Fully fuzzy supervised classification of land cover from remotely sensed imagery with an artificial neural network

Mixed pixels are a major problem in mapping land cover from remotely sensed imagery. Unfortunately, such imagery may be dominated by mixed pixels, and the conventional hard image classification techniques used in mapping applications are unable to appropriately represent the land cover of mixed pixels. Fuzzy classification techniques can, however, accommodate the partial and multiple class membership of mixed pixels, and be used to derive an appropriate land cover representation. This is, however, only a partial solution to the mixed pixel problem in supervised image classification. It must be reognised that the land cover on the ground is fuzzy, at the scale of the pixel, and so it is inappropriate to use procedures designed for hard data in the training and testing stages of the classification. Here an approach for land cover classification in which fuzziness is accommodated in all three stages of a supervised classification is presented. Attention focuses on the classification of airborne thematic mapper data with an artificial neural network. Mixed pixels could be accommodated in training the artificial neural network, since the desired output for each training pixel can be specified. A fuzzy land cover representation was derived by outputting the activation level of the network's output units. The activation level of each output unit was significantly correlated with the proportion of the area represented by a pixel which was covered with the class associated with the unit (r>0.88, significant at the 99% level of confidence). Finally, the distance between the fuzzy land cover classification derived from the artificial neural network and the fuzzy ground data was used to illustrate the accuracy of the land cover representation derived. The dangers of hardening the classification output and ground data sets to enable a conventional assessment of classification accuracy are also illustrated; the hardened data sets were over three times more distant from each other than the fuzzy data sets.     

Approaches for the production and evaluation of fuzzy land cover classifications from remotely-sensed data

Remote sensing is an attractive source of data for land cover mapping applications. Mapping is generally achieved through the application of a conventional statistical classification, which allocates each image pixel to a land cover class. Such approaches are inappropriate for mixed pixels, which contain two or more land cover classes, and a fuzzy classification approach is required. When pixels may have multiple and partial class membership measures of the strength of class membership may be output and, if strongly related to the land cover composition, mapped to represent such fuzzy land cover. This type of representation can be derived by softening the output of a conventional ‘hard’ classification or using a fuzzy classification. The accuracy of the representation provided by a fuzzy classification is, however, difficult to evaluate. Conventional measures of classification accuracy cannot be used as they are appropriate only for ‘hard’ classifications. The accuracy of a classification may, however, be indicated by the way in which the strength of class membership is partitioned between the classes and how closely this represents the partitioning of class membership on the ground. In this paper two measures of the closeness of the land cover representation derived from a classification to that on the ground were used to evaluate a set of fuzzy classifications. The latter were based on measures of the strength of class membership output from classifications by a discriminant analysis, artificial neural network and fuzzy c-means classifiers. The results show the importance of recognising and accommodating for the fuzziness of the land cover on the ground. The accuracy assessment methods used were applicable to pure and mixed pixels and enabled the identification of the most accurate land cover representation derived. The results showed that the fuzzy representations were more accurate than the ‘hard’ classifications. Moreover, the outputs derived from the artificial neural network and the fuzzy c-means algorithm in particular were strongly related to the land cover on the ground and provided the most accurate land cover representations. The ability to appropriately represent fuzzy land cover and evaluate the accuracy of the representation should facilitate the use of remote sensing as a source of land cover data.     

The effect of training set size and composition on artificial neural network classification

Training set characteristics can have a significant effect on the performance of an image classification. In this paper the effect of variations in training set size and composition on the accuracy of classifications of synthetic and remotely sensed data sets by an artificial neural network and discriminant analysis are assessed. Attention is focused on the effects of variations in the overall size of the training set, in terms of the number of training samples, as well as on variations in the size of individual classes in the training set. The results showed that higher classification accuracies were generally derived from the artificial neural network, especially when small training sets only were available. It was also apparent that the opportunity of the artificial neural network to learn class appearance was influenced by the composition of the training set. The results indicated that the size of each class in the training set had an effect similar to. that of including a priori probabilities of class membership into the discriminant analysis. In the classification of the remotely sensed data set the classification accuracy was increased significantly as a result of increasing the number of training cases for abundant classes in the image.     

