PLSR)、人工神经网络(ANN)和支持向量机(SVM),选取主成分数为5,相似样本为40时,模型验证的决定系数(RP2)为0.83,均方根误差(RMSEP)为0.25 g kg-1,测定值标准偏差与标准预测误差的比值(RPD)达到2.41。LWR从建模集中选取与验证样本相似的土样作为局部建模样本,降低了差别大的样本对模型的干扰,从而提高了模型的预测能力。因此,LWR建模方法通过大范围、大样本土壤光谱数据进行大尺度区域的全氮等土壤属性预测时能够发挥更好的作用。     

基于多源环境变量的渭–库绿洲土壤颗粒含量预测研究

本文以渭干河–库车河绿洲(简称渭–库绿洲)土壤颗粒为研究对象,采集了绿洲内50个典型表层(0～10 cm)土壤样本,通过相关软件,提取到遥感指数变量、地形和气候等环境变量,经过相关性分析确定环境变量和预测目标间的关系,使用R语言构建了预测土壤颗粒含量的随机森林(random forest,RF)模型和极端梯度提升(extreme gradient boosting,XGBoost)模型。研究结果表明：XGBoost模型的预测结果整体好于RF模型,其中相关系数介于0.39～0.78;土壤pH、高程及衍生变量、光谱变换变量均是两个模型预测土壤颗粒含量的重要因子;将模型预测结果、实测数据和世界土壤数据库(HWSD)中的3种土壤颗粒数据作对比分析,结果表现出模型预测数据的误差小于HWSD与实测数据的误差。综上所述,通过筛选环境变量建立的XGBoost模型,是预测渭–库绿洲土壤颗粒含量的有效方法。     

基于多源辅助变量和随机森林模型的表层土壤全氮分布预测

土壤全氮与土壤质量和肥力密切相关,准确掌握土壤全氮的空间分布特征对精准农业管理措施的实施具有重要意义。以寻乌县为研究区域,利用随机森林(RF)和随机森林残差克里格(RFRK)方法,结合地形因子、地理坐标、遥感因子、气候因子、距离因子和土壤理化因子等多源辅助变量,对寻乌县表层土壤全氮的空间分布进行预测和制图,并在迭代100次模型后对比了两种模型的预测精度。结果表明,在选择的4种模型精度指标中,RF模型的决定系数均值(R~2=0.629 1)和一致性相关系数均值(CCC=0.7613)均高于RFRK模型(R~2=0.5719,CCC=0.6881),而RF模型的平均相对误差均值(MAE=0.157 0 g·kg^-1)和均方根误差均值(RMSE=0.210 8 g·kg^-1)均小于RFRK模型(MAE=0.168 2 g·kg^-1,RMSE=0.226 7 g·kg^-1)。将残差作为误差项加入RF模型并未提高其预测精度,因此,RF模型可以作为土壤属性预测的一种有效方法,为农业管理措施的实施提供技术支撑。     

东北地区稻谷储藏期间脂肪酸含量的预测模型

粮食储备是保障国家粮食安全的重要物质基础,谷物中脂肪酸含量是粮食储藏过程中品质变化的敏感性指标。为了实现绿色储粮安全管理,该文采用多元线性回归(multiple linear regression,MLR)、人工神经网络(artificial neural network, ANN)、支持向量回归(support vector regression, SVR)、最小二乘支持向量回归(least square support vector regression,LSSVR)等机器学习算法模型,对东北地区稻谷储藏过程中的脂肪酸含量(以KOH计)进行预测。通过主成分分析(principal component analysis,PCA)方法筛选稻谷关键储藏参数,得到4个影响稻谷脂肪酸含量的关键因子,分别为稻谷入仓水分、入仓脂肪酸含量、储藏有效积温、检测粮温。然后,将得到的关键因子进行归一化处理,再分别输入到MLR、ANN、SVR、LSSVR模型,采用决定系数R2、平均绝对误差(MAE)、平均绝对百分比误差(MAPE)、均方根误差(RMSE)等评价指标对不同模型的预测性能进行对比,探讨稻谷脂肪酸含量预测的最优模型算法。研究结果表明,LSSVR模型的决定系数R2、MAE、MAPE、RMSE分别为0.911、0.275 mg/100 g、1.604%、0.348 mg/100 g,预测效果略优于MLR,明显优于ANN和SVR,LSSVR和MLR模型可作为稻谷储藏期间脂肪酸含量预测的方法。该研究实现了稻谷脂肪酸含量的预测,为科学储粮、安全绿色储粮提供参考。     

Assessment of gully erosion susceptibility using different DEM-derived topographic factors in the black soil region of Northeast China

As a primary sediment source, gully erosion leads to severe land degradation and poses a threat to food and ecological security. Therefore, identification of susceptible areas is critical to the prevention and control of gully erosion. This study aimed to identify areas prone to gully erosion using four machine learning methods with derived topographic attributes. Eight topographic attributes (elevation, slope aspect, slope degree, catchment area, plan curvature, profile curvature, stream power index, and topographic wetness index) were derived as feature variables controlling gully occurrence from digital elevation models with four different pixel sizes (5.0 m, 12.5 m, 20.0 m, and 30.0 m). A gully inventory map of a small agricultural catchment in Heilongjiang, China, was prepared through a combination of field surveys and satellite imagery. Each topographic attribute dataset was randomly divided into two portions of 70% and 30% for calibrating and validating four machine learning methods, namely random forest (RF), support vector machines (SVM), artificial neural network (ANN), and generalized linear models (GLM). Accuracy (ACC), area under the receiver operating characteristic curve (AUC), root mean square error (RMSE), and mean absolute error (MAE) were calculated to assess the performance of the four machine learning methods in predicting spatial distribution of gully erosion susceptibility (GES). The results suggested that the selected topographic attributes were capable of predicting GES in the study catchment area. A pixel size of 20.0 m was optimal for all four machine learning methods. The RF method described the spatial relationship between the feature variables and gully occurrence with the greatest accuracy, as it returned the highest values of ACC (0.917) and AUC (0.905) at a 20.0 m resolution. The RF was also the least sensitive to resolutions, followed by SVM (ACC = 0.781–0.891, AUC = 0.724–0.861) and ANN (ACC = 0.744–0.808, AUC = 0.649–0.847). GLM performed poorly in this study (ACC = 0.693–0.757, AUC = 0.608–0.703). Based on the spatial distribution of GES determined using the optimal method (RF + pixel size of 20.0 m), 16% of the study area has very high level susceptibility classes, whereas areas with high, moderate, and low levels of susceptibility make up approximately 24%, 30%, and 31% of the study area, respectively. Our results demonstrate that GES assessment with machine learning methods can successfully identify areas prone to gully erosion, providing reference information for future soil conservation plans and land management. In addition, pixel size (resolution) is the key consideration when preparing suitable datasets of feature variables for GES assessment.     

基于机器学习的离心泵气液两相压升预测

针对离心泵气液两相压升难以准确预测的问题,该研究构建了基于机器学习的离心泵压升预测模型.通过试验获得入口体积含气率、转速和液相流量对离心泵两相压升性能的影响规律,建立气液两相运行条件下离心泵性能基础数据库.根据试验结果,确定以入口体积含气率、转速和液相流量作为输入特征,构建基于线性回归、BP神经网络、支持向量机和随机森...     

Pattern recognition of geophysical data

A new rock classification method for ground penetrating radar (GPR) data is presented for cases where no additional geological information is available from boreholes. There are non-linear relationships between petrophysical properties of rocks and electromagnetic waves which can be handled using two methods derived from statistical learning theory on pattern recognition. An investigation was carried out looking at proving the feasibility of the method in principle for use on synthetic models as well as measurement data. The different learning methods were also compared.The method is based on multivariate statistical learning algorithms for the discrimination of layer boundaries between different rocks. The discrimination developed works with artificial neural networks (ANN) and support vector machines (SVM). The processing procedure starts with geological models with varying petrophysical rock parameters, which are to be sought in the measurement data. The models are used to generate synthetic radargrams from which rock properties can be derived using wave attributes. The calculated values of the wave attributes are stored in a multivariate data pool. This data pool is used to train the ANN and the SVM. The same wave attributes are derived from the GPR data and also saved in a data pool. This generates two data sets for pattern recognition with which to directly classify rock layers. Wave attributes can therefore be used to derive the non-linear correlative relationships between rock properties and GPR data by the weighted matrices of ANN and SVM.The presented method can be used to match reflections in the GRR data directly with the layer boundaries of rock formations. The classification of a boundary horizon between rock salt and anhydrite is demonstrated on synthetic GPR traces and measurement data from a rock salt mine. The advantage of this method is that rock classification is not a priori dependent on borehole data.     

Uncertainty in the spatial prediction of soil texture: Comparison of regression tree and Random Forest models

Within the southern Ecuadorian Andes, landslides have an impact on landscape development. Landslide risk estimation as well as hydrological modelling requires physical soil data. Statistical models were adapted to predict the spatial distribution of soil texture from terrain parameters. For this purpose, 56 soil profiles were analysed horizon-wise by pipette and laser method. Results by pipette compared to laser method showed the expected shift to higher silt and lower clay contents. Linear regression equations were adapted. The performance of regression tree (RT) and Random Forest (RF) models was compared by hundredfold model runs on random Jackknife partitions. Digital soil maps of sand, silt and clay percentage mean and standard deviation indicate model variability and prediction uncertainty.RF models performed better than RT models. All terrain factors considered in the analysis influenced soil texture of the surface horizon, but altitude a.s.l. was assigned the highest variable importance during model construction. Shallow subsurface flow is considered responsible for increasing sand/clay ratios with increasing altitude, on steep slopes and with overland flow distance to the channel network by removing clay particles downslope. Deeper soil layers are not influenced by this process and therefore, did not show the same texture properties. However, the influence of parent material and landslides on the spatial distribution of soil texture cannot be neglected. Model performance, most probably, could be improved by a bigger dataset.     

机器学习用于耕地土壤有机碳空间预测对比研究——以亚热带复杂地貌区为例

耕地土壤有机碳(SOC)是土壤质量的重要指标,也是生态系统健康的重要表征。当前机器学习(Machine Learning, ML)用于SOC数字制图日益热门,但不同算法在高空间分辨率SOC数字制图中的对比研究尚有欠缺。本研究以福建省东北部复杂地形地貌区为例,采用10m空间分辨率Sentinel-2影像数据,选取地形、气候、遥感植被变量为驱动因子,重点分析当前常用的机器学习算法——支持向量机(SupportVector Machine,SVM)、随机森林(RandomForest,RF)在SOC预测中的差异,并与传统普通克里格模型(Ordinary Kriging, OK)进行比较。结果表明:基于地形、遥感植被因子和气候因子构建的RF模型表现最佳(RMSE=2.004,r=0.897),其精度优于OK模型(RMSE=4.571, r=0.623),而SVM模型预测精度相对最低(RMSE=5.190, r=0.431);3种模型预测SOC空间分布趋势总体相似,表现为西高东低、北高南低,其中RF模型呈现的空间分异信息更加精细;最优模型反演得到耕地土壤有机碳平均含量为15.33 g·kg-1; RF模型和SVM模型变量重要性程度表明:高程和降水是影响复杂地貌区SOC空间分布的重要变量,而遥感植被因子重要性程度低于高程。     

Stacking集成模型模拟膜下滴灌玉米逐日蒸散量和作物系数

为准确模拟膜下滴灌玉米逐日蒸散量和作物系数,该研究以4个经典机器学习模型:随机森林(Random Forest,RF)、支持向量机(Support Vector Machine,SVM)、BP神经网络(Back Propagation Neural Network,BP)和Adaboost集成学习模型(Adaboost,ADA)为基础,基于Stacking算法建立了集成学习模型(Linear Stacking Model,LSM)对膜下滴灌玉米逐日蒸散量和作物系数进行模拟。并将LSM的模拟精度与RF、SVM、BP和ADA模型的模拟精度相比较,结果表明:1)RF、SVM、BP和ADA模型模拟膜下滴灌玉米的逐日蒸散量和作物系数时的相对均方根误差均大于0.2;2)相比RF、SVM、BP和ADA模型,LSM模型提高了玉米逐日蒸散量和作物系数模拟精度。LSM模拟的膜下滴灌玉米的作物系数相比于FAO推荐值更接近实测值;3)日序数、平均温度、株高、叶面积指数和短波辐射5个特征对玉米膜下滴灌玉米日蒸散量和作物系数影响最高,基于这5个特征建立的LSM模型模拟膜下滴灌玉米的蒸散量和作物系数的R2分别为0.9和0.89,相对均方根误差分别为0.23和0.16。因此,建议在该研究区使用日序数、平均温度、株高、叶面积指数和短波辐射5个特征参数建立LSM模型模拟膜下滴灌玉米蒸散量和作物系数。该研究可为高效节水条件下作物蒸散量和作物系数的精准模拟和合理制定灌溉制度提供参考。     

Using support vector machines to predict cation exchange capacity of different soil horizons in Qingdao City,China

Agricultural, environmental and ecological modeling requires soil cation exchange capacity (CEC) that is difficult to measure. Pedotransfer functions (PTFs) are thus routinely applied to predict CEC from easily measured physicochemical properties (e.g., texture, soil organic matter, pH). This study developed the support vector machines (SVM)‐based PTFs to predict soil CEC based on 208 soil samples collected from A and B horizons in Qingdao City, Shandong Province, China. The database was randomly split into calibration and validation datasets in proportions of 3:1 using the bootstrap method. The optimal SVM parameters were searched by applying the genetic algorithm (GA). The performance of SVM models was compared to those of multiple stepwise regression (MSR) and artificial neural network (ANN) models. Results show that the accuracy of CEC predicted by SVM improves considerably over those predicted by MSR and ANN. The performance of SVM for B horizon (R² = 0.85) is slightly better than that for A horizon (R² = 0.81). The SVM is a powerful approach in the simulation of nonlinear relationship between CEC and physicochemical properties of widely distributed samples from different soil horizons. Sensitivity analysis was also conducted to explore the influence of each input parameter on the CEC predictions by SVM. The clay content is the most sensitive parameter, followed by soil organic matter and pH, while sand content has the weakest influence. This suggests that clay is the most important predictor for predicting CEC of both soil horizons.     

Estimation of Soil Cation Exchange Capacity using Multiple Regression,Artificial Neural Networks,and Adaptive Neuro-fuzzy Inference System Models in Golestan Province,Iran

Artificial neural networks (ANN) and adaptive neuro-fuzzy inference systems (ANFIS) provide an alternative by estimating soil parameters from more readily available data. In this article, multilayer perceptron (MLP) and radial basis function (RBF) of ANN and ANFIS models were described to estimate soil cation exchange capacity and compared to traditional multiple regression (MR). Moreover, to test the accuracy of previous functions that estimate cation exchange capacity (CEC), five pedotransfer functions (PTFs) were surveyed. The results showed that the accuracies of ANN and ANFIS models were similar in relation to their statistical parameters. It was also found that ANFIS model exhibited greater performance than RBF, MLP, MR, and PTFs to estimate soil CEC, respectively. Finally, sensitivity analysis was conducted to determine the most and the least influential variables affecting soil CEC. The performance comparisons of used models showed that the soft computing system is a good tool to predict soil characteristics.     

采用AIS计算中西太平洋延绳钓渔船捕捞努力量

对渔船捕捞行为和捕捞强度空间高分辨率的估计可以作为海洋资源管理和生态脆弱性评估的重要信息。为识别远洋延绳钓渔船作业状态,该文基于2017年10-11月中西太平洋延绳钓渔船卫星船舶自动识别系统(automatic identification system,AIS)数据和捕捞日志数据,采用支持向量机(support vector machine,SVM)学习方法,构建了中国中西太平洋延绳钓渔船捕捞作业状态(捕捞/非捕捞)分类模型。通过计算模型分类准确率、精确率、敏感度和特异度来评价模型对渔船作业状态分类能力。结果表明,模型训练数据的准确率为95.24%(Kappa系数为0.9),验证数据的准确率为93.85%(Kappa系数为0.87)。采用构建好的模型识别2017年10月和11月中西太平洋延绳钓渔船共计125624条AIS记录数据,模型准确率在83.3%(Kappa系数为0.67)。2017年10、11月所有数据分类精确率为82.33%,灵敏度为88.32%,特异度为77.27%。渔船主要作业空间在168°E^173°E,12°S^18°S,有3个明显的作业强度较高区域。基于SVM模型和日志记录的捕捞强度信息在空间上相关性很高(r>0.98),SVM模型识别的渔船捕捞努力量空间分布特征和实际吻合。捕捞努力量与单位捕捞努力量渔获量(catch per unit of effort,CPUE)、渔获尾数、渔获质量和投钩数的相关系数分别是0.68、0.93、0.93和0.94。基于AIS信息挖掘的渔船空间捕捞努力量可用于渔业资源分析。