EVALUATION OF SOFT CLASSIFIERS FOR REMOTE SENSING DATA

Abstract

For a large country like India, mapping is often carried out at regional level. This necessitates the use of remote sensing images at coarse spatial resolution acquired from sensors onboard a number of satellites. However, in coarse resolution images, occurrence of mixed pixels at the scale of measurement is a major problem. The mixed pixels are treated as noise or uncertainty in class allocation of a pixel. Conventional hard classification may thus produce inaccurate classification. The application of sub-pixel or soft classification methods such as those based on spectral mixture analysis, fuzzy set theory and artificial neural network may be adopted for classification of images acquired in complex and uncertain environment. The outputs from these methods are a set of class membership values for each pixel, also called as soft, fuzzy or sub-pixel classification. The main objective of this research endeavour is to evaluate the utility of some recently introduced soft classifiers for remote sensing image classification. From amongst a number of soft classifiers, this study has focussed on two statistical classifiers - maximum likelihood classifier (MLC) and linear mixture model (LMM), two fuzzy set theory based classifiers - fuzzy c-means (FCM) and possibilistic c-means (PCM), and two neural network classifiers - back propagation neural network (BPNN) and competitive learning neural network followed by learning vector quantizers (CLNN-LVQ). Due to its wide acceptance, MLC has been used as a bench mark to evaluate other soft classifiers studied in this research. A new fuzzy set clustering algorithm, namely possibilistic c-means (PCM) has been introduced in this research to get over the major limitation of FCM (i.e. probabilistic sum to one constraint). BPNN has commonly been used in remote sensing Evaluation of Soft Classifiers for Remote Sensing Data as hard classifier but has been implemented here to produce soft classification outputs. The CLNