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.     

Automatic land cover analysis for Tenerife by supervised classification using remotely sensed data

Automatic land cover classification from satellite images is an important topic in many remote sensing applications. In this paper, we consider three different statistical approaches to tackle this problem: two of them, namely the well-known maximum likelihood classification (ML) and the support vector machine (SVM), are noncontextual methods. The third one, iterated conditional modes (ICM), exploits spatial context by using a Markov random field. We apply these methods to Landsat 5 Thematic Mapper (TM) data from Tenerife, the largest of the Canary Islands. Due to the size and the strong relief of the island, ground truth data could be collected only sparsely by examination of test areas for previously defined land cover classes.We show that after application of an unsupervised clustering method to identify subclasses, all classification algorithms give satisfactory results (with statistical overall accuracy of about 90%) if the model parameters are selected appropriately. Although being superior to ML theoretically, both SVM and ICM have to be used carefully: ICM is able to improve ML, but when applied for too many iterations, spatially small sample areas are smoothed away, leading to statistically slightly worse classification results. SVM yields better statistical results than ML, but when investigated visually, the classification result is not completely satisfying. This is due to the fact that no a priori information on the frequency of occurrence of a class was used in this context, which helps ML to limit the unlikely classes.     

Boosting a fast neural network for supervised land cover classification

It is demonstrated that the use of an ensemble of neural networks for routine land cover classification of multispectral satellite data can lead to a significant improvement in classification accuracy. Specifically, the AdaBoost.M1 algorithm is applied to a sequence of three-layer, feed-forward neural networks. In order to overcome the drawback of long training time for each network in the ensemble, the networks are trained with an efficient Kalman filter algorithm. On the basis of statistical hypothesis tests, classification performance on multispectral imagery is compared with that of maximum likelihood and support vector machine classifiers. Good generalization accuracies are obtained with computation times of the order of 1 h or less. The algorithms involved are described in detail and a software implementation in the ENVI/IDL image analysis environment is provided.     

A linearised pixel-swapping method for mapping rural linear land cover features from fine spatial resolution remotely sensed imagery

Accurate maps of rural linear land cover features, such as paths and hedgerows, would be useful to ecologists, conservation managers and land planning agencies. Such information might be used in a variety of applications (e.g., ecological, conservation and land management applications). Based on the phenomenon of spatial dependence, sub-pixel mapping techniques can be used to increase the spatial resolution of land cover maps produced from satellite sensor imagery and map such features with increased accuracy. Aerial photography with a spatial resolution of 0.25 m was acquired of the Christchurch area of Dorset, UK. The imagery was hard classified using a simple Mahalanobis distance classifier and the classification degraded to simulate land cover proportion images with spatial resolutions of 2.5 and 5 m. A simple pixel-swapping algorithm was then applied to each of the proportion images. Sub-pixels within pixels were swapped iteratively until the spatial correlation between neighbouring sub-pixels for the entire image was maximised. Visual inspection of the super-resolved output showed that prediction of the position and dimensions of hedgerows was comparable with the original imagery. The maps displayed an accuracy of 87%. To enhance the prediction of linear features within the super-resolved output, an anisotropic modelling component was added. The direction of the largest sums of proportions was calculated within a moving window at the pixel level. The orthogonal sum of proportions was used in estimating the anisotropy ratio. The direction and anisotropy ratio were then used to modify the pixel-swapping algorithm so as to increase the likelihood of creating linear features in the output map. The new linear pixel-swapping method led to an increase in the accuracy of mapping fine linear features of approximately 5% compared with the conventional pixel-swapping method.     

The use of contextual information to improve land cover classification of digital remotely sensed data

All land cover classifications which use remotely sensed data contain error. Where this error is assumed to conform to particular spatial patterns, then it may be possible to apply automated correction procedures. Tests were carried out on urban:non-urban classifications of four sets of Landsat data of the U.K. Confusion between roads and urban areas was reduced by adding the results of linear feature detection to the urban classification. The results were then smoothed and remaining objects below a given size were removed. Results showed that increases in accuracy were obtained which were statistically significant at the 95 per cent confidence level.     

An artificial neural network model for estimating crop yields using remotely sensed information

Crop yield forecasting is a very important task for researchers in remote sensing. Problems exist with traditional statistical modelling (especially regression models) of nonlinear functions with multiple factors in the cropland ecosystem. This paper describes the successful application of an artificial neural network in developing a model for crop yield forecasting using back-propagation algorithms. The model has been adapted and calibrated using on the ground survey and statistical data, and it has proven to be stable and highly accurate.     

Comparison of maximum likelihood classification method with supervised artificial neural network algorithms for land use activities

More than most European cities, Istanbul is experiencing considerable pressure from urban development due to a rapidly increasing population. As a consequence the land use activities in urban and suburban areas are changing dramatically. To provide cost-effective information about the current state and how it is changing in order to develop integrated policies, multi-temporal remotely sensed data, with its synoptic and regular coverage, is being used. Nevertheless, the mapping and monitoring of urban change through remote sensing is difficult owing to the complex urban land use patterns. Although many image processing techniques have been developed for this purpose, they are complicated by differences amongst images caused by differences in the effects of the atmosphere, illumination, and surface moisture. One technique which is relatively unaffected by these problems is based on artificial neural network (ANN) classification algorithms. The main objective of this study was to examine the performance of two ANN classifiers for land use classification using Landsat TM data. Two different supervised ANN approaches were used: the multi layer perceptron (MLP) and the learning vector quantization (LVQ). The performance of these classifiers was compared to the more conventional maximum likelihood approach.     

Technical note: Rotational transformation of remotely sensed data for land use classification

In this paper, the rotational transformation process is explained as a problem of rotation of remotely sensed data in the variance-covariance space. In particular, the rotation which maximizes the covariance-variance ratio is examined in detail for various land use and land cover classes in the Bombay suburban and Thana district area in India. A statistical approach to determine transformed components and other statistical variables on different band combinations is discussed. The results are analysed, and the best possible combinations selected for accurate classification are presented.     

An artificial immune network approach to multi-sensor land use/land cover classification

An optimized artificial immune network-based classification model, namely OPTINC, was developed for remote sensing-based land use/land cover (LULC) classification. Major improvements of OPTINC compared to a typical immune network-based classification model (aiNet) include (1) preservation of the best antibodies of each land cover class from the antibody population suppression, which ensures that each land cover class is represented by at least one antibody; (2) mutation rates being self-adaptive according to the model performance between training generations, which improves the model convergence; and (3) incorporation of both Euclidean distance and spectral angle mapping distance to measure affinity between two feature vectors using a genetic algorithm-based optimization, which helps the model to better discriminate LULC classes with similar characteristics. OPTINC was evaluated using two sites with different remote sensing data: a residential area in Denver, CO with high-spatial resolution QuickBird image and LiDAR data, and a suburban area in Monticello, UT with HyMap hyperspectral imagery. A decision tree, a multilayer feed-forward back-propagation neural network, and aiNet were also tested for comparison. Classification accuracy, local homogeneity of classified images, and model sensitivity to training sample size were examined. OPTINC outperformed the other models with higher accuracy and more spatially cohesive land cover classes with limited salt-and-pepper noise. OPTINC was relatively less sensitive to training sample size than the neural network, followed by the decision tree.     

Pruning artificial neural networks: An example using land cover classification of multi-sensor images

The use of three techniques for pruning artificial neural networks (magnitude-based pruning, optimum brain damage and optimal brain surgeon) is investigated, using microwave SAR and optical SPOT data to classify land cover in a test area located in eastern England. Results show that it is possible to reduce network size significantly without compromising overall classification accuracy; indeed, accuracy may rise as the number of links decreases. However, individual class accuracies and the spatial distribution of the pixels forming the individual classes may change significantly. If the network is pruned too severely some classes may be eliminated altogether. In terms of maintaining overall classification accuracy the optimal brain surgeon algorithm gave the best results, and magnitude-based pruning also gave good results despite its simplicity. The optimum brain damage algorithm performed least well of the three methods tested.     

Thematic labelling from hyperspectral remotely sensed imagery: trade-offs in image properties

The effect of spatial, spectral and noise degradations on the accuracy of two highly contrasting thematic labelling scenarios was investigated. The study used hyperspectral imagery of a site near Falmouth, UK, to assess the effect of the data degradations on the accuracy of supervised classification when the H-resolution scene model was applicable and on labelling when an L-resolution scene model was applicable and no ground data were available. In both scenarios, the spatial, spectral and noise degradations affected the accuracy of labelling. However, over the range of degradations investigated, the noise content of the data was consistently noted to be a major variable affecting the accuracy of labelling.     

