Exploring the potential role of feature selection in global land-cover mapping

Global land cover has been acknowledged as a fundamental variable in several global-scale studies for environment and climate change. Recent developments in global land-cover mapping focused on spatial resolution improvement with more heterogeneous features to integrate the spatial, spectral, and temporal information. Although the high dimensional input features as a whole lead to discriminatory strengths to produce more accurate land-cover maps, it comes at the cost of an increased classification complexity. The feature selection method has become a necessity for dimensionality reduction in classification with large amounts of input features. In this study, the potential of feature selection in global land-cover mapping is explored. A total of 63 features derived from the Landsat Thematic Mapper (TM) spectral bands, Moderate Resolution Imaging Spectroradiometer (MODIS) time series enhanced vegetation index (EVI) data, digital elevation model (DEM), and many climate-ecological variables and global training samples are input to k-nearest neighbours (k-NN) and Random Forest (RF) classifiers. Two filter feature selection algorithms, i.e. Relieff and max-min-associated (MNA), were employed to select the optimal subsets of features for the whole world and different biomes. The mapping accuracies with/without feature selection were evaluated by a global validation sample set. Overall, the result indicates no significant accuracy improvement in global land-cover mapping after dimensionality reduction. Nevertheless, feature selection has the capability of identifying useful features in different biomes and improves the computational efficiency, which is valuable in global-scale computing.     

基于遥感和陆表信息集成的广东省土地覆盖分类方法研究

土地覆盖信息是估算地-气间的生物物理过程和能量交换的关键参数,也是区域和全球尺度气候和生态系统过程模型所需要的重要参量。如何高效地利用遥感数据提取土地覆盖信息是当前研究迫切需要解决的问题。面向对象的分类方法不但充分利用了遥感数据的光谱信息,同时也利用了影像的纹理结构信息和更多的地物分布信息关系,在遥感分类中具有较大的潜力。研究基于2010年多时相的环境卫星数据、TM数据以及DEM数据,并结合地表采集的4000多个样点数据,采用面向对象的分类方法对广东省土地覆盖进行分类。经采样验证,广东省土地覆盖平均精度为85%,分类结果精度远高于常规的分类算法,说明结合陆表信息的面向对象分类方法比常规的分类算法更具有优势,可以实现高精度的土地覆盖分类。     

Multi-scale object-based image analysis and feature selection of multi-sensor earth observation imagery using random forests

The random forest (RF) classifier is a relatively new machine learning algorithm that can handle data sets with large numbers and types of variables. Multi-scale object-based image analysis (MOBIA) can generate dozens, and sometimes hundreds, of variables used to classify earth observation (EO) imagery. In this study, a MOBIA approach is used to classify the land cover in an area undergoing intensive agricultural development. The information derived from the elevation data and imagery from two EO satellites are classified using the RF algorithm. Using a wrapper feature selection algorithm based on the RF, a large initial data set consisting of 418 variables was reduced by ～60%, with relatively little loss in the overall classification accuracy. With this feature-reduced data set, the RF classifier produced a useable depiction of the land cover in the selected study area and achieved an overall classification accuracy of greater than 90%. Variable importance measures produced by the RF algorithm provided an insight into which object features were relatively more important for classifying the individual land-cover types. The MOBIA approach outlined in this study achieved the following: (i) consistently high overall classification accuracies (>85%) using the RF algorithm in all models examined, both before and after feature reduction; (ii) feature selection of a large data set with little expense to the overall classification accuracy; and (iii) increased interpretability of classification models due to the feature selection process and the use of variable importance scores generated by the RF algorithm.     

Development of a global land cover characteristics database and IGBP DISCover from 1 km AVHRR data

Researchers from the U.S. Geological Survey, University of Nebraska-Lincoln and the European Commission's Joint Research Centre, Ispra, Italy produced a 1 km resolution global land cover characteristics database for use in a wide range of continental-to global-scale environmental studies. This database provides a unique view of the broad patterns of the biogeographical and ecoclimatic diversity of the global land surface, and presents a detailed interpretation of the extent of human development. The project was carried out as an International Geosphere-Biosphere Programme, Data and Information Systems (IGBP-DIS) initiative. The IGBP DISCover global land cover product is an integral component of the global land cover database. DISCover includes 17 general land cover classes defined to meet the needs of IGBP core science projects. A formal accuracy assessment of the DISCover data layer will be completed in 1998. The 1 km global land cover database was developed through a continent-by-continent unsupervised classification of 1 km monthly Advanced Very High Resolution Radiometer (AVHRR) Normalized Difference Vegetation Index (NDVI) composites covering 1992-1993. Extensive post-classification stratification was necessary to resolve spectral/temporal confusion between disparate land cover types. The complete global database consists of 961 seasonal land cover regions that capture patterns of land cover, seasonality and relative primary productivity. The seasonal land cover regions were aggregated to produce seven separate land cover data sets used for global environmental modelling and assessment. The data sets include IGBP DISCover, U.S. Geological Survey Anderson System, Simple Biosphere Model, Simple Biosphere Model 2, Biosphere-Atmosphere Transfer Scheme, Olson Ecosystems and Running Global Remote Sensing Land Cover. The database also includes all digital sources that were used in the classification. The complete database can be sourced from the website: http://edcwww.cr.usgs.gov/landdaac/glcc/glcc.html.     

Mapping Land Cover from Remotely Sensed Data with a Softened Feedforward Neural Network Classification

Remote sensing has considerable potential as a source of data for land cover mapping. This potential remains to be fully realised due, in part, to the methods used to extract land cover information from the remotely sensed data. Widely used statistical classifiers provide a poor representation of land cover, make untenable assumptions about the data and convey no information on the quality of individual class allocations. This paper shows that a softened classification, providing information on the strength of membership to all classes for each image pixel, may be derived from a neural network. This information may be used to indicate classification quality on a per-pixel basis. Moreover, a soft or fuzzy classification may be derived to more appropriately represent land cover than the conventional hard classification.     

NDVI-derived land cover classifications at a global scale

Phenological differences among vegetation types, reflected in temporal variations in the Normalized Difference Vegetation Index (NDVI) derived from satellite data, have been used to classify land cover at continental scales. Extending this technique to global scales raises several issues: identifying land cover types that are spectrally distinct and applicable at the global scale; accounting for phasing of seasons in different parts of the world; validating results in the absence of reliable information on global land cover; and acquiring high quality global data sets of satellite sensor data for input to land cover classifications. For this study, a coarse spatial resolution (one by one degree) data set of monthly NDVI values for 1987 was used to explore these methodological issues. A result of a supervised, maximum likelihood classification of eleven cover types is presented to illustrate the feasibility of using satellite sensor data to increase the accuracy of global land cover information, although the result has not been validated systematically. Satellite sensor data at finer spatial resolutions that include other bands in addition to NDVI, as well as methodologies to better identify and describe gradients between cover types, could increase the accuracy of results of global land cover data sets derived from satellite sensor data.     

基于3S技术的土地利用/土地覆盖变化野外采样框架设计 ——以东北地区黑龙江省为例

土地利用/土地覆盖变化是全球环境变化的重要组成部分,随着3S技术的不断成熟和发展,运用RS、GPS和GIS技术进行土地利用/土地覆盖变化研究已成为一种越来越成熟的方式和手段。从空间抽样模型理论出发,以我国黑龙江省为例,运用RS、GPS和GIS技术,通过对黑龙江省道路网、土地利用区划、土地利用/土地覆盖类型、土地利用/土地覆盖1 km×1 km格网数据等空间信息分布的综合考虑、分析,设计了土地利用/土地覆盖变化的综合野外采样框架。框架主要包括采样区的布设、采样路线和采样点的选择等。由于以多层空间信息为采样依据,经实践检验,该采样框架具有经济实用等优点。     

样本特征对参数/非参数分类器分类精度的影响分析

为验证理论训练数量(10~30 p)对参数分类器(如最大似然分类)、非参数分类器(如支撑向量机)的适用性以及样本特征(光谱统计、空间分布特征)对分类器分类精度的影响,选择不同规模的训练样本进行最大似然分类和支撑向量机分类,分析分类精度与样本之间的关系。实验结果表明:随着样本量的增加,最大似然、支撑向量机分类精度均随样本量增多而提高并趋于稳定,最大似然分类精度的增长速度要快于支撑向量机。MLC受样本量的影响较大,在小样本的时候(5个),分类精度不稳定,超过30个样本的时候,分类精度稳定下来;对于SVM分类器,在小样本的时候(5个),分类精度较高且稳定,因此SVM分类适合于小样本分类,不受限于理论样本量的影响。当样本量超过最小理论样本量值(30个)的时候,最大似然分类精度要优于支撑向量机,主要是由于当样本量增加后,最大似然更易于获得有效的信息量样本,而对于支撑向量机边缘信息样本的增加数量不大。研究结果为进一步优化样本进行分类打下前期的实验基础。     

Improving global scale land cover classifications with multi-directional POLDER data and a decision tree classifier

Several investigations indicate that the Bidirectional Reflectance Distribution Function (BRDF) contains information that can be used to complement spectral information for improved land cover classification accuracies. Prior studies on the addition of BRDF information to improve land cover classifications have been conducted primarily at local or regional scales. Thus, the potential benefits of adding BRDF information to improve global to continental scale land cover classification have not yet been explored. Here we examine the impact of multidirectional global scale data from the first Polarization and Directionality of Earth Reflectances (POLDER) spacecraft instrument flown on the Advanced Earth Observing Satellite (ADEOS-1) platform on overall classification accuracy and per-class accuracies for 15 land cover categories specified by the International Geosphere Biosphere Programme (IGBP).

A set of 36,648 global training pixels (7 × 6 km spatial resolution) was used with a decision tree classifier to evaluate the performance of classifying POLDER data with and without the inclusion of BRDF information. BRDF ‘metrics’ for the eight-month POLDER on ADEOS-1 archive (10/1996–06/1997) were developed that describe the temporal evolution of the BRDF as captured by a semi-empirical BRDF model. The concept of BRDF ‘feature space’ is introduced and used to explore and exploit the bidirectional information content. The C5.0 decision tree classifier was applied with a boosting option, with the temporal metrics for spectral albedo as input for a first test, and with spectral albedo and BRDF metrics for a second test. Results were evaluated against 20 random subsets of the training data.

Examination of the BRDF feature space indicates that coarse scale BRDF coefficients from POLDER provide information on land cover that is different from the spectral and temporal information of the imagery. The contribution of BRDF information to reducing classification errors is also demonstrated: the addition of BRDF metrics reduces the mean, overall classification error rates by 3.15% (from 18.1% to 14.95% error) with larger improvements for producer's accuracies of individual classes such as Grasslands (+ 8.71%), Urban areas (+ 8.02%), and Wetlands (+ 7.82%). User's accuracies for the Urban (+ 7.42%) and Evergreen Broadleaf Forest (+ 6.70%) classes are also increased. The methodology and results are widely applicable to current multidirectional satellite data from the Multi-angle Imaging Spectroradiometer (MISR), and to the next generation of POLDER-like multi-directional instruments.     

Random Sampling for Subspace Face Recognition

Subspace face recognition often suffers from two problems: (1) the training sample set is small compared with the high dimensional feature vector; (2) the performance is sensitive to the subspace dimension. Instead of pursuing a single optimal subspace, we develop an ensemble learning framework based on random sampling on all three key components of a classification system: the feature space, training samples, and subspace parameters. Fisherface and Null Space LDA (N-LDA) are two conventional approaches to address the small sample size problem. But in many cases, these LDA classifiers are overfitted to the training set and discard some useful discriminative information. By analyzing different overfitting problems for the two kinds of LDA classifiers, we use random subspace and bagging to improve them respectively. By random sampling on feature vectors and training samples, multiple stabilized Fisherface and N-LDA classifiers are constructed and the two groups of complementary classifiers are integrated using a fusion rule, so nearly all the discriminative information is preserved. In addition, we further apply random sampling on parameter selection in order to overcome the difficulty of selecting optimal parameters in our algorithms. Then, we use the developed random sampling framework for the integration of multiple features. A robust random sampling face recognition system integrating shape, texture, and Gabor responses is finally constructed.     

