FUZZY HYBRID APPROACHES FOR SOFT CLASSIFICATION OF SATELLITE IMAGES

Abstract

This study is intends to provide an exhaustive treatise on hybrid soft classification algorithm design, analysis and testing using multi-spectral remote sensing imagery. Satellite image classification is a complex process that is affected by various parameters. On the basis of current practices of image classification, the problem area has been examined and better approach of classification using spectral and spatial information have been proposed. The major emphasis has been placed on the use of advanced classification approaches and the techniques used for improving the accuracy of classification. In medium and course resolution images, occurrences of mixed pixels at the scale of measurement are a major problem. Due to this problem, there has been an increasing interest to use contextual information (spectral, spatial or temporal) to eliminate possible ambiguities. Fuzzy set based classifier such as Fuzzy c-Mean (FCM), Possibilistic c-Mean (PCM) and Noise Clustering (NC) classifier can be used to handle mixed pixels. Although, these classifiers have the capability to classify mixed pixels by assigning membership value but unable to incorporate entropy and spatial contextual information of an image. Use of entropy as a regularizer and context, eliminates the problem of isolated pixels and improves the accuracy of a classifier. In this study, three Fuzzy set based classifier (FCM, PCM, NC), two entropy based i.e. Fuzzy c-Mean with Entropy (FCMWE) and Noise Clustering with Entropy (NCWE) and five contextual based i.e. FCM with contextual, PCM with contextual, NC with contextual, FCMWE with contextual and NCWE with contextual classifier has been considered by using MRF models. To incorporate contextual information of an image, Smoothness Adaptive (SA) prior and four Discontinuity Adaptive (DA), MRF models have been used. The hybridized SA and DA prior have been tested on coarse and medium resolution dataset i.e. AWiFS, LISS-III and LISS-IV of Resourcesat-1 with spatial resolution of 60m, 20m and 5m respectively. The classified fraction images of AWiFS have been tested using finer resolution dataset of LISS-III or LISS-IV as reference. To resolve the sub-pixel area allocation problem, class membership, Sub-pixel Confusion Uncertainty Matrix (SCM), FERM, MIN-MIN, MIN-LEAST and ENTROPY methods has been Fuzzy Hybrid Approaches for Soft Classification of Satellite Images ii introduced to assess the accuracy of fuzzy, entropy and contextual based hybrid soft classifiers. All the classification algorithms of this study have been tested in supervised mode using Euclidian weighted norm to classify the remote sensing imagery and entropy has been used to measure the accuracy in terms of uncertainty without using any kind of ground reference data. In FCM classifier, it has been found that irrespective of datasets, the generalized optimum value of weighting exponent (m) has also been fixed as 2.4 for classification using FCM classifier. For PCM classifier, it has been found that irrespective of dataset, m=2.1 have been found to be more suitable to classify agriculture. However, for barren land, moist land and water body, m=2.2 yield good classification result while for sal forest and eucalyptus plantation, m=2.3 gives the best result. It has also been found that MIN-LEAST operator is not suitable to assess the accuracy of PCM. NC classifier has been used to overcome the problem of noisy data points. The performance of NC classifier is dependent upon the optimized value of weighting exponent (m=2.4) and by varying value of resolution parameter (δ). It is found that irrespective of datasets, the optimized value of (δ) for agriculture, sal forest, eucalyptus plantation, barren land, moist land and water body is 3.2 105, 0.8 105, 3.1 105, 4.1 105, 27.8 105, and 3.1 105 respectively. Further, it is found that for δ =105, all accuracy measures have highest value and achieve the threshold criterion of accuracy measures of 85% with least uncertainty values when LISS-III and LISS-IV dataset is used as reference data. Thus, the optimum value of δ has been fixed as 105 for NC classifier. The joint effect of two purely fuzzy models (FCM and NC) and entropy models which are similar to statistical model have been investigated. The NCWE classifier is able to extract the multiple land cover class at a time, at sub-pixel level. The performance of this classifier is dependent on the constant value of resolution parameter δ=105 and regularizing parameter (ν). It is observed that irrespective of datasets, the optimized value of (ν) for agriculture and barren land is 2.12 102, while for sal forest, eucalyptus plantation, moist land and water body, the optimized value is 102. In the process of identifying the generalized optimized value of ν, it is found that for ν =102, where all the accuracy indices have high value. Abstract iii To perform the FCMWE classification, a fixed value of m=1 has been used for different values of ν. It has been found that irrespective of datasets used, ν= 6.6 102 is found to be most suitable in classifying agriculture and eucalyptus plantation. However, for barren land, moist land and water body, ν =102 is found to be suitable for classification using FCMWE classification approach. For sal forest, ν= 7.7 102 is found to be most appropriate for classification. To incorporate the spatial contextual information along with spectral information, hybridized model has been devised to resolve the problem of mixed pixel. Contextual information has been added into FCM, PCM, NC, NCWE and FCMWE classifiers to generate smoothness effect, preserving the edges and reduce the classification uncertainty. In FCM-S classifier, it has been observed that for optimized values of λ and β are 0.7 and 3.5 respectively for AWiFS, LISS-III and LISS-IV datasets. To incorporate spatial contextual information with FCM, four DA-MRF models have been implemented. The hybridized model of DA approach has been devised to resolve the problems of mixed pixel and over-smoothing. Four objective functions of FCM with Discontinuity Adaptive (DA) prior known as FDM-(H1), FDM-(H2), FDM-(H3), and FDM-(H4) have been defined. It is observed that FDM-(H3) model produces highest accuracy of 95%. The basic advantage of hybridization DA model with FCM classifier is that classes are well classified and edges are not over smoothed. In PCM-S classifier, for λ=0.6 and β=3.0, is found to be more suitable to classify multiple land cover classes. In case PCM-DA-MRF model, it has been found that PDM-(H4) model produces higher accuracy for all cases of λ=0.5 and γ=0.5. The idea of using hybrid approach for soft classification i.e. Noise Clustering (NC) with contextual is a new approach which helps significantly to eliminate noise pixels while incorporating spatial contextual information. It has been found that for NC-S classifier, for λ=0.7 and β =3.5, class membership lies between 0.90 to 0.99 for all six classes selected for this study and that the computed entropy is lies between 0.005 to 0.65. This trend indicates that pixels of particular interest have been classified properly using context. To perform Noise Clustering with contextual classification with all four DA-MRF models, the optimized value of resolution parameter δ=105 has been taken. For NDM-(H1), NDM-(H2), NDM-(H3), NDMFuzzy Hybrid Approaches for Soft Classification of Satellite Images iv (H4) classifier, it has been found that the optimized value of λ and γ is 0.4. Further, it has been found that NDM-(H4) model produces highest accuracy in all cases. This classifier uses spatial contextual information in an appropriate manner which helps significantly in the removal of noisy pixels in remote sensing imagery. For FCMWE-S classifier, it has been found that for λ=0.4 and β =2.0, agriculture, barren land, moist land and water body produces highest membership. However, for λ=0.5 and β =2.5 and λ=0.6 and β=3.0, sal forest and eucalyptus plantation produces highest membership respectively. It has been found that for FCMWE with Discontinuity Adaptive Prior, the optimized value of λ and γ is 0.7. The accuracy values have been compared for all four DA models and it has been found that FDEM-(H1) model produces highest accuracy in all cases. While performing NCWE-S classification, it has been found that for λ=0.6 and β =3.0, all land cover classes produces the highest membership. In case of NCWE-DA-MRF model, the accuracy values have been compared for all four DA models and it has been found that NDEM-(H2) model produces highest accuracy in all cases. On the basis of SCM accuracy, the comparative performance analysis has been done for all the classifiers, and it is found that NDM (H4) classifiers produce the highest accuracy (99.79%) with minimum entropy (0.005). This output reflects that the combination of noise clustering classifier with contextual information using DA model produces least uncertainty value when compared to other classifiers. Thus, NDM-(H4) model is less affected by uncertainty and hence, it can be used to generate spectrally and spatially consistent thematic maps which preserve the edges between classes. The hybrid approach of soft classification based upon contextual is effective for the appropriate land cover identification and applicable for the multiple land cover identification at the same time. Thus, this study has explored the applicability of SA and DA MRF models for incorporating spatial contextual information of an image. The finding of the study illustrate that by utilizing appropriate hybrid classification strategy, accurate and meaningful land cover classifications can be produced from remote sensing imagery with minimum level of uncertainty. This study also suggests that suitable use of context allows the elimination of ambiguities, recovery of missing information and helps in correction of errors.