Reusing back-propagation artificial neural networks for land cover classification in tropical savannahs

This paper focuses on a method to overcome some of the disadvantages that are related with the use of artificial neural networks (ANNs) as supervised classifiers. The proposed method aims at speeding up network learning, improving classification accuracies and reducing variability on classification performance due to random weight initialization. This can be realized by transferring implicit knowledge from a previously learned source task to a new target task using the proposed algorithm, Discriminality Based Transfer (DBT). The presented approach is compared with conventional network training and a literal transfer method in a 13-class tropical savannah classification experiment using Landsat Thematic Mapper (TM) data. Knowledge was extracted from a network trained on the Kara experimental site in Togo. This information was used to classify the Savanes-L'Oti area which differs in terms of geographical position, image acquisition date, climatological condition and land cover. It was possible to speed up network learning 5.2, 4.3 and 1.8 times using, respectively, 5-, 10- and 20-pixels-per-class training sets. Larger training sets showed less speed improvement. After applying DBT, average classification accuracies were not significantly different from accuracies obtained after training random initialized networks, although DBT tended to show better performance on smaller training sets. It was possible to explain differences in individual class accuracies by analysing Bhattacharyya (BH) distances calculated between all Kara and Savanes-L'Oti classes. Finally, variability on classification performance decreased significantly when training with 5-, 10- and 20-pixels-per-class training sets after DBT application.     

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.     

Applying neural networks in rare vegetation communities classification of remotely sensed images

Artificial neural networks (ANNs) are used for rare vegetation communities’ classification using remotely sensed data. Training of a neural network requires that the user specifies the network structure and sets the learning parameters. Heuristics proposed by a number of researchers to determine the optimum values of network parameters are compared using datasets. Training and test samples were collected for each class type (12 classes). After preliminary statistical tests for training samples, two modification algorithms of the classification scheme were defined: the first one led to creating a scheme which consisted of 7 classes, and the second one led us to creating of 5 class’s scheme. Testing results show that the use of ANNs on the based of 5 class’s scheme can produce higher classification accuracies than either alternative. The visual analysis of the results of the classification is described using Geoinformation Technologies in details. The text was submitted by the authors in English.     

Continuum-based classification of remotely sensed imagery to describe urban sprawl on a watershed scale

Remote sensed imagery can both describe urban sprawl on a watershed scale and provide essential information for modeling the impact of sprawl on watercourses. This paper looks at six watersheds in greater Cleveland, OH: two urban; two rural; and two undergoing rapid urbanization. Thematic Mapper imagery from 1984, 1988, 1994, to 1999 was classified into functional classes describing each watershed in terms of the position of each pixel along continua of [1] percentage permeability and [2] canopy cover. Because the functional classes represent positions along independent continua rather than thematic land-cover classes, they can easily be compared from image to image, and they provide quantitative estimates of parameters at 30-m resolution suitable for spatial simulation models. The imagery classified in this way makes it possible to observe the progress of urban sprawl both within these watersheds and over a study area which extends from the inner city to its rural surroundings.     

A multistage genetic fuzzy classifier for land cover classification from satellite imagery

A linguistic boosted genetic fuzzy classifier (LiBGFC) is proposed in this paper for land cover classification from multispectral images. The LiBGFC is a three-stage process, aiming at effectively tackling the interpretability versus accuracy tradeoff problem. The first stage iteratively generates fuzzy rules, as directed by a boosting algorithm that localizes new rules in uncovered subspaces of the feature space, implicitly preserving the cooperation with previously derived ones. Each rule is able to select the required features, further improving the interpretability of the obtained model. Special provision is taken in the formulation of the fitness function to avoid the creation of redundant rules. A simplification stage follows the first one aiming at further improving the interpretability of the initial rule base, providing a more compact and interpretable solution. Finally, a genetic tuning stage fine tunes the fuzzy sets database improving the classification performance of the obtained model. The LiBGFC is tested using an IKONOS multispectral VHR image, in a lake-wetland ecosystem of international importance. The results indicate the effectiveness of the proposed system in handling multidimensional feature spaces, producing easily understandable fuzzy models.     

