基于机器学习的贫困地区识别算法对比——以陕西省为例

以传统社会经济指标为主导的贫困识别依赖于详尽的普查抽查数据,收集和处理不同质量和数量的普查抽查数据来研究区域贫困需要耗费大量的人力物力和时间,难以快速动态地监测贫困状态。然而时间分辨率高且客观易获取的夜间灯光数据可以在一定程度上弥补统计数据的劣势,即时地反映地表社会经济现象。机器学习算法能够从这些数据中学习出规律和模式,从中挖掘出潜在信息来识别贫困地区。基于陕西省NPP-VIIRS夜间灯光数据,通过构造多维统计变量,利用逻辑回归、支持向量机、K近邻、随机森林、决策树和梯度提升树6种监督分类算法识别贫困地区。结果表明从夜间灯光数据提取的多维特征能够更好的应用于贫困地区的识别,6种算法都能够准确的识别贫困地区,分类结果在空间上具有相似性,且表现出一定的地域性,分类准确度达到76.82%~83.20%。根据混淆矩阵进一步对比各个算法的特点,认为随机森林算法在误差偏移和分类精度等方面综合表现最佳。     

基于遥感和机器学习的内陆水体水深反演技术

文章主要根据机器学习算法（随机森林算法和极端梯度提升算法）和遥感水深反演的原理,利用Sentinel_2多光谱卫星数据和无人船实测水深数据,对内陆水体——梅州水库建立了随机森林（RF）、极端梯度提升（XGBoost）和支持向量机（SVM）水深反演模型,并对反演结果进行对比分析。结果表明：1）RF的训练精度为97%,测试精度为0.80;XGBoost模型的训练精度为97%,测试精度为0.79;SVM的训练精度为90%,测试精度为0.78。说明了在水深预测方面RF模型和XGBoost模型比SVM模型表现更好,对各个区段的水深值较为敏感。2）根据运行时间考察各个模型的效率,其中RF模型从读取数据至输出结果耗时3.92 s;XGBoost模型4.26 s;SVM模型6.66 s。因此,在反演精度和效率上RF模型优于XGBoost模型优于SVM模型,且RF模型的预测结果图细节更加丰富,轮廓更加分明;XGBoost模型次之,但总体效果也较好;SVM模型表现最差。由此可知,机器学习水深反演模型获得的水深结果精度明显提高,解决了传统水深反演模型精度不高的问题。     

基于低空无人机的草原灌丛遥感辨识方法

以位于中国科学院内蒙古草原生态研究定位站灌丛化样地实验平台为研究区，基于低空无人机遥感影像，结合实地调查，开展草原灌丛遥感辨识方法研究。通过对灌丛、草地和裸地归一化植被指数（NDVI）的方差统计分析，确定了裸地与植被的分割阈值为-0.08，并使用该阈值提取植被覆盖区，然后分别利用面向对象的决策树（DT）、贝叶斯（Bayes）、K最邻近（KNN）、支持向量机（SVM）机器学习分类器进行灌丛辨识。研究表明：借助Estimation of Scale Parameter（ESP）最优分割尺度评价工具可以快速确定分割参数，获取灌丛、草地影像对象；利用特征空间优化工具选取了18个的对象特征，可以有效避免盲目选择而导致的计算量增大；通过对不同分类器分类结果的对比和样本数量敏感性实验得出：Bayes分类器精度稳定、无需设置参数，灌丛分类精度最高，总体精度和Kappa系数分别达到92%和0.83，结果与影像地物嵌合最好，能够精确识别单株灌丛；根据Bayes分类器分类结果统计得研究区灌丛盖度为14.74%，平均冠幅为0.6 m²，与样方调查结果基本一致。由于4种分类器的算法特征以及对训练样本数量的敏感性各不相同，因此选择合适的分类器还需根据具体影像的地物特征、空间分辨率和研究区范围来确定。     

Goodnews Bay Platinum Resource Estimation Using Least Squares Support Vector Regression with Selection of Input Space Dimension and Hyperparameters

Resource estimation of a placer deposit is always a difficult and challenging job because of high variability in the deposit. The complexity of resource estimation increases when drill-hole data are sparse. Since sparsely sampled placer deposits produce high-nugget variograms, a traditional geostatistical technique like ordinary kriging sometimes fails to produce satisfactory results. In this article, a machine learning algorithm—the support vector machine (SVM)—is applied to the estimation of a platinum placer deposit. A combination of different neighborhood samples is selected for the input space of the SVM model. The trade-off parameter of the SVM and the bandwidth of the kernel function are selected by genetic algorithm learning, and the algorithm is tested on a testing data set. Results show that if eight neighborhood samples and their distances and angles from the estimated point are considered as the input space for the SVM model, the developed model performs better than other configurations. The proposed input space-configured SVM model is compared with ordinary kriging and the traditional SVM model (location as input) for resource estimation. Comparative results reveal that the proposed input space-configured SVM model outperforms the other two models.     

