An evaluation of unsupervised and supervised learning algorithms for clustering landscape types in the United States

Knowledge of landscape type can inform cartographic generalization of hydrographic features, because landscape characteristics provide an important geographic context that affects variation in channel geometry, flow pattern, and network configuration. Landscape types are characterized by expansive spatial gradients, lacking abrupt changes between adjacent classes; and as having a limited number of outliers that might confound classification. The US Geological Survey (USGS) is exploring methods to automate generalization of features in the National Hydrography Data set (NHD), to associate specific sequences of processing operations and parameters with specific landscape characteristics, thus obviating manual selection of a unique processing strategy for every NHD watershed unit. A chronology of methods to delineate physiographic regions for the United States is described, including a recent maximum likelihood classification based on seven input variables. This research compares unsupervised and supervised algorithms applied to these seven input variables, to evaluate and possibly refine the recent classification. Evaluation metrics for unsupervised methods include the Davies–Bouldin index, the Silhouette index, and the Dunn index as well as quantization and topographic error metrics. Cross validation and misclassification rate analysis are used to evaluate supervised classification methods. The paper reports the comparative analysis and its impact on the selection of landscape regions. The compared solutions show problems in areas of high landscape diversity. There is some indication that additional input variables, additional classes, or more sophisticated methods can refine the existing classification.     

Does environmental data increase the accuracy of land use and land cover classification?

Optical image classification converts spectral data into thematic information from the spectral signature of each object in the image. However, spectral separability is influenced by intrinsic characteristics of the targets, as well as the characteristics of the images used. The classification process will present more reliable results when aspects associated with natural environments (climate, soil, relief, water, etc.) and anthropic environments (roads, constructions, urban area) begin to be considered, as they determine and guide land use and land cover (LULC). The objectives of this study are to evaluate the integration of environmental variables with spectral variables and the performance of the Random Forest algorithm in the classification of Landsat-8 OLI images, of a watershed in the Eastern Amazon, Brazil. The classification process used 96 predictive variables, involving spectral, geological, pedological, climatic and topographic data and Euclidean distances. The selection of variables to construct the predictive models was divided into two approaches: (i) data set containing only spectral variables, and (ii) set of environmental variables added to the spectral data. The variables were selected through nonlinear correlation analysis, with the Randomized Dependence Coefficient and the Recursive Feature Elimination (RFE) method, using the Random Forest classifier algorithm. The spectral variables NDVI, bands 2, 4, 5, 6 and 7 of the dry season and band 4 of the rainy season were selected in both approaches (i and ii). The Euclidean distance from the urban area, Arenosol soil class, annual precipitation, precipitation in February and precipitation of the wettest quarter were the variables selected from the auxiliary data set. This study showed that the addition of environmental data to the spectral data reduces the limitation of the latter, regarding the discrimination of the different classes of LULC, in addition to improving the accuracy of the classification. The addition of soil classes to spectral variables provided a reduction in errors for vegetation classification (Evergreen Forest and Cerrado Sensu Stricto), as it was able to inform about nutrient availability and water storage capacity. The study demonstrates that the addition of environmental variables to the spectral variables can be an alternative to improve monitoring in areas of ecotone in Neotropical regions.     

Classification of vegetation in an open landscape using full-waveform airborne laser scanner data

Airborne laser scanning (ALS) is increasingly being used for the mapping of vegetation, although the focus so far has been on woody vegetation, and ALS data have only rarely been used for the classification of grassland vegetation. In this study, we classified the vegetation of an open alkali landscape, characterized by two Natura 2000 habitat types: Pannonic salt steppes and salt marshes and Pannonic loess steppic grasslands. We generated 18 variables from an ALS dataset collected in the growing (leaf-on) season. Elevation is a key factor determining the patterns of vegetation types in the landscape, and hence 3 additional variables were based on a digital terrain model (DTM) generated from an ALS dataset collected in the dormant (leaf-off) season. We classified the vegetation into 24 classes based on these 21 variables, at a pixel size of 1 m. Two groups of variables with and without the DTM-based variables were used in a Random Forest classifier, to estimate the influence of elevation, on the accuracy of the classification. The resulting classes at Level 4, based on associations, were aggregated at three levels — Level 3 (11 classes), Level 2 (8 classes) and Level 1 (5 classes) — based on species pool, site conditions and structure, and the accuracies were assessed. The classes were also aggregated based on Natura 2000 habitat types to assess the accuracy of the classification, and its usefulness for the monitoring of habitat quality. The vegetation could be classified into dry grasslands, wetlands, weeds, woody species and man-made features, at Level 1, with an accuracy of 0.79 (Cohen’s kappa coefficient, κ). The accuracies at Levels 2–4 and the classification based on the Natura 2000 habitat types were κ: 0.76, 0.61, 0.51 and 0.69, respectively. Levels 1 and 2 provide suitable information for nature conservationists and land managers, while Levels 3 and 4 are especially useful for ecologists, geologists and soil scientists as they provide high resolution data on species distribution, vegetation patterns, soil properties and on their correlations. Including the DTM-based variables increased the accuracy (κ) from 0.73 to 0.79 for Level 1. These findings show that the structural and spectral attributes of ALS echoes can be used for the classification of open landscapes, especially those where vegetation is influenced by elevation, such as coastal salt marshes, sand dunes, karst or alluvial areas; in these cases, ALS has a distinct advantage over other remotely sensed data.     

Model Building to Investigate the Role of Spatial Location in Classifying Satellite Image Using SVM,CART and mBACT: A Case Study

This paper investigates the importance of spatial location of pixels in terms of row-column as an additional explanatory variable in classification along with available spectral bands of remotely sensed data. In view of this, a forward step-wise variable selection algorithm is used to select significant bands/variables and build an optimal model to extract the maximum accuracy. Author performed a case study on the area of town of Wolfville acquired by LANDSAT 5 TM data containing six 30 m resolution spectral bands and pixel location as an additional variable. Data are classified into seven classes using three advanced classifiers i.e. classification and regression trees (CART), support vector machines (SVM) and multi-class Bayesian additive classification tree (mBACT). Traditionally, it is assumed that addition of more explanatory variables always increase the accuracy of classified satellite images. However, results of this study show that adding more variables may sometimes confuse the classifier, that is, if selected carefully, fewer variables can provide the more accurate classification. Importance of row-column information turns out to be more beneficial for mBACT followed by SVM. Interestingly, spatial locations did not turn out to be useful for CART. Based on the findings of this study, mBACT appears to be a slightly better classifier than SVM and a substantially better than CART.     

Classifiers vs. input variables—The drivers in image classification for land cover mapping

The study investigates the performance of image classifiers for landscape-scale land cover mapping and the relevance of ancillary data for the classification success in order to assess and to quantify the importance of these components in image classification. Specifically tested are the performance of maximum likelihood classification (MLC), artificial neural networks (ANN) and discriminant analysis (DA) based on Landsat7 ETM+ spectral data in combination with topographic measures and NDVI. ANN produced high accuracies of more than 75% also with limited input information, while MLC and DA produced comparable results only by incorporating ancillary data into the classification process. The superiority of ANN classification was less pronounced on the level of the single land cover classes. The use of ancillary data generally increased classification accuracy and showed a similar potential for increasing classification accuracy than the selection of the classifier. Therefore, a stronger focus on the development of appropriate and optimised sets of input variables is suggested. Also the definition and selection of land cover classes has shown to be crucial and not to be simply adaptable from existing land cover class schemes. A stronger research focus towards discriminating land cover classes by their typical spectral, topographic or seasonal properties is therefore suggested to advance image classification.     

A study of supervised classification accuracy in fuzzy topological methods

The multiple classifier system (MCS) is an effective automatic classification method, useful in connection with remote sensing analysis techniques. Combining MSC with induced fuzzy topology enables a decomposition of image classes. This fuzzy topological MCS then provides a new and improved approach to classification. The basic classification methods discussed in this paper include maximum likelihood classification (MLC), minimum distance classification (MIND) and Mahalanobis distance classification (MAH).     

全球地表覆盖制图研究新范式

全球土地覆盖制图在过去的10年中取得重要进展,空间分辨率从300 m增加至30 m,分类详细程度也有所提高,从10余个一级类到包含29类的二级分类体系。然而,利用光学遥感数据在大空间范围制图方面仍有诸多挑战。本文主要介绍在农田、居住区、水体和湿地制图方面的挑战,讨论在使用多时相和多传感器遥感数据上的困难,这将是未来遥感应用的趋势。由于各种地表覆盖数据产品有自己定义的地表覆盖类型体系和处理流程,通过调和以及集成各种全球土地覆盖制图产品能够满足新的应用目的,并且可以最大程度地利用已有的土地覆盖数据。然而,未来全球土地覆盖制图需要能够按照新应用需求动态生成地表覆盖数据产品的能力。过去的研究表明有效地提高局部尺度制图的分类精度,更好的算法、更多种特征变量(新类型的数据或特征)以及更具代表性的训练样本都非常重要。我们却认为特征变量的使用更重要。本文提出了一个全球土地覆盖制图的新范式。在这个新范式中,地表覆盖类型的定义被分解为定性指标的类、定量指标的植被郁闭度和高度。非植被类型通过它们的光谱和纹理信息提取。复合考虑类、郁闭度和高度3种指标来定义和区别包含植被的地表覆盖类型。郁闭度和高度不能在分类算法中提取,需要借助其他直接测量或间接反演方法。新的范式还表明,一个普遍适用的训练样本集有效地提高了在非洲大陆尺度土地覆盖分类。为了确保更加容易地实现从传统的土地覆盖制图到全球土地覆盖制图新范式的转变,建议构建一体化的数据管理和分析系统。通过集成相关的观测数据、样本数据和分析算法,逐步建成全球土地覆盖制图在线系统,构建全球地表覆盖制图门户网站,为数据生产者、数据用户、专业研究人员、决策人员搭建合作互助的平台。     

Classification of forest land attributes using multi-source remotely sensed data

The aim of the study was to (1) examine the classification of forest land using airborne laser scanning (ALS) data, satellite images and sample plots of the Finnish National Forest Inventory (NFI) as training data and to (2) identify best performing metrics for classifying forest land attributes. Six different schemes of forest land classification were studied: land use/land cover (LU/LC) classification using both national classes and FAO (Food and Agricultural Organization of the United Nations) classes, main type, site type, peat land type and drainage status. Special interest was to test different ALS-based surface metrics in classification of forest land attributes. Field data consisted of 828 NFI plots collected in 2008–2012 in southern Finland and remotely sensed data was from summer 2010. Multinomial logistic regression was used as the classification method. Classification of LU/LC classes were highly accurate (kappa-values 0.90 and 0.91) but also the classification of site type, peat land type and drainage status succeeded moderately well (kappa-values 0.51, 0.69 and 0.52). ALS-based surface metrics were found to be the most important predictor variables in classification of LU/LC class, main type and drainage status. In best classification models of forest site types both spectral metrics from satellite data and point cloud metrics from ALS were used. In turn, in the classification of peat land types ALS point cloud metrics played the most important role. Results indicated that the prediction of site type and forest land category could be incorporated into stand level forest management inventory system in Finland.     

Utility of image point cloud data towards generating enhanced multitemporal multisensor land cover maps

Multitemporal land cover classification over urban areas is challenging, especially when using heterogeneous data sources with variable quality attributes. A prominent challenge is that classes with similar spectral signatures (such as trees and grass) tend to be confused with one another. In this paper, we evaluate the efficacy of image point cloud (IPC) data combined with suitable Bayesian analysis based time-series rectification techniques to improve the classification accuracy in a multitemporal context. The proposed method uses hidden Markov models (HMMs) to rectify land covers that are initially classified by a random forest (RF) algorithm. This land cover classification method is tested using time series of remote sensing data from a heterogeneous and rapidly changing urban landscape (Kuopio city, Finland) observed from 2006 to 2014. The data consisted of aerial images (5 years), Landsat data (all 9 years) and airborne laser scanning data (1 year). The results of the study demonstrate that the addition of three-dimensional image point cloud data derived from aerial stereo images as predictor variables improved overall classification accuracy, around three percentage points. Additionally, HMM-based post processing reduces significantly the number of spurious year-to-year changes. Using a set of 240 validation points, we estimated that this step improved overall classification accuracy by around 3.0 percentage points, and up to 6 to 10 percentage points for some classes. The overall accuracy of the final product was 91% (kappa = 0.88). Our analysis shows that around 1.9% of the area around Kuopio city, representing a total area of approximately 0.61 km², experienced changes in land cover over the nine years considered.     

An object-based approach for mapping forest structural types based on low-density LiDAR and multispectral imagery

Mapping forest structure variables provides important information for the estimation of forest biomass, carbon stocks, pasture suitability or for wildfire risk prevention and control. The optimization of the prediction models of these variables requires an adequate stratification of the forest landscape in order to create specific models for each structural type or strata. This paper aims to propose and validate the use of an object-oriented classification methodology based on low-density LiDAR data (0.5 m^?2) available at national level, WorldView-2 and Sentinel-2 multispectral imagery to categorize Mediterranean forests in generic structural types. After preprocessing the data sets, the area was segmented using a multiresolution algorithm, features describing 3D vertical structure were extracted from LiDAR data and spectral and texture features from satellite images. Objects were classified after feature selection in the following structural classes: grasslands, shrubs, forest (without shrubs), mixed forest (trees and shrubs) and dense young forest. Four classification algorithms (C4.5 decision trees, random forest, k-nearest neighbour and support vector machine) were evaluated using cross-validation techniques. The results show that the integration of low-density LiDAR and multispectral imagery provide a set of complementary features that improve the results (90.75% overall accuracy), and the object-oriented classification techniques are efficient for stratification of Mediterranean forest areas in structural- and fuel-related categories. Further work will be focused on the creation and validation of a different prediction model adapted to the various strata.     

Band Selection for Hyperspectral Image Classification Using Mutual Information

Spectral band selection is a fundamental problem in hyperspectral data processing. In this letter, a new band-selection method based on mutual information (MI) is proposed. MI measures the statistical dependence between two random variables and can therefore be used to evaluate the relative utility of each band to classification. A new strategy is described to estimate the MI using a priori knowledge of the scene, reducing reliance on a "ground truth" reference map, by retaining bands with high associated MI values (subject to the so-called "complementary" conditions). Simulations of classification performance on 16 classes of vegetation from the AVIRIS 92AV3C data set show the effectiveness of the method, which outperforms an MI-based method using the associated reference map, an entropy-based method, and a correlation-based method. It is also competitive with the steepest ascent algorithm at much lower computational cost     

Mapping multiple tree species classes using a hierarchical procedure with optimized node variables and thresholds based on high spatial resolution satellite data

ABSTRACT

Tree species distribution mapping using remotely sensed data has long been an important research area. However, previous studies have rarely established a comprehensive and efficient classification procedure to obtain an accurate result. This study proposes a hierarchical classification procedure with optimized node variables and thresholds to classify tree species based on high spatial resolution satellite imagery. A classification tree structure consisting of parent and leaf nodes was designed based on user experience and visual interpretation. Spectral, textural, and topographic variables were extracted based on pre-segmented images. The random forest algorithm was used to select variables by ranking the impact of all variables. An iterating approach was used to optimize variables and thresholds in each loop by comprehensively considering the test accuracy and selected variables. The threshold range for each selected variable was determined by a statistical method considering the mean and standard deviation for two subnode types at each parent node. Classification of tree species was implemented using the optimized variables and thresholds. The results show that (1) the proposed procedure can accurately map the tree species distribution, with an overall accuracy of over 86% for both training and test stages; (2) critical variables for each class can be identified using this proposed procedure, and optimal variables of most tree plantation nodes are spectra related; (3) the overall forest classification accuracy using the proposed method is more accurate than that using the random forest (RF) and classification and regression tree (CART). The proposed approach provides results with 3.21% and 7.56% higher overall land cover classification accuracy and 4.68% and 10.28% higher overall forest classification accuracy than RF and CART, respectively.     

AN IMPROVED ALGORITHM FOR SUPERVISED FUZZY C-MEANS CLUSTERING OF REMOTELY SENSED DATA

l IntroductionClassification pIays an imPOrtant role for rernotelysensed data tO be intngrated into gapraphical infOr-mation systems(GISs), and is increasingly comPut-eriZed with soPhisticated hardware and software(Cambell l987; Lillesand and Kiefer l994). Pnd-ucts Of classification are usua[ly represented in formof contiguous patches of pixels,with each being la-belled as belonging to a discrete and dominantclass. Such tyPe of classification is termed as crispor discrete. The accuracie…     

Comparing object-based and pixel-based classifications for mapping savannas

The development of robust object-based classification methods suitable for medium to high resolution satellite imagery provides a valid alternative to ‘traditional’ pixel-based methods. This paper compares the results of an object-based classification to a supervised per-pixel classification for mapping land cover in the tropical north of the Northern Territory of Australia. The object-based approach involved segmentation of image data into objects at multiple scale levels. Objects were assigned classes using training objects and the Nearest Neighbour supervised and fuzzy classification algorithm. The supervised pixel-based classification involved the selection of training areas and a classification using the maximum likelihood classifier algorithm. Site-specific accuracy assessment using confusion matrices of both classifications were undertaken based on 256 reference sites. A comparison of the results shows a statistically significant higher overall accuracy of the object-based classification over the pixel-based classification. The incorporation of a digital elevation model (DEM) layer and associated class rules into the object-based classification produced slightly higher accuracies overall and for certain classes; however this was not statistically significant over the object-based using spectral information solely. The results indicate object-based analysis has good potential for extracting land cover information from satellite imagery captured over spatially heterogeneous land covers of tropical Australia.     

An improved algorithm for supervised fuzzyC-means clustering of remotely sensed data

This paper describes an improved algorithm for fuzzyc-means clustering of remotely sensed data, by which the degree of fuzziness of the resultant classification is decreased as comparing with that by a conventional algorithm: that is, the classification accuracy is increased. This is achieved by incorporating covariance matrices at the level of individual classes rather than assuming a global one. Empirical results from a fuzzy classification of an Edinburgh suburban land cover confirmed the improved performance of the new algorithm for fuzzyc-means clustering, in particular when fuzziness is also accommodated in the assumed reference data.     

Object-oriented extraction of beach morphology from video images

The ARGUS system is a shore-based, optical video system offering a suitable remote sensing technique for the purpose of long-term, high-resolution monitoring of coastal morphodynamics. Ten-minute time-exposure (timex) images obtained by the ARGUS cameras during low tide show the intertidal morphology (bars, troughs and rips) by the differences between water, wet sand and dry sand, where dry sand represents bars, and wet sand and water represent troughs and rips. A semi-automatic object-oriented algorithm was developed for classification of intertidal beach in low-tide video images and was tested on 13 low-tide ARGUS images collected at Noordwijk aan Zee, The Netherlands. Because of the strong relation between the visual observations and object-oriented image analysis, the ARGUS images are subdivided in small homogeneous areas (i.e. objects) by segmentation. Maximum likelihood classification creates a model for each day using a random selection of the objects, which are manually labelled, and their accompanying variables. Of the three classes, class wet sand had a classification fit of 43.4% when compared to an in situ classification; class water was correctly classified for 90.1% and dry sand could be classified best (92.8%). By combining their cross-shore position and their classification, objects can be directly linked with the respective morphological features.     