Integrating remotely sensed data,GIS and expert knowledge to update object-based land use/land cover information

Remote-sensing technology provides a powerful means for land use/land cover (LU/LC) monitoring at global and regional scales. However, it is more efficient and effective to combine remote-sensing measurements with a geographic information system (GIS) database and expert knowledge for change updating than to use remote-sensing technology alone. In this article, these different sources of information are integrated in the proposed framework, which is able to provide rapid updating of LU/LC information. An object-based data analysis is adopted for thematic mapping, taking both spectral and spatial properties into consideration. An expert knowledge coding is introduced and combined quantitatively with other evidence provided by remotely sensed data and the GIS database. A case study using Landsat Thematic Mapper (TM) datasets demonstrated an overall successful LU/LC map updating and a satisfactory change detection using the proposed change-updating framework.     

Multi-scale crown closure retrieval for moso bamboo forest using multi-source remotely sensed imagery based on geometric-optical and Erf-BP neural network models

This article focuses on retrieving the multi-scale crown closure (CC) of Moso bamboo forest using Système Pour l’Observation de la Terre (SPOT5) and Landsat Thematic Mapper (TM) satellite remotely sensed imagery based on the geometric-optical model and the artificial neural network (ANN) model. CC at local scale was first retrieved using the Li-Strahler geometric-optical model (LSGM) and images from an unmanned aerial vehicle (UAV). Then, multi-scale CC was retrieved using the Erf-BP model (a kind of back-propagation (BP) feed-forward neural network, which takes a Gaussian error function (Erf) as an activation function of the hidden layer) based on a combination of SPOT5 and Landsat TM images. The results show that by combining multi-source remotely sensed data, the CC of Moso bamboo forest can be retrieved at the local region, township area, and county scale with high accuracy using the Erf-BP model. Estimated values have a linear relationship with the observed values at a significance level of 0.05. The highest accuracy of the retrieval of CC (referred to as LSGM-UAV-CC) was observed at the local region based on LSGM and UAV, with the coefficient of determination (R²) of 0.63, followed by that at the township area with an R² of 0.0.55 based on LSGM-UAV-CC and SPOT5 data using the Erf-BP model (Erf-BP-SPOT5-CC), and that at the county scale with an R² of 0.54 based on Erf-BP-SPOT5-CC and Landsat TM data using the Erf-BP model (Erf-BP-TM-CC).     

Classification of remotely sensed imagery using stochastic gradient boosting as a refinement of classification tree analysis

Classification tree analysis (CTA) provides an effective suite of algorithms for classifying remotely sensed data, but it has the limitations of (1) not searching for optimal tree structures and (2) being adversely affected by outliers, inaccurate training data, and unbalanced data sets. Stochastic gradient boosting (SGB) is a refinement of standard CTA that attempts to minimize these limitations by (1) using classification errors to iteratively refine the trees using a random sample of the training data and (2) combining the multiple trees iteratively developed to classify the data. We compared traditional CTA results to SGB for three remote sensing based data sets, an IKONOS image from the Sierra Nevada Mountains of California, a Probe-1 hyperspectral image from the Virginia City mining district of Montana, and a series of Landsat ETM+ images from the Greater Yellowstone Ecosystem (GYE). SGB improved the overall accuracy of the IKONOS classification from 84% to 95% and the Probe-1 classification from 83% to 93%. The worst performing classes using CTA exhibited the largest increases in class accuracy using SGB. A slight decrease in overall classification accuracy resulted from the SGB analysis of the Landsat data.     

A multi-objective neural network based method for cover crop identification from remote sensed data

One of the objectives of conservation agriculture to reduce soil erosion in olive orchards is to protect the soil with cover crops between rows. Andalusian and European administrations have developed regulations to subsidise the establishment of cover crops between rows in olive orchards. Current methods to follow-up the cover crops systems by administrations consist of sampling and on ground visits of around 1% of the total olive orchards surface at any time from March to late June. This paper outlines a multi-objective neural network based method for the classification of olive trees (OT), bare soil (BS) and different cover crops (CC), using remote sensing data taken in spring and summer.The main findings of this paper are: (1) the proposed models performed well in all seasons (particularly during the summer, where only 48 pixels of CC are confused with BS and 10 of BS with CC with the best model obtained. This model obtained a 97.80% of global classification, 95.20% in the class with the worst classification rate and 0.9710 in the KAPPA statistics), and (2) the best-performing models could potentially decrease the number of complaints made to the Andalusian and European administrations. The complaints in question concern the poor performance of current on-ground methods to address the presence or absence of cover crops in olive orchards.