Landslide susceptibility mapping based on Support Vector Machine: A case study on natural slopes of Hong Kong, China

The Support Vector Machine (SVM) is an increasingly popular learning procedure based on statistical learning theory, and involves a training phase in which the model is trained by a training dataset of associated input and target output values. The trained model is then used to evaluate a separate set of testing data. There are two main ideas underlying the SVM for discriminant-type problems. The first is an optimum linear separating hyperplane that separates the data patterns. The second is the use of kernel functions to convert the original non-linear data patterns into the format that is linearly separable in a high-dimensional feature space. In this paper, an overview of the SVM, both one-class and two-class SVM methods, is first presented followed by its use in landslide susceptibility mapping. A study area was selected from the natural terrain of Hong Kong, and slope angle, slope aspect, elevation, profile curvature of slope, lithology, vegetation cover and topographic wetness index (TWI) were used as environmental parameters which influence the occurrence of landslides. One-class and two-class SVM models were trained and then used to map landslide susceptibility respectively. The resulting susceptibility maps obtained by the methods were compared to that obtained by the logistic regression (LR) method. It is concluded that two-class SVM possesses better prediction efficiency than logistic regression and one-class SVM. However, one-class SVM, which only requires failed cases, has an advantage over the other two methods as only “failed” case information is usually available in landslide susceptibility mapping.     

Predictive modelling of gold potential with the integration of multisource information based on random forest: a case study on the Rodalquilar area,Southern Spain

Mineral exploration activities require robust predictive models that result in accurate mapping of the probability that mineral deposits can be found at a certain location. Random forest (RF) is a powerful machine data-driven predictive method that is unknown in mineral potential mapping. In this paper, performance of RF regression for the likelihood of gold deposits in the Rodalquilar mining district is explored. The RF model was developed using a comprehensive exploration GIS database composed of: gravimetric and magnetic survey, a lithogeochemical survey of 59 elements, lithology and fracture maps, a Landsat 5 Thematic Mapper image and gold occurrence locations. The results of this study indicate that the use of RF for the integration of large multisource data sets used in mineral exploration and for prediction of mineral deposit occurrences offers several advantages over existing methods. Key advantages of RF include: (1) the simplicity of parameter setting; (2) an internal unbiased estimate of the prediction error; (3) the ability to handle complex data of different statistical distributions, responding to nonlinear relationships between variables; (4) the capability to use categorical predictors; and (5) the capability to determine variable importance. Additionally, variables that RF identified as most important coincide with well-known geologic expectations. To validate and assess the effectiveness of the RF method, gold prospectivity maps are also prepared using the logistic regression (LR) method. Statistical measures of map quality indicate that the RF method performs better than LR, with mean square errors equal to 0.12 and 0.19, respectively. The efficiency of RF is also better, achieving an optimum success rate when half of the area predicted by LR is considered.     

Strength of Stacking Technique of Ensemble Learning in Rockburst Prediction with Imbalanced Data: Comparison of Eight Single and Ensemble Models

Rockburst is a common dynamic geological hazard, severely restricting the development and utilization of underground space and resources. As the depth of excavation and mining increases, rockburst tends to occur frequently. Hence, it is necessary to carry out a study on rockburst prediction. Due to the nonlinear relationship between rockburst and its influencing factors, artificial intelligence was introduced. However, the collected data were typically imbalanced. Single algorithms trained by such data have low recognition for minority classes. In order to handle the problem, this paper employed stacking technique of ensemble learning to establish rockburst prediction models. In total, 246 sets of data were collected. In the preprocessing stage, three data mining techniques including principal component analysis, local outlier factor and expectation maximization algorithm were used for dimension reduction, outlier detection and outlier substitution, respectively. Then, the pre-processed data were split into a training set (75%) and a test set (25%) with stratified sampling. Based on the four classical single intelligent algorithms, namely k-nearest neighbors (KNN), support vector machine (SVM), deep neural network (DNN) and recurrent neural network (RNN), four ensemble models (KNN–RNN, SVM–RNN, DNN–RNN and KNN–SVM–DNN–RNN) were built by stacking technique of ensemble learning. The prediction performance of eight models was evaluated, and the differences between single models and ensemble models were analyzed. Additionally, a sensitivity analysis was conducted, revealing the importance of input variables on the models. Finally, the impact of class imbalance on the prediction accuracy and fitting effect of models was quantitatively discussed. The results showed that stacking technique of ensemble learning provides a new and promising way for rockburst prediction, which exhibits unique advantages especially when using imbalanced data.