Computational Technique for an Efficient Classification of Protein Sequences With Distance‐Based Sequence Encoding Algorithm

Machine learning is being implemented in bioinformatics and computational biology to solve challenging problems emerged in the analysis and modeling of biological data such as DNA, RNA, and protein. The major problems in classifying protein sequences into existing families/superfamilies are the following: the selection of a suitable sequence encoding method, the extraction of an optimized subset of features that possesses significant discriminatory information, and the adaptation of an appropriate learning algorithm that classifies protein sequences with higher classification accuracy. The accurate classification of protein sequence would be helpful in determining the structure and function of novel protein sequences. In this article, we have proposed a distance‐based sequence encoding algorithm that captures the sequence's statistical characteristics along with amino acids sequence order information. A statistical metric‐based feature selection algorithm is then adopted to identify the reduced set of features to represent the original feature space. The performance of the proposed technique is validated using some of the best performing classifiers implemented previously for protein sequence classification. An average classification accuracy of 92% was achieved on the yeast protein sequence data set downloaded from the benchmark UniProtKB database.     

SVM在多源遥感图像分类中的应用研究

在利用遥感图像进行土地利用/覆盖分类过程中,可采用以下两种途径来提高分类精度:一是通过增加有利于分类的数据源,引入地理辅助数据和归一化植被指数(NDVI)来进行多源信息融合;二是选择更好的分类方法,例如支持向量机(SVM)学习方法,由于该方法克服了最大似然法和神经网络的弱点,非常适合高维、复杂的小样本多源数据的分类。为了提高多源遥感图像分类的精度,还研究了支持向量机在遥感图像分类中模型的选择,包括多类模型和核函数的选择。分类结果表明,支持向量机比传统的分类方法具有更高的精度,尤其是基于径向基核函数和一对一多类方法的支持向量机模型更适合多源遥感图像分类,因此,基于支持向量机的多源土地利用/覆盖分类能大大提高分类精度。     

Contribution of multispectral and multitemporal information from MODIS images to land cover classification

The goal of this study is to evaluate the relative usefulness of high spectral and temporal resolutions of MODIS imagery data for land cover classification. In particular, we highlight the individual and combinatorial influence of spectral and temporal components of MODIS reflectance data in land cover classification. Our study relies on an annual time series of twelve MODIS 8-days composited images (MOD09A1) monthly acquired during the year 2000, at a 500 m nominal resolution. As our aim is not to propose an operational classifier directed at thematic mapping based on the most efficient combination of reflectance inputs — which will probably change across geographical regions and with different land cover nomenclatures — we intentionally restrict our experimental framework to continental Portugal. Because our observation data stream contains highly correlated components, we need to rank the temporal and the spectral features according not only to their individual ability at separating the land cover classes, but also to their differential contribution to the existing information. To proceed, we resort to the median Mahalanobis distance as a statistical separability criterion. Once achieved this arrangement, we strive to evaluate, in a classification perspective, the gain obtained when the dimensionality of the input feature space grows. We then successively embedded the prior ranked measures into the multitemporal and multispectral training data set of a Support Vector Machines (SVM) classifier. In this way, we show that, only the inclusion of the approximately first three dates substantially increases the classification accuracy. Moreover, this multitemporal factor has a significant effect when coupled with combinations of few spectral bands, but it turns negligible as soon as the full spectral information is exploited. Regarding the multispectral factor, its beneficence on classification accuracy remains more constant, regardless of the number of dates being used.     

