典型遥感数据分类方法的适用性分析——以遥感图像场景分类为例

目前普遍采用的分类器通常都是针对单一或小量任务而设计的,在小数据量的处理中能取得比较满意的结果。但对于海量遥感数据的处理,其在处理时效和分类精度方面还有待研究。本文以遥感图像场景分类任务为例,着重对遥感数据分类问题中几种典型分类方法的适用性进行比较研究,包括K近邻(KNN)、随机森林(RF),支持向量机(SVM)和稀疏表达分类器(SRC)等。分别从参数敏感性,训练样本数据量,待分类样本数据量和样本特征维数对分类器性能的影响等几个方面进行比较分析。实验结果表明:(1)KNN,RF和L0-SRC方法相比RBF-SVM,Linear-SVM和L1-SRC,受参数影响的程度更弱;(2)待分类样本固定的情况下,随着训练样本数目的增加,SRC类型分类方法的分类性能最佳,SVM类型方法次之,然后是RF和KNN,在总体分类时间上呈现出L0-SRCL1-SRCRFRBF-SVM/Linear-SVMKNN/L0-SRC-Batch的趋势;(3)训练样本固定的情况下,所有分类方法的分类精度几乎都不受待分类样本数目变化的影响,RBF-SVM方法性能最佳,其次是L1-SRC,然后是Linear-SVM,最后是RF和L0-SRC/L0-SRC-Batch,在总体分类时间上,L1-SRC和L0-SRC相比其他分类方法最为耗时;(4)样本特征维数的变化不仅影响分类器的运行效率,同时也影响其分类精度,其中SRC和KNN分类器器无需较高的特征维数即可获得较好的分类结果,SVM对高维特征具有较强的包容性和学习能力,RF分类器对特征维数增加则表现得并不敏感,特征维数的增加并不能对其分类精度的提升带来更多的贡献。总的来说,在大数据量的遥感数据分类任务中,现有分类方法具有良好的适用性,但是对于分类器的选择应当基于各自的特点和优势,结合实际应用的特点进行权衡和选择,选择参数敏感性较小,分类总体时间消耗低但分类精度相对较高的分类方法。     

Efficacy of multi-season Sentinel-2 imagery for compositional vegetation classification

Vegetation maps are essential tools for the conservation and management of landscapes as they contain essential information for informing conservation decisions. Traditionally, maps have been created using field-based approaches which, due to limitations in costs and time, restrict the size of the area for which they can be created and frequency at which they can be updated. With the increasing availability of satellite sensors providing multi-spectral imagery with high temporal frequency, new methods for efficient and accurate vegetation mapping have been developed. The objective of this study was to investigate to what extent multi-seasonal Sentinel-2 imagery can assist in mapping complex compositional classifications at fine spatial scales. We deliberately chose a challenging case study, namely a visually and structurally homogenous scrub vegetation (known as kwongan) of Western Australia. The classification scheme consists of 24 target classes and a random 60/40 split was used for model building and validation. We compared several multi-temporal (seasonal) feature sets, consisting of numerous combinations of spectral bands, vegetation indices as well as principal component and tasselled cap transformations, as input to four machine learning classifiers (Support Vector Machines; SVM, Nearest Neighbour; NN, Random Forests; RF, and Classification Trees; CT) to separate target classes. The results show that a multi-temporal feature set combining autumn and spring images sufficiently captured the phenological differences between the classes and produced the best results, with SVM (74%) and NN (72%) classifiers returning statistically superior results compared to RF (65%) and CT (50%). The SWIR spectral bands captured during spring, the greenness indices captured during spring and the tasselled cap transformations derived from the autumn image emerged as most informative, which suggests that ecological factors (e.g. shared species, patch dynamics) occurring at a sub-pixel level likely had the biggest impact on class confusion. However, despite these challenges, the results are auspicious and suggest that seasonal Sentinel-2 imagery has the potential to predict compositional vegetation classes with high accuracy. Further work is needed to determine whether these results are replicable in other vegetation types and regions.     

Sentinel-2与Landsat 8数据组合下的多特征冬小麦面积提取

遥感卫星的波段设置、信噪比及传感器观测角度等因素都会影响作物提取精度。为充分挖掘与发挥Sentinel-2卫星多光谱成像仪(MSI)与Landsat 8陆地成像仪(OLI)在冬小麦信息提取方面的优势,本文以商河县为研究区,基于两数据源的光谱特征、纹理特征、植被指数特征组合数据,利用随机森林(RF)与支持向量机(SVM)对冬小麦进行提取。结果表明:基于单一影像的最优Kappa系数与最优OA分别为0.89和95.13%,基于组合数据源的最优Kappa系数为0.92,最优OA为95.28%,两数据源组合的精度优于单一数据源提取精度;数据组合效果与分类器的性能有关,RF的Kappa系数相对于SVM分别提升0.04、0.20和0.11,OA分别提升2.41%、11.31%和6%,RF对冬小麦提取精度优于SVM。本文研究结果对于构建中高分辨率影像组合的典型农作物分类提取体系具有重要意义。     

Assessment of Sentinel-1A data for rice crop classification using random forests and support vector machines

This study developed an approach to map rice-cropping systems in An Giang and Dong Thap provinces, South Vietnam using multi-temporal Sentinel-1A (S1A) data. The data were processed through four steps: (1) data pre-processing, (2) constructing smooth time series VH backscatter data, (3) rice crop classification using random forests (RF) and support vector machines (SVM) and (4) accuracy assessment. The results indicated that the smooth VH backscatter profiles reflected the temporal characteristics of rice-cropping patterns in the study region. The comparisons between the classification results and the ground reference data indicated that the overall accuracy and Kappa coefficient achieved from RF were 86.1% and 0.72, respectively, which were slightly more accurate than SVM (overall accuracy of 83.4% and Kappa coefficient of 0.67). These results were reaffirmed by the government’s rice area statistics with the relative error in area (REA) values of 0.2 and 2.2% for RF and SVM, respectively.     

利用深层卷积神经网络实现地形辅助的多波束海底底质分类

海底底质分类对于海洋资源开发与利用、海洋科学研究等多方面具有重要意义。目前,多波束探测是实现大范围海底底质分类的有效手段之一,通常基于多波束反向散射强度提取角度响应(AR)特征及反向散射图像特征进行底质分类。由于特征来源较单一,分类器结构简单,往往分类精度不高。为此,本文提出了一种基于深层卷积神经网络(CNN)的多波束海底底质分类方法。除反向散射强度特征外,还利用地形特征,将特征向量转换为波形图,再输入卷积神经网络进行训练和分类。试验对比不同特征组合以及BP网络、支持向量机(SVM)、K近邻(KNN)、随机森林(RF)4种常规分类器,本文模型算法总体分类精度达到94.86%,Kappa系数为0.93,精度具有明显优势,效率也比较高。表明该方法有效利用两种数据类型所蕴含的海底底质信息,充分发挥卷积神经网络权值共享、高效率等特点,实现高分辨率海底底质分类,可对海底底质分类研究提供参考。     

Is it possible to discern Striga weed (Striga hermonthica) infestation levels in maize agro-ecological systems using in-situ spectroscopy?

The invasion by Striga in most cereal crop fields in Africa has posed a significant threat to food security and has caused substantial socioeconomic losses. Hyperspectral remote sensing is an effective means to discriminate plant species, providing possibilities to track such weed invasions and improve precision agriculture. However, essential baseline information using remotely sensed data is missing, specifically for the Striga weed in Africa. In this study, we investigated the spectral uniqueness of Striga compared to other co-occurring maize crops and weeds. We used the in-situ FieldSpec® Handheld 2™ analytical spectral device (ASD), hyperspectral data and their respective narrow-band indices in the visible and near infrared (VNIR) region of the electromagnetic spectrum (EMS) and four machine learning discriminant algorithms (i.e. random forest: RF, linear discriminant analysis: LDA, gradient boosting: GB and support vector machines: SVM) to discriminate among different levels of Striga (Striga hermonthica) infestations in maize fields in western Kenya. We also tested the utility of Sentinel-2 waveband configurations to map and discriminate Striga infestation in heterogenous cereal crop fields. The in-situ hyperspectral reflectance data were resampled to the spectral waveband configurations of Sentinel-2 using published spectral response functions. We sampled and detected seven Striga infestation classes based on three flowering Striga classes (low, moderate and high) against two background endmembers (soil and a mixture of maize and other co-occurring weeds). A guided regularized random forest (GRRF) algorithm was used to select the most relevant hyperspectral wavebands and vegetation indices (VIs) as well as for the resampled Sentinel-2 multispectral wavebands for Striga infestation discrimination. The performance of the four discriminant algorithms was compared using classification accuracy assessment metrics. We were able to positively discriminate Striga from the two background endmembers i.e. soil and co-occurring vegetation (maize and co-occurring weeds) based on the few GRRF selected hyperspectral vegetation indices and the GRRF selected resampled Sentinel-2 multispectral bands. RF outperformed all the other discriminant methods and produced the highest overall accuracy of 91% and 85%, using the hyperspectral and resampled Sentinel-2 multispectral wavebands, respectively, across the four different discriminant models tested in this study. The class with the highest detection accuracy across all the four discriminant algorithms, was the “exclusively maize and other co-occurring weeds” (>70%). The GRRF reduced the dimensionality of the hyperspectral data and selected only 9 most relevant wavebands out of 750 wavebands, 6 VIs out of 15 and 6 out of 10 resampled Sentinel-2 multispectral wavebands for discriminating among the Striga and co-occurring classes. Resampled Sentinel-2 multispectral wavebands 3 (green) and 4 (red) were the most crucial for Striga detection. The use of the most relevant hyperspectral features (i.e. wavebands and VIs) significantly (p ≤ 0.05) increased the overall classification accuracy and Kappa scores (±5% and ±0.2, respectively) in all the machine learning discriminant models. Our results show the potential of hyperspectral, resampled Sentinel-2 multispectral datasets and machine learning discriminant algorithms as a tool to accurately discern Striga in heterogenous maize agro-ecological systems.     

Evaluation of different machine learning methods for land cover mapping of a Mediterranean area using multi-seasonal Landsat images and Digital Terrain Models

Land cover monitoring using digital Earth data requires robust classification methods that allow the accurate mapping of complex land cover categories. This paper discusses the crucial issues related to the application of different up-to-date machine learning classifiers: classification trees (CT), artificial neural networks (ANN), support vector machines (SVM) and random forest (RF). The analysis of the statistical significance of the differences between the performance of these algorithms, as well as sensitivity to data set size reduction and noise were also analysed. Landsat-5 Thematic Mapper data captured in European spring and summer were used with auxiliary variables derived from a digital terrain model to classify 14 different land cover categories in south Spain. Overall, statistically similar accuracies of over 91% were obtained for ANN, SVM and RF. However, the findings of this study show differences in the accuracy of the classifiers, being RF the most accurate classifier with a very simple parameterization. SVM, followed by RF, was the most robust classifier to noise and data reduction. Significant differences in their performances were only reached for thresholds of noise and data reduction greater than 20% (noise, SVM) and 25% (noise, RF), and 80% (reduction, SVM) and 50% (reduction, RF), respectively.     

An effective ensemble classification framework using random forests and a correlation based feature selection technique

Accurate classification of heterogeneous land surfaces with homogeneous land cover classes is a challenging task as satellite images are characterized by a large number of features in the spectral and spatial domains. The identifying relevance of a feature or feature set is an important task for designing an effective classification scheme. Here, an ensemble of random forests (RF) classifiers is realized on the basis of relevance of features. Correlation‐based Feature Selection (CFS) was utilized to assess the relevance of a subset of features by studying the individual predictive ability of each feature along with the degree of redundancy between them. Predictability of RF was greatly improved by random selection of the relevant features in each of the splits. An investigation was carried out on different types of images from the Landsat Enhanced Thematic Mapper Plus (Landsat ETM+) and QuickBird sensors. It has been observed that the performance of the RF classifier was significantly improved while using the optimal set of relevant features compared with a few of the most advanced supervised classifiers such as maximum likelihood classifier (MLC), Navie Bayes, multi‐layer perception (MLP), support vector machine (SVM) and bagging.     

基于Sentinel-1A数据的多种机器学习算法识别冰山的比较

冰山识别对于海洋环境监测和船只安全运行等具有重要的意义,是北极航道开通和北极开发过程中的重要内容。采用合成孔径雷达(SAR)影像进行冰山识别具有独特的优势,多种机器学习算法均可用于SAR影像的冰山识别中。为了最大限度地发挥机器学习算法的性能,有必要对不同机器学习算法及其搭配使用的特征与特征标准化方法进行评估,从而进行最优冰山识别方法的选择。因此,本文基于Sentinel-1A SAR影像,采用多种机器学习方法、多种特征组合及多种特征标准化方法进行冰山识别,并比较各流程方法的识别性能差异。采用的机器学习算法包括贝叶斯分类器(Bayes)、反向神经网络(BPNN)、线性判别分析(LDA)、随机森林(RF)以及支持向量机(SVM);特征标准化方法包括Min-max标准化、Z-score标准化及log函数标准化;数据集是含有12个SAR影像特征的969个冰山与非冰山样本,样本主要位于格陵兰岛东海岸。分类效果采用接收者操作特性(ROC)曲线下的面积(AUC)进行衡量。结果显示,最佳搭配下的RF的AUC值最高,达到了0.945,比最差的Bayes高出0.09。从识别率上来看,RF在冰山查全率为80%的情况下非冰山查全率达到92.6%,效果最好,比第2位的BPNN高出1.4%,比最差的Bayes高出2.6%;BPNN在冰山查全率为90%的情况下非冰山查全率达到87.4%,比第2位的RF高出0.8%,比最差的Bayes高出2.7%。上述结果表明,对冰山识别而言,选择最优的机器学习算法和最佳的特征与特征标准化方法都是十分重要的。     

黄河三角洲地区GF-3雷达数据与Sentinel-2多光谱数据湿地协同分类研究

黄河三角洲湿地的动态变化监测对湿地资源合理利用、开发保护具有重要意义。采用C波段全极化高分三号（GF-3）合成孔径雷达数据与欧洲空间局哨兵二号（Sentinel-2B）多光谱数据,分析了黄河三角洲湿地7类地物的光谱、指数、极化散射以及纹理等特征信息,分别基于最大似然法（maximum likelihood,ML）、决策树（decision tree,DT）、支持向量机（support vector machine,SVM）方法实现了有监督分类,评估了两者协同与单独应用于湿地地物分类与识别的能力,结果表明,两者协同分类时,其总体精度分别可达90.4%、95.4%、95.7%,均明显高于两者单独分类的结果,证明了GF-3雷达数据与多光谱数据在湿地协同分类方面的可靠性和应用潜力。     

Assessing the optimal preprocessing steps of MODIS time series to map cropping systems in Mato Grosso,Brazil

The adoption of new cropping practices such as integrated Crop-Livestock systems (iCL) aims at improving the land use sustainability of the agricultural sector in the Brazilian Amazon. The emergence of such integrated systems, based on crop and pasture rotations over and within years, challenges the remote sensing community who needs to implement accurate and efficient methods to process satellite image time series (SITS) in order to come up with a monitoring protocol. These methods generally include a SITS preprocessing step which can be time consuming. The aim of this study is to assess the importance of preprocessing operations such as temporal smoothing and computation of phenological metrics on the mapping of main cropping systems (i.e. pasture, single cropping, double cropping and iCL), with a special emphasis on the iCL class. The study area is located in the state of Mato Grosso, an important producer of agriculture commodities located in the Southern Brazilian Amazon. SITS were composed of a set of 16-day composites of MODIS Vegetation Indices (MOD13Q1 product) covering a one year period between 2014 and 2015. Two widely used classifiers, i.e. Random Forest (RF) and Support Vector Machine (SVM), were tested using five data sets issued from a same SITS but with different preprocessing levels: (i) raw NDVI; (ii) raw NDVI + raw EVI; (iii) smoothed NDVI; (iv) NDVI-derived phenometrics; (v) raw NDVI + phenometrics. Both RF and SVM classification results showed that the “raw NDVI + raw EVI” data set achieved the highest performance (RF OA = 0.96, RF Kappa = 0.94, SVM OA = 0.95, SVM Kappa = 0.93), followed closely by the “raw NDVI” and the “raw NDVI + phenometrics” datasets. The “NDVI-derived phenometrics” alone achieved the lowest accuracies (RF OA = 0.58 and SVM OA = 0.66). Considering that the implementation of preprocessing steps is computationally expensive and does not provide significant gains in terms of classification accuracy, we recommend to use raw vegetation indices for mapping cropping practices in Mato Grosso, including the integrated Crop-Livestock systems.     

Comparative performance of convolutional neural network,weighted and conventional support vector machine and random forest for classifying tree species using hyperspectral and photogrammetric data

ABSTRACT

The classification of tree species can significantly benefit from high spatial and spectral information acquired by unmanned aerial vehicles (UAVs) associated with advanced classification methods. This study investigated the following topics concerning the classification of 16 tree species in two subtropical forest fragments of Southern Brazil: i) the potential integration of UAV-borne hyperspectral images with 3D information derived from their photogrammetric point cloud (PPC); ii) the performance of two machine learning methods (support vector machine – SVM and random forest – RF) when employing different datasets at a pixel and individual tree crown (ITC) levels; iii) the potential of two methods for dealing with the imbalanced sample set problem: a new weighted SVM (wSVM) approach, which attributes different weights to each sample and class, and a deep learning classifier (convolutional neural network – CNN), associated with a previous step to balance the sample set; and finally, iv) the potential of this last classifier for tree species classification as compared to the above mentioned machine learning methods. Results showed that the inclusion of the PPC features to the hyperspectral data provided a great accuracy increase in tree species classification results when conventional machine learning methods were applied, between 13 and 17% depending on the classifier and the study area characteristics. When using the PPC features and the canopy height model (CHM), associated with the majority vote (MV) rule, the SVM, wSVM and RF classifiers reached accuracies similar to the CNN, which outperformed these classifiers for both areas when considering the pixel-based classifications (overall accuracy of 84.4% in Area 1, and 74.95% in Area 2). The CNN was between 22% and 26% more accurate than the SVM and RF when only the hyperspectral bands were employed. The wSVM provided a slight increase in accuracy not only for some lesser represented classes, but also some major classes in Area 2. While conventional machine learning methods are faster, they demonstrated to be less stable to changes in datasets, depending on prior segmentation and hand-engineered features to reach similar accuracies to those attained by the CNN. To date, CNNs have been barely explored for the classification of tree species, and CNN-based classifications in the literature have not dealt with hyperspectral data specifically focusing on tropical environments. This paper thus presents innovative strategies for classifying tree species in subtropical forest areas at a refined legend level, integrating UAV-borne 2D hyperspectral and 3D photogrammetric data and relying on both deep and conventional machine learning approaches.     

Crop type mapping using LiDAR,Sentinel-2 and aerial imagery with machine learning algorithms

LiDAR data are becoming increasingly available, which has opened up many new applications. One such application is crop type mapping. Accurate crop type maps are critical for monitoring water use, estimating harvests and in precision agriculture. The traditional approach to obtaining maps of cultivated fields is by manually digitizing the fields from satellite or aerial imagery and then assigning crop type labels to each field - often informed by data collected during ground and aerial surveys. However, manual digitizing and labeling is time-consuming, expensive and subject to human error. Automated remote sensing methods is a cost-effective alternative, with machine learning gaining popularity for classifying crop types. This study evaluated the use of LiDAR data, Sentinel-2 imagery, aerial imagery and machine learning for differentiating five crop types in an intensively cultivated area. Different combinations of the three datasets were evaluated along with ten machine learning. The classification results were interpreted by comparing overall accuracies, kappa, standard deviation and f-score. It was found that LiDAR data successfully differentiated between different crop types, with XGBoost providing the highest overall accuracy of 87.8％. Furthermore, the crop type maps produced using the LiDAR data were in general agreement with those obtained by using Sentinel-2 data, with LiDAR obtaining a mean overall accuracy of 84.3％ and Sentinel-2 a mean overall accuracy of 83.6％. However, the combination of all three datasets proved to be the most effective at differentiating between the crop types, with RF providing the highest overall accuracy of 94.4％. These findings provide a foundation for selecting the appropriate combination of remotely sensed data sources and machine learning algorithms for operational crop type mapping.     

TanDEM-X multiparametric data features in sea ice classification over the Baltic sea

In this study, we assess the potential of X-band Interferometric Synthetic Aperture Radar imagery for automated classification of sea ice over the Baltic Sea. A bistatic SAR scene acquired by the TanDEM-X mission over the Bothnian Bay in March of 2012 was used in the analysis. Backscatter intensity, interferometric coherence magnitude, and interferometric phase have been used as informative features in several classification experiments. Various combinations of classification features were evaluated using Maximum likelihood (ML), Random Forests (RF) and Support Vector Machine (SVM) classifiers to achieve the best possible discrimination between open water and several sea ice types (undeformed ice, ridged ice, moderately deformed ice, brash ice, thick level ice, and new ice). Adding interferometric phase and coherence-magnitude to backscatter-intensity resulted in improved overall classification per- formance compared to using only backscatter-intensity. The RF algorithm appeared to be slightly superior to SVM and ML due to higher overall accuracies, however, at the expense of somewhat longer processing time. The best overall accuracy (OA) for three methodologies were achieved using combination of all tested features were 71.56, 72.93, and 72.91％ for ML, RF and SVM classifiers, respectively. Compared to OAs of 62.28, 66.51, and 63.05％ using only backscatter intensity, this indicates strong benefit of SAR interferometry in discriminating different types of sea ice. In contrast to several earlier studies, we were particularly able to successfully discriminate open water and new ice classes.     

High-resolution mapping of forest canopy height using machine learning by coupling ICESat-2 LiDAR with Sentinel-1, Sentinel-2 and Landsat-8 data

Forest canopy height is an important indicator of forest carbon storage, productivity, and biodiversity. The present study showed the first attempt to develop a machine-learning workflow to map the spatial pattern of the forest canopy height in a mountainous region in the northeast China by coupling the recently available canopy height (H_canopy) footprint product from ICESat-2 with the Sentinel-1 and Sentinel-2 satellite data. The ICESat-2 H_canopy was initially validated by the high-resolution canopy height from airborne LiDAR data at different spatial scales. Performance comparisons were conducted between two machine-learning models – deep learning (DL) model and random forest (RF) model, and between the Sentinel and Landsat-8 satellites. Results showed that the ICESat-2 H_canopy showed the highest correlation with the airborne LiDAR canopy height at a spatial scale of 250 m with a Pearson’s correlation coefficient (R) of 0.82 and a mean bias of -1.46 m, providing important evidence on the reliability of the ICESat-2 vegetation height product from the case in China’s forest. Both DL and RF models obtained satisfactory accuracy on the upscaling of ICESat-2 H_canopy assisted by Sentinel satellite co-variables with an R-value between the observed and predicted H_canopy equalling 0.78 and 0.68, respectively. Compared to Sentinel satellites, Landsat-8 showed relatively weaker performance in H_canopy prediction, suggesting that the addition of the backscattering coefficients from Sentinel-1 and the red-edge related variables from Sentinel-2 could positively contribute to the prediction of forest canopy height. To our knowledge, few studies have demonstrated large-scale vegetation height mapping in a resolution ≤ 250 m based on the newly available satellites (ICESat-2, Sentinel-1 and Sentinel-2) and DL regression model, particularly in the forest areas in China. Thus, the present work provided a timely and important supplementary to the applications of these new earth observation tools.     

Evaluation of classifiers for processing Hyperion (EO-1) data of tropical vegetation

There is an urgent necessity to monitor changes in the natural surface features of earth. Compared to broadband multispectral data, hyperspectral data provides a better option with high spectral resolution. Classification of vegetation with the use of hyperspectral remote sensing generates a classical problem of high dimensional inputs. Complexity gets compounded as we move from airborne hyperspectral to Spaceborne technology. It is unclear how different classification algorithms will perform on a complex scene of tropical forests collected by spaceborne hyperspectral sensor. The present study was carried out to evaluate the performance of three different classifiers (Artificial Neural Network, Spectral Angle Mapper, Support Vector Machine) over highly diverse tropical forest vegetation utilizing hyperspectral (EO-1) data. Appropriate band selection was done by Stepwise Discriminant Analysis. The Stepwise Discriminant Analysis resulted in identifying 22 best bands to discriminate the eight identified tropical vegetation classes. Maximum numbers of bands came from SWIR region. ANN classifier gave highest OAA values of 81% with the help of 22 selected bands from SDA. The image classified with the help SVM showed OAA of 71%, whereas the SAM showed the lowest OAA of 66%. All the three classifiers were also tested to check their efficiency in classifying spectra coming from 165 processed bands. SVM showed highest OAA of 80%. Classified subset images coming from ANN (from 22 bands) and SVM (from 165 bands) are quite similar in showing the distribution of eight vegetation classes. Both the images appeared close to the actual distribution of vegetation seen in the study area. OAA levels obtained in this study by ANN and SVM classifiers identify the suitability of these classifiers for tropical vegetation discrimination.     

Sentinel-2A数据支持下的雷州半岛植被类型识别

本文以雷州半岛为研究区,利用Sentinel-2A影像数据和真实植被样本数据,综合探讨了机器学习中随机森林与支持向量机的分类效果,并与传统的最大似然法进行比较。提取Sentinel-2A影像9个波段、7个植被指数、72个纹理特征,通过递归特征消除法挑选了10个特征组合,并将其应用于3种分类方法中,对其分类效果进行比较。结果表明:①有效使用多种特征变量是提高植被类型识别精度的关键,就不同特征对植被类型识别的重要性而言,光谱特征与纹理特征相当且大于植被指数,三者重要性相差不大;②随机森林分类效果最佳,不但能对特征进行有效选择,而且能保证植被类型提取精度,提高运行效率;③基于随机森林特征选择的递归特征消除法得到的特征组合不能对其他分类器性能进行优化,对随机森林模型本身的优化效果也有限。     

Fusion of RADARSAT-2 and multispectral optical remote sensing data for LULC extraction in a tropical agricultural area

In this study, we investigated the performance of different fusion and classification techniques for land cover mapping in Hilir Perak, Peninsula Malaysia using RADAR and Landsat-8 images in a predominantly agricultural area. The fusion methods used are Brovey Transform, Wavelet Transform, Ehlers and Layer Stacking and their results classified into seven different land cover classes which include (1) pixel-based classifiers (spectral angle mapper (SAM), maximum likelihood (ML), support vector machine (SVM)) and (2) Object-based (rule-based and standard nearest neighbour (NN)) classifiers. The result shows that pixel-based classification achieved maximum accuracy of the optical data classification using SVM in Landsat-8 with 74.96% accuracy compared to SAM and ML. For multisource data classification, the highest overall accuracy recorded for layer stacking (SVM) was 79.78%, Ehlers fusion (SVM) with 45.57%, Brovey fusion (SVM) with 63.70% and Wavelet fusion (SVM) 61.16%. And for object-based classifiers, the overall classification accuracy is 95.35% for rule-based and 76.33% for NN classifier, respectively. Based on the analysis of their performances, object-based and the rule-based classifiers produced the best classification accuracy from the fused images.     

Land cover and land use classification performance of machine learning algorithms in a boreal landscape using Sentinel-2 data

ABSTRACT

In recent years, the data science and remote sensing communities have started to align due to user-friendly programming tools, access to high-end consumer computing power, and the availability of free satellite data. In particular, publicly available data from the European Space Agency’s Sentinel missions have been used in various remote sensing applications. However, there is a lack of studies that utilize these data to assess the performance of machine learning algorithms in complex boreal landscapes. In this article, I compare the classification performance of four non-parametric algorithms: support vector machines (SVM), random forests (RF), extreme gradient boosting (Xgboost), and deep learning (DL). The study area chosen is a complex mixed-use landscape in south-central Sweden with eight land-cover and land-use (LCLU) classes. The satellite imagery used for the classification were multi-temporal scenes from Sentinel-2 covering spring, summer, autumn and winter conditions. Using stratified random sampling, each LCLU class was allocated 1477 samples, which were divided into training (70%) and evaluation (30%) subsets. Accuracy was assessed through metrics derived from an error matrix, but primarily overall accuracy was used in allocating algorithm hierarchy. A two-proportion Z-test was used to compare the proportions of correctly classified pixels of the algorithms and a McNemar’s chi-square test was used to compare class-wise predictions. The results show that the highest overall accuracy was produced by support vector machines (0.758 ± 0.017), closely followed by extreme gradient boosting (0.751 ± 0.017), random forests (0.739 ± 0.018), and finally deep learning (0.733 ± 0.0023). The Z-test comparison of classifiers showed that a third of algorithm pairings were statistically different. On a class-wise basis, McNemar’s test results showed that 62% of class-wise predictions were significant from one another at the 5% level or less. Variable importance metrics show that nearly half of the top twenty Sentinel-2 bands belonged to the red edge (25%) and shortwave infrared (23%) portions of the electromagnetic spectrum, and were dominated by scenes from spring (38%) and summer (40%). The results are discussed within the scope of recent studies involving machine learning and Sentinel-2 data and key knowledge gaps identified. The article concludes with recommendations for future research.