     

Geochemical Fingerprinting of Coltan Ores by Machine Learning on Uneven Datasets

Two modern machine learning techniques, Linear Programming Boosting (LPBoost) and Support Vector Machines (SVMs), are introduced and applied to a geochemical dataset of niobium–tantalum (“coltan”) ores from Central Africa to demonstrate how such information may be used to distinguish ore provenance, i.e., place of origin. The compositional data used include uni- and multivariate outliers and elemental distributions are not described by parametric frequency distribution functions. The “soft margin” techniques of LPBoost and SVMs can be applied to such data. Optimization of their learning parameters results in an average accuracy of up to c. 92%, if spot measurements are assessed to estimate the provenance of ore samples originating from two geographically defined source areas. A parameterized performance measure, together with common methods for its optimization, was evaluated to account for the presence of uneven datasets. Optimization of the classification function threshold improves the performance, as class importance is shifted towards one of those classes. For this dataset, the average performance of the SVMs is significantly better compared to that of LPBoost.     

THE EFFECT OF MISLABELED SAMPLES ON THE PERFORMANCE OF THE LINEAR LEARNING MACHINE

Over the past 15 years the linear learning machine has been applied to a large number of chemicalproblems.The learning machine approach is conceptually simple and does not require knowledge aboutthe statistical distribution of the data.However,there are problems associated with this approach.Oneproblem which has not been investigated is the influence of mislabeled samples on the positioning of thehyerplane in feature space.If a few samples in a data set are incorrectly tagged prior to training(i.e.thesamples are labeled as members of class 2 even though they are actually members of class 1),it is stilIpossible using the linear learning machine to achieve a classification success rate of 100% for the trainingset.However,unfavorable results will be obtained for the prediction set.The magnitude of this effect andits potential implications regarding the proper use of the linear learning machine are discussed.     

Machine learning predictions of passive microwave brightness temperature over snow-covered land using the special sensor microwave imager (SSM/I)

Abstract

Two different forms of machine learning – an artificial neural network (ANN) and a support vector machine (SVM) – are used to estimate passive microwave (PMW) brightness temperatures (Tb) as observed by the special sensor microwave imager (SSM/I) satellite sensor over snow- covered land in North America. Both techniques reasonably reproduce unbiased estimates of SSM/I observations at 19.35 and 37.0 GHz for both vertically- and horizontally-polarized channels. When compared against SSM/I observations not used during training, domain-averaged statistics from 1 September 1987 to 1 September 2002 yielded a root mean squared error (RMSE) of less than 9 K for all frequency and polarization combinations examined in this study. Even though both ML techniques reasonably reproduced SSM/I Tb observations, the SVM outperformed the ANN because the SVM: (1) better captured the high-frequency (i.e. day-to-day) temporal characteristics in the Tb observations across the majority of the study domain, (2) better reproduced the spatial variability as a function of snow classification, and (3) yielded greater sensitivity to snow-related input variables during the estimation of PMW Tb. These findings reinforce previous research of SVM-based estimation of PMW Tb employing observations from the advanced microwave scanning radiometer.     

多源遥感影像土地覆盖分类结果一致性评价与集成应用

针对多源遥感影像土地覆盖分类结果一致性与分类精度改进的要求,对两组中等空间分辨率的光学影像进行土地覆盖分类,以支持向量机分类结果为基础,采用Kappa统计量、双错误测量、Q统计量、相同错误率从不同角度评价了不同分类结果的一致性。实验表明,多源遥感数据分类结果总体上常规一致性程度较好,二值先验一致性程度尚可,错误一致性程度较小;不同土地覆盖类别的一致性程度并不相同,有的类别甚至出现不一致现象。提出组合法和替换法两种策略以综合数据优点、实现多传感器数据集成应用,能够有效提高分类精度。     

Application of Artificial Neural Network for Gold–Silver Deposits Potential Mapping: A Case Study of Korea