Integrating Landsat TM and SRTM-DEM derived variables with decision trees for habitat classification and change detection in complex neotropical environments

Remotely sensed images and processing techniques are a primary tool for mapping changes in tropical forest types important to biodiversity and environmental assessment. Detailed land cover data are lacking for most wet tropical areas that present special challenges for data collection. For this study, we utilize decision tree (DT) classifiers to map 32 land cover types of varying ecological and economic importance over an 8000 km² study area and biological corridor in Costa Rica. We assess multivariate QUEST DTs with unbiased classification rules and linear discriminant node models for integrated vegetation mapping and change detection. Predictor variables essential to accurate land cover classification were selected using importance indices statistically derived with classification trees. A set of 35 variables from SRTM-DEM terrain variables, WorldClim grids, and Landsat TM bands were assessed.

Of the techniques examined, QUEST trees were most accurate by integrating a set of 12 spectral and geospatial predictor variables for image subsets with an overall cross-validation accuracy of 93% ± 3.3%. Accuracy with spectral variables alone was low (69% ± 3.3%). A random selection of training and test set pixels for the entire landscape yielded lower classification accuracy (81%) demonstrating a positive effect of image subsets on accuracy. A post-classification change comparison between 1986 and 2001 reveals that two lowland forest types of differing tree species composition are vulnerable to agricultural conversion. Tree plantations and successional vegetation added forest cover over the 15-year time period, but sometimes replaced native forest types, reducing floristic diversity. Decision tree classifiers, capable of combining data from multiple sources, are highly adaptable for mapping and monitoring land cover changes important to biodiversity and other ecosystem services in complex wet tropical environments.     

ECOCLIMAP-II: An ecosystem classification and land surface parameters database of Western Africa at 1 km resolution for the African Monsoon Multidisciplinary Analysis (AMMA) project

This work is devoted to a presentation of the ECOCLIMAP-II database for Western Africa, which is an upgrade for this region of the former initiative, ECOCLIMAP-I, implemented at global scale. ECOCLIMAP-II is a dual database at 1-km resolution that comprises an ecosystem classification and a coherent set of land surface parameters. This new physiographic information (e.g. leaf area index, fractional vegetation cover, albedo and land cover classification), was especially developed in the framework of the African Monsoon Multidisciplinary Analysis (AMMA) programme in order to support the modelling of land-atmosphere interactions, which stresses the importance of the present study. Criteria for coherence between prevalent land cover classifications and the analysis of time series of the satellite leaf area index (LAI) between 2000 and 2007 constitute the analysis tools for setting up ECOCLIMAP-II. The LAI and inferred fraction of vegetation cover are spatially distributed per land cover unit. The fraction of vegetation cover is handled to split the land surface albedo into vegetation and bare soil albedo components, as is required for a large number of applications. The new ECOCLIMAP-II land cover product is improved with regard to the spatial coherence compared to former version. The reliability of the physiographic details is also confirmed through verification with land cover products at higher resolution.     

A Dempster Shafer-based fuzzy multisensor fusion system using airborne lidar and hyperspectral imagery

ABSTRACT

Nowadays, accurate spectral reflectance information is provided by hyperspectral (HS) data while light detection and ranging (lidar) data provides precise information about the height and geometrical properties of the surfaces. In the most research papers, data fusion of disparate sensors significantly improves object classification performance compared to that of just an individual sensor. Previous researches on fusion of these two sensors had problems such as crisp classifiers or simple fuzzy decision-making systems. This article tries to overcome these weaknesses by accurate support vector machine (SVM) and Fuzzy SVM as classifiers in crisp and fuzzy decision fusion system and fusion of two sensors by two different methods based on precise theories of Bayesian and Shafer. Also, the proposed method tries to compare the results of fusion of both data using decision fusion system with stacked features strategy. This study focuses on HS and lidar fusion through three main phases. The first phase is based on the using of Noise Weighted Harsanyi-Farrand-Chang method and principal component analysis to overcome the high dimensionality problem of HS data. The second phase is based on the feature extraction and selection strategy on lidar data. Finally, fuzzy SVM and Dempster Shafer methods are applied as fuzzy classification and fuzzy decision fusion strategies on the feature spaces. A co-registered HS and lidar data set from Houston of U.S.A. by 15 classes was available to examine the effectiveness of the proposed method. The results of this study highlight that the combination of HS and lidar data enable reliable mapping of land cover.     

Improving 30 m global land-cover map FROM-GLC with time series MODIS and auxiliary data sets: a segmentation-based approach

FROM-GLC (Fine Resolution Observation and Monitoring of Global Land Cover) is the first 30 m resolution global land-cover map produced using Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) data. Due to the lack of temporal features as inputs in producing FROM-GLC, considerable confusion exists among land-cover types (e.g. agriculture lands, grasslands, shrublands, and bareland). The Moderate Resolution Imaging Spectrometer (MODIS) provides high-temporal frequency information on surface cover. Other auxiliary bioclimatic, digital elevation model (DEM), and world maps on soil-water conditions are possible sources for improving the accuracy of FROM-GLC. In this article, a segmentation-based approach was applied to Landsat imagery to down-scale coarser-resolution MODIS data (250 m) and other 1 km resolution auxiliary data to the segment scale based on TM data. Two classifiers (support vector machine (SVM) and random forest (RF)) and two different strategies for use of training samples (global and regional samples based on a spatial temporal selection criterion) were performed. Results show that RF based on the global use of training samples achieves an overall classification accuracy of 67.08% when assessed by test samples collected independently. This is better than the 64.89% achieved by FROM-GLC based on the same set of test samples. Accuracies for vegetation cover types are most substantially improved.     

Land cover classification in rugged areas using simulated moderateresolution remote sensor data and an artificial neural network

Rugged land cover classification accuracies produced by an artificial neural network (ANN) using simulated moderate-resolution remote sensor data exceed overall accuracies produced using the maximum likelihood rule (MLR). Land cover in spatially-complex areas and at broad spatial scales may be difficult to monitor due to ambiguities in spectral reflectance information produced from cloud-related and topographic effects, or from sampling constraints. Such ambiguities may produce inconsistent estimates of changes in vegetation status, surface energy balance, run-off yields, or other land cover characteristics. By use of a 'back-classification' protocol, which uses the same pixels for testing as for training the classifier, tests of ANN versus MLR-based classifiers demonstrated the ANNbased classifier equalled or exceeded classification accuracies produced by the MLR-based classifier in five of six land cover classes evaluated.     

面向宏观地表分类的特征选择算法比较研究

为了成功将土地覆盖进行分类,选择合适的特征是至关重要的。针对利用MODIS数据进行宏观土地覆盖的分类问题,对三种典型的特征选择方法进行了比较研究。研究结果表明：分支定界法（BB）最适合于该土地覆盖分类问题,与此同时,ReliefF和mRMR方法在目标应用中的精度非常接近。研究结果同样表明进行特征选择是非常必要的,它不仅能够大大地降低计算复杂度,而且分类精度能够保持不变,甚至更高。     

MODIS Collection 5 global land cover: Algorithm refinements and characterization of new datasets

Information related to land cover is immensely important to global change science. In the past decade, data sources and methodologies for creating global land cover maps from remote sensing have evolved rapidly. Here we describe the datasets and algorithms used to create the Collection 5 MODIS Global Land Cover Type product, which is substantially changed relative to Collection 4. In addition to using updated input data, the algorithm and ancillary datasets used to produce the product have been refined. Most importantly, the Collection 5 product is generated at 500-m spatial resolution, providing a four-fold increase in spatial resolution relative to the previous version. In addition, many components of the classification algorithm have been changed. The training site database has been revised, land surface temperature is now included as an input feature, and ancillary datasets used in post-processing of ensemble decision tree results have been updated. Further, methods used to correct classifier results for bias imposed by training data properties have been refined, techniques used to fuse ancillary data based on spatially varying prior probabilities have been revised, and a variety of methods have been developed to address limitations of the algorithm for the urban, wetland, and deciduous needleleaf classes. Finally, techniques used to stabilize classification results across years have been developed and implemented to reduce year-to-year variation in land cover labels not associated with land cover change. Results from a cross-validation analysis indicate that the overall accuracy of the product is about 75% correctly classified, but that the range in class-specific accuracies is large. Comparison of Collection 5 maps with Collection 4 results show substantial differences arising from increased spatial resolution and changes in the input data and classification algorithm.