Classifying historical remotely sensed imagery using a tempo-spatial feature evolution (T-SFE) model

Large and growing archives of orbital imagery of the earth’s surface collected over the past 40 years provide an important resource for documenting past and current land cover and environmental changes. However uses of these data are limited by the lack of coincident ground information with which either to establish discrete land cover classes or to assess the accuracy of their identification. Herein is proposed an easy-to-use model, the Tempo-Spatial Feature Evolution (T-SFE) model, designed to improve land cover classification using historical remotely sensed data and ground cover maps obtained at later times. This model intersects (1) a map of spectral classes (S-classes) of an initial time derived from the standard unsupervised ISODATA classifier with (2) a reference map of ground cover types (G-types) of a subsequent time to generate (3) a target map of overlaid patches of S-classes and G-types. This model employs the rules of Count Majority Evaluation, and Subtotal Area Evaluation that are formulated on the basis of spatial feature evolution over time to quantify spatial evolutions between the S-classes and G-types on the target map. This model then applies these quantities to assign G-types to S-classes to classify the historical images. The model is illustrated with the classification of grassland vegetation types for a basin in Inner Mongolia using 1985 Landsat TM data and 2004 vegetation map. The classification accuracy was assessed through two tests: a small set of ground sampling data in 1985, and an extracted vegetation map from the national vegetation cover data (NVCD) over the study area in 1988. Our results show that a 1985 image classification was achieved using this method with an overall accuracy of 80.6%. However, the classification accuracy depends on a proper calibration of several parameters used in the model.     

Classification of remote sensing images from urban areas using a fuzzy possibilistic model

The classification of very high-resolution remotely sensed images from urban areas is addressed. Previous studies have shown the interest of exploiting the local geometrical information of each pixel to improve the classification. This is performed using the derivative morphological profile (DMP) obtained with a granulometric approach, using opening and closing operators. For each pixel, this DMP constitutes the feature vector on which the classification is based. In this letter, we present an interpretation of the DMP in terms of a fuzzy measurement of the characteristic size and contrast of each structure. This fuzzy measure can be compared to predefined possibility distributions to derive a membership degree for a set of given classes. The decision is taken by selecting the class with the highest membership degree. This model is illustrated and validated in a classification problem using IKONOS images.     

顾及空间上下文关系的JointBoost点云分类及特征降维

随着激光雷达技术的发展及广泛应用,点云数据的分类及理解成为了目前一个研究热点。本文研究了较复杂的电力线路走廊场景的点云自动分类方法,目标类别为地面、植被、建筑物、电力塔、电力线等。本文首先归纳、定义了点云分类所需的关键特征,并利用JointBoost实现地物分类;同时,考虑到点云数据量大,其分类速度较慢,本文结合地物空间上的相互关联关系,提出了一种序列化的点云分类及特征降维方法。该方法在保证分类精度的前提下,使分类所需特征维数降低,缩短了分类所需时间。实际的电力线路走廊的激光扫描点云数据分类实验证明本文研究的分类方法的有效性。     

A heuristic multi-criteria classification approach incorporating data quality information for choropleth mapping

ABSTRACT

Despite conceptual and technology advancements in cartography over the decades, choropleth map design and classification fail to address a fundamental issue: estimates that are statistically indifferent may be assigned to different classes on maps or vice versa. Recently, the class separability concept was introduced as a map classification criterion to evaluate the likelihood that estimates in two classes are statistical different. Unfortunately, choropleth maps created according to the separability criterion usually have highly unbalanced classes. To produce reasonably separable but more balanced classes, we propose a heuristic classification approach to consider not just the class separability criterion but also other classification criteria such as evenness and intra-class variability. A geovisual-analytic package was developed to support the heuristic mapping process to evaluate the trade-off between relevant criteria and to select the most preferable classification. Class break values can be adjusted to improve the performance of a classification.