Optimizing unsupervised classifications of remotely sensed imagery with a data-assisted labeling approach

The quality of remotely sensed land use and land cover (LULC) maps is affected by the accuracy of image data classifications. Various efforts have been made in advancing supervised or unsupervised classification methods to increase the repeatability and accuracy of LULC mapping. This study incorporates a data-assisted labeling approach (DALA) into the unsupervised classification of remotely sensed imagery. The DALA-unsupervised classification algorithm consists of three steps: (1) creation of N spectral-class maps using Iterative Self-Organizing Data Analysis Technique Algorithm (ISODATA); (2) development of LULC maps with assistance of reference data; and (3) accuracy assessments of all the LULC maps using independent reference data and selection of one LULC map with the highest accuracy. Classification experiments with a composite image of a Landsat Thematic Mapper (TM) image and an Enhanced Thematic Mapper Plus (ETM+) image suggest that DALA was effective in making unsupervised classification process more objective, automatic, and accurate. A comparison between the DALA-unsupervised classifications and some conventional classifications suggests that the DALA-unsupervised classification algorithm yielded better classification accuracies compared to these conventional approaches. Such a simple, effective approach has not been systematically examined before but has great potential for many applications in the geosciences.     

FasART模糊神经网络用于遥感图象监督分类的研究

说明了遥感图象数据的非线性性质，目视的图象分类实践是一个模糊推理的过程，模糊神经网络遥感图象分类符合其事物的内在规律，具有理论优势，分析了模糊ART，模糊ARTMAP和FasART模型的结构和原理，详细地阐述了FasART是一种基于模糊逻辑系统的神经网络，提出了一种简化的FasART模型，改变了一般遥感数据的模糊化方法，采用中巴资源一号卫星数据进行测试实验，结果表明，该简化的FasART模型能用于遥感图象的监督分类，其分类精度高于模糊ARTMAP神经网络和K均值算法，且性能稳定，有较好的抗干扰能力，尤其具有良好的处理两组相似程度比较接近的，和同组数据模式变化较大的非线性数据的能力。     

Previously trained neural networks as ensemble members: Knowledge extraction and transfer

The use of Artificial Neural Networks (ANNs) for the classification of remotely sensed imagery offers several advantages over more conventional methods. Yet their training still requires a set of pixels with known land cover. To increase ANN classification accuracy when few training data are available, an algorithm was applied that allows experience gained in previous classifications to be reused. The proposed method was evaluated by classifying a tropical savannah region in northern Togo using Landsat Thematic Mapper (TM) imagery. The presented approach reached a mean kappa coefficient that was significantly larger (at the 95% level) than that obtained after training networks with randomly initialized weights. Also, the observed variances on the obtained accuracies were significantly lower when compared to networks that were randomly initialized. Finally, Bhattacharyya (BH) distances were used to explain why some land cover classes benefit more from knowledge transfer than others.     

Performance characteristics of the indicator classifier on simulated image data

Classification of remotely sensed imagery into groups of pixels having similar spectral reflectance characteristics is conducted classically by comparing the spectral signature of unknown pixels with those of training pixels of known ground cover type. Thus classification methods use only the spectral characteristics of image data without considering the spatial aspects or the relative location of an unknown pixel with respect to pixels from the training data set. An indicator classifier was introduced in 1992 that combines spatial and spectral information in a decision model. In this Letter the performance of this classifier is tested on simulated image data with known mineral targets and varying spatial variability and noise. It is demonstrated that incorporating spatial continuity into the classification process may largely increase the accuracy of the resulting classified images.     

The application of ensemble techniques for land-cover classification in arid lands

Robust classification approaches are required for accurate classification of complex land-use/land-cover categories of desert landscapes using remotely sensed data. Machine-learning ensemble classifiers have proved to be powerful for the classification of remotely sensed data. However, they have not been evaluated for classifying land-cover categories in desert regions. In this study, the performance of two machine-learning ensemble classifiers – random forests (RF) and boosted artificial neural networks – is explored in the context of classification of land use/land cover of desert landscapes. The evaluation is based on the accuracy of classification of remotely sensed data, with and without integration of ancillary data. Landsat-5 Thematic Mapper data captured for a desert landscape in the north-western coastal desert of Egypt are used with ancillary variables derived from a digital terrain model to classify 13 different land-use/land-cover categories. Results show that the two ensemble methods produce accurate land-cover classifications, with and without integrating spectral data with ancillary data. In general, the overall accuracy exceeded 85% and the kappa coefficient (κ) attained values over 0.83. The integration of ancillary data improved the performance of the boosted artificial neural networks by approximately 5% and the random forests by 9%. The latter showed overall higher accuracy; however, boosted artificial neural networks showed better generalization ability and lower overfitting tendencies. The results reveal the merit of applying ensemble methods to integrated spectral and ancillary data of similar desert landscapes for achieving high classification accuracies.     

A cartographic and database approach for land cover/use mapping and generalization from remotely sensed data

Classification of remotely sensed data involves a set of generalization processes, i.e. reality is simplified to a set of a few classes that are relevant to the application under consideration. This article introduces an approach to image classification that uses a class hierarchy structure for mapping unit definition at different generalization levels. This structure is implemented as an operational relational database and allows querying of more detailed land cover/use information from a higher abstraction level, which is that viewed by the map user. Elementary mapping units are defined on the basis of an unsupervised classification process in order to determine the land cover/use classes registered in the remotely sensed data. Mapping unit composition at different generalization levels is defined on the basis of membership values of sampled pixels to land cover/use classes. Unlike fuzzy classifications, membership values are presented to the user at mapping unit level.     

Increasing the accuracy of neural network classification using refined training data

Image classification is a complex process affected by some uncertainties and decisions made by the researchers. The accuracy achieved by a supervised classification is largely dependent upon the training data provided by the analyst. The use of representative training data sets is of significant importance for the performance of all classification methods. However, this issue is more important for neural network classifiers since they take each sample into consideration in the training stage. The representativeness is related to the size and quality of the training data that are highly important in assessing the accuracy of the thematic maps derived from remotely sensed data. Quality analysis of training data helps to identify outlier and mixed pixels that can undermine the reliability and accuracy of a classification resulting from an incorrect class boundary definition. Training data selection can be thought of as an iterative process conducted to form a representative data set after some refinements. Unfortunately, in many applications the quality of the training data is not questioned, and the data set is directly employed in the training stage. In order to increase the representativeness of the training data, a two-stage approach is presented, and performance tests are conducted for a selected region. Multi-layer perceptron model trained with backpropagation learning algorithm is employed to classify major land cover/land use classes present in the study area, the city of Trabzon in Turkey. Results show that the use of representative training data can help the classifier to produce more accurate and reliable results. An improvement of several percent in classification accuracy can make significant effect on the quality of the classified image. Results also confirm the value of visualization tools for the assessment of training pixels through decision boundary analysis.     

基于径向基函数(RBF)映射理论的遥感影像分类模型研究

与传统统计方法的分类器相比较,人工神经网络（ＡＮＮ）方法应用于遥感影像分类,不需预先假设样本空间的参数化统计分布,具有复杂的映射能力。大多数ＡＮＮ分类器采用误差反向传播（ＢＰ）学习算法的多层感知器模型（ＢＰＮＮ）,其主要缺陷是学习速度缓慢、容易陷入局部极小而导致难以收敛等。基于径向基函数（ＲＢＦ）映射理论的神经网络模型融合了参数化统计分布模型和非参经线性感知器映射模型的优点,在实现快速学习的同时,     

Non-linear mixture modelling without end-members using an artificial neural network

Many methods of analysing remotely sensed data assume that pixels are pure, and so a failure to accommodate mixed pixels may result in significant errors in data interpretation and analysis. The analysis of data containing a large proportion of mixed pixels may therefore benefit from the decomposition of the pixels into their component parts. Methods for unmixing the composition of pixels have been used in a range of studies and have often increased the accuracy of the analyses. However, many of the methods assume linear mixing and require end-member spectra, but mixing is often non-linear and end-member spectra are difficult to obtain. In this paper, an alternative approach to unmixing the composition of image pixels, which makes no assumptions about the nature of the mixing and does not require end-member spectra, is presented. The method is based on an artificial neural network (ANN) and shown in a case study to provide accurate estimates of sub-pixel land cover composition. The results of this case study showed that accurate estimates of the proportional cover of a class and its areal extent may be made. It was also shown that there was a tendency for the accuracy of the unmixing to increase with the complexity of the network and the intensity of training. The results indicate the potential to derive accurate information from remotely sensed data sets dominated by mixed pixels.     

A New Methodology for Efficient Classification of Multispectral Satellite Images Using Neural Network Techniques

A methodology based on self-organizing feature maps and indexing techniques for time and memory efficient neural network training and classification of large volumes of remotely sensed data is presented. Results on land-cover classification of multispectral satellite images using two popular neural models show orders of magnitude of speedup with respect to both training and classification times. The generality of the proposed methodology is demonstrated with a dramatic improvement of the classification time of the k-nearest neighbors statistical classifier.     

A hybrid approach to urban land use/cover mapping using Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images

A hybrid method that incorporates the advantages of supervised and unsupervised approaches as well as hard and soft classifications was proposed for mapping the land use/cover of the Atlanta metropolitan area using Landsat 7 Enhanced Thematic Mapper Plus (ETM+) data. The unsupervised ISODATA clustering method was initially used to segment the image into a large number of clusters of pixels. With reference to ground data based on 1?:?40?000 colour infrared aerial photographs in the form of Digital Orthophoto Quarter Quadrangle (DOQQ), homogeneous clusters were labelled. Clusters that could not be labelled because of mixed pixels were clipped out and subjected to a supervised fuzzy classification. A final land use/cover map was obtained by a union overlay of the two partial land use/cover maps. This map was evaluated by comparing with maps produced using unsupervised ISODATA clustering, supervised fuzzy and supervised maximum likelihood classification methods. It was found that the hybrid approach was slightly better than the unsupervised ISODATA clustering in land use/cover classification accuracy, most probably because of the supervised fuzzy classification, which effectively dealt with the mixed pixel problem in the low-density urban use category of land use/cover. It was suggested that this hybrid approach can be economically implemented in a standard image processing software package to produce land use/cover maps with higher accuracy from satellite images of moderate spatial resolution in a complex urban environment, where both discrete and continuous land cover elements occur side by side.     

Application of the MODIS global supervised classification model to vegetation and land cover mapping of Central America

While mapping vegetation and land cover using remotely sensed data has a rich history of application at local scales, it is only recently that the capability has evolved to allow the application of classification models at regional, continental and global scales. The development of a comprehensive training, testing and validation site network for the globe to support supervised and unsupervised classification models is fraught with problems imposed by scale, bioclimatic representativeness of the sites, availability of ancillary map and high spatial resolution remote sensing data, landscape heterogeneity, and vegetation variability. The System for Terrestrial Ecosystem Parameterization (STEP) - a model for characterizing site biophysical, vegetation and landscape parameters to be used for algorithm training and testing and validation - has been developed to support supervised land cover mapping. This system was applied in Central America using two classification systems based on 428 sites. The results indicate that: (1) it is possible to generate site data efficiently at the regional scale; (2) implementation of a supervised model using artificial neural network and decision tree classification algorithms is feasible at the regional level with classification accuracies of 75-88%; and (3) the STEP site parameter model is effective for generating multiple classification systems and thus supporting the development of global surface biophysical parameters.     

Using genetic algorithms in sub-pixel mapping

In remotely sensed images, mixed pixels will always be present. Soft classification defines the membership degree of these pixels for the different land cover classes. Sub-pixel mapping is a technique designed to use the information contained in these mixed pixels to obtain a sharpened image. Pixels are divided into sub-pixels, representing the land cover class fractions. Genetic algorithms combined with the assumption of spatial dependence assign a location to every sub-pixel. The algorithm was tested on synthetic and degraded real imagery. Obtained accuracy measures were higher compared with conventional hard classifications.     

Managing uncertainty when aggregating from pixels to objects: habitats,context-sensitive mapping and possibility theory

Object-oriented remote sensing software provides the user with flexibility in the way that remotely sensed data are classified through segmentation routines and user-specified fuzzy rules. This paper explores the classification and uncertainty issues associated with aggregating detailed ‘sub-objects’ to spatially coarser ‘super-objects’ in object-oriented classifications. We show possibility theory to be an appropriate formalism for managing the uncertainty commonly associated with moving from ‘pixels to parcels’ in remote sensing. A worked example with habitats demonstrates how possibility theory and its associated necessity function provide measures of certainty and uncertainty and support alternative realizations of the same remotely sensed data that are increasingly required to support different applications.     

Super‐resolution mapping of the waterline from remotely sensed data

Methods for mapping the waterline at a subpixel scale from a soft image classification of remotely sensed data are evaluated. Unlike approaches based on hard classification, these methods allow the waterline to run through rather than between image pixels and so have the potential to derive accurate and realistic representations of the waterline from imagery with relatively large pixels. The most accurate predictions of waterline location were made from a geostatistical approach applied to the output of a soft classification (RMSE = 2.25 m) which satisfied the standards for mapping at 1 : 5000 scale from imagery with a 20 m spatial resolution.     

Implementation of machine-learning classification in remote sensing: an applied review

Machine learning offers the potential for effective and efficient classification of remotely sensed imagery. The strengths of machine learning include the capacity to handle data of high dimensionality and to map classes with very complex characteristics. Nevertheless, implementing a machine-learning classification is not straightforward, and the literature provides conflicting advice regarding many key issues. This article therefore provides an overview of machine learning from an applied perspective. We focus on the relatively mature methods of support vector machines, single decision trees (DTs), Random Forests, boosted DTs, artificial neural networks, and k-nearest neighbours (k-NN). Issues considered include the choice of algorithm, training data requirements, user-defined parameter selection and optimization, feature space impacts and reduction, and computational costs. We illustrate these issues through applying machine-learning classification to two publically available remotely sensed data sets.     

A comparison of gradient ascent, gradient descent and genetic-algorithm-based artificial neural networks for the binary classification problem

Abstract: We compare log maximum likelihood gradient ascent, root-mean-square error minimizing gradient descent and genetic-algorithm-based artificial neural network procedures for a binary classification problem. We use simulated data and real-world data sets, and four different performance metrics of correct classification, sensitivity, specificity and reliability for our comparisons. Our experiments indicate that a genetic-algorithm-based artificial neural network that maximizes the total number of correct classifications generally fares well for the binary classification problem. However, if the training data set contains inconsistent decisions or noise then the log maximum likelihood maximizing gradient ascent may be the best classification approach to use. The root-mean-square minimizing gradient descent approach appears to overfit training data and has the lowest reliability among the approaches considered for our research. At the end of the paper, we provide a few guidelines, including computational complexity, for selection of an appropriate technique for a given binary classification problem.