The aim of this study is to analyze hydrothermal gold–silver mineral deposits potential in the Taebaeksan mineralized district, Korea, using an artificial neural network (ANN) and a geographic information system (GIS) environment. A spatial database considering 46 Au and Ag deposits, geophysical, geological, and geochemical data was constructed for the study area using the GIS. The geospatial factors were used with the ANN to analyze mineral potential. The Au and Ag mineral deposits were randomly divided into a training set (70%) to analyze mineral potential using ANN and a test set (30%) to validate predicted potential map. Four different training datasets determined from likelihood ratio and weight of evidence models were applied to analyze and validate the effect of training. Then, the mineral potential index (MPI) was calculated using the trained back-propagation weights, and mineral potential maps (MPMs) were constructed from GIS data for the four training cases. The MPMs were then validated by comparison with the test mineral occurrences. The validation results gave respective accuracies of 73.06, 73.52, 70.11, and 73.10% for the training cases. The comparison results of some training cases showed less sensitive to training data from likelihood ratio than weight of evidence. Overall, the training cases selected from 10% area with low and high index value of MPM_L and MPM_W gave higher accuracy (73.52 and 73.10%) for MPMs than those (73.06 and 70.11%, respectively) from known deposits and 10% area with low index value of MPI_L and MPI_W.     

Self-Learning Random Forests Model for Mapping Groundwater Yield in Data-Scarce Areas

Globally, groundwater plays a major role in supplying drinking water for urban and rural population and is used for irrigation to grow crops and in many industrial processes. A novel self-learning random forest (SLRF) model is developed and validated for groundwater yield zonation within the Yeondong Province in South Korea. This study was conducted with an inventory data initially divided randomly into 70% for training and 30% for testing and 13 groundwater-conditioning factors. SLRF was optimized using Bayesian optimization method. We also compared our method to other machine learning methods including support vector machine (SVM), artificial neural networks (ANN), decision trees (DT), and voting ensemble models. Model validation was accomplished using several methods, including a confusion matrix, receiver operating characteristics, cross-validation, and McNemar’s test. Our proposed self-learning method improves random forest (RF) generalization performance by about 23%, with SLRF success rates of 0.76 and prediction rates of 0.83. In addition, the optimized SLRF performed better [according to a threefold cross-validated AUC (area under curve) of 0.75] than that using randomly initialized parameters (0.57). SLRF outperformed all of the other models for the testing dataset (RF, SVM, ANN, DT, and Voted ANN-RF) when the overall accuracy, prediction rate, and cross-validated AUC metrics were considered. The SLRF also estimated the contribution of individual groundwater conditioning factors and showed that the three most influential factors were geology (1.00), profile curvature (0.97), and TWI (0.95). Overall, SLRF effectively modeled groundwater potential, even within data-scarce regions.

     

Mapping abandoned agricultural land in Kyzyl-Orda,Kazakhstan using satellite remote sensing

In many regions worldwide, cropland abandonment is growing, which has strong and known environmental and socio-economic consequences. Yet, spatially explicit information on the spatial pattern of abandonment is sparse, particularly in post-Soviet countries of Central Asia. When thriving reaching for key Millennium Development Goals such as food security and poverty reduction, the issue of cropland abandonment is critical and therefore must be monitored and limited, or land use transformed into an alternative one. Central Asia experienced large changes of its agricultural system after the collapse of the Soviet Union in 1991. Land degradation, which started already before independence, and cropland abandonment is growing in extent, but their spatial pattern remains ill-understood. The objective of this study was to map and analyse agricultural land use in the irrigated areas of Kyzyl-Orda, southern Kazakhstan, Central Asia. For mapping land use and identifying abandoned agricultural land, an object-based classification approach was applied. Random forest (RF) and support vector machines (SVM) algorithms permitted classifying Landsat and RapidEye data from 2009 to 2014. Overlaying these maps with information about irrigated land parcels, installed during the Soviet period, allowed indicating abandoned fields. Fusing the results of the two approaches, RF and SVM, resulted in classification accuracies of up to 97%. This was statistically significantly higher than with RF or SVM alone. Through the analysis of the land use trajectories, abandoned agricultural fields and a clear indication of abandoned land were identified on almost 50% of all fields in Kyzyl-Orda with an accuracy of approximately 80%. The outputs of this study may provide valuable information for planners, policy- and decision-makers to support better-informed decision-making like reducing possible environmental impacts of land abandonment, or identifying areas for sustainable intensification or re-cultivation.     

基于支持向量机和Getis因子的高分辨率遥感图像分类

采用支持向量机对具有RGB 3个波段、分辨率为0.32 m的航空摄影图像进行实验,首次根据表示空间聚集程度的局部Getis因子完成分类。结果表明:1)当应用基于线性、多项式、径向基和Sigmoid 4种常用核函数的SVM进行分类时,基于径向基的SVM分类精度最高,总体精度超过91%。2)从原始图像计算出局部Getis因子,该指标可用于图像分类,且分类精度与局部Getis因子的步长有关;在步长小于变异函数变程的条件下,应用径向基SVM的总体分类精度达95.66%,高于直接使用原始图像RGB波段光谱信息的分类精度,因此局部Getis因子在高空间分辨率遥感图像分类中具有应用和研究价值。     

基于SVM的泥石流危险度评价研究

选取泥石流一次(可能)最大冲出量(L₁)、泥石流发生频率(L₂)、流域面积(S₁)、主沟长度(S₂)、流域最大相对高差(S₃)、流域切割密度(S₆)和泥沙补给段长度比(S₉)7个因子作为泥石流沟谷危险度评价因子,运用支持向量机理论,以云南省37条泥石流沟的259个基础数据为样本进行学习训练和测试,建立泥石流危险度评价的支持向量机模型,通过实例验证,取得良好效果。     

Fast Nearest Neighbor Classification Methods for Multispectral Imagery

Nearest neighbor classifiers have not been widely used by remote sensing practitioners. The lack of acceptance of these classifiers may be partially due to their notoriously slow speed of execution which makes them impractical for the classification of mega-pixel images. However, training data reduction, distance measure optimization, and neighbor searching algorithms based on the modified k-d tree can speed nearest neighbor classification substantially.     

Application of Discriminant Analysis and Support Vector Machine in Mapping Gold Potential Areas for Further Drilling in the Sari-Gunay Gold Deposit,NW Iran

In this contribution, we used discriminant analysis (DA) and support vector machine (SVM) to model subsurface gold mineralization by using a combination of the surface soil geochemical anomalies and earlier bore data for further drilling at the Sari-Gunay gold deposit, NW Iran. Seventy percent of the data were used as the training data and the remaining 30 % were used as the testing data. Sum of the block grades, obtained by kriging, above the cutoff grade (0.5 g/t) was multiplied by the thickness of the blocks and used as productivity index (PI). Then, the PI variable was classified into three classes of background, medium, and high by using fractal method. Four classification functions of SVM and DA methods were calculated by the training soil geochemical data. Also, by using all the geochemical data and classification functions, the general extension of the gold mineralized zones was predicted. The mineral prediction models at the Sari-Gunay hill were used to locate high and moderate potential areas for further infill systematic and reconnaissance drilling, respectively. These models at Agh-Dagh hill and the area between Sari-Gunay and Agh-Dagh hills were used to define the moderate and high potential areas for further reconnaissance drilling. The results showed that the nu-SVM method with 73.8 % accuracy and c-SVM with 72.3 % accuracy worked better than DA methods.     

基于GF-1与Sentinel-2融合数据的地膜识别方法研究

随着我国地膜使用面积的增加和人们对土壤微塑料污染问题的日益关注,大尺度的地膜遥感识别已成为农业生产管理、土壤污染防治的必要手段。针对地膜光谱反射特征的复杂性以及基于单一遥感影像光谱特征识别方法错分率高等问题,该文以河北省邯郸市邱县为试验区,利用GF-1数据的空间细节与Sentinel-2数据的光谱信息进行NN Diffuse Pan Sharpening融合,据此建立地膜识别的特征矩阵(NDVI、MNDWI、NDBI、IBI、PSI),基于该特征矩阵可实现自动阈值地膜分层分类识别。多种方法的地膜识别结果精度对比表明:多源光学遥感数据融合方法的总体精度为94.87%,Kappa系数达0.89,显著优于基于单一数据源的深度学习法的精度(93.14%)以及基于传统机器学习分类方法的支持向量机(85.91%)和随机森林分类法(86.78%)的精度;通过与Sentinel-2多光谱影像融合,弥补了GF-1数据光谱分辨率低的缺陷,实现了多源数据在地膜识别中的优势互补,可为相关部门农业规划与管理以及生态环境保护等研究提供大尺度、高精度的地膜分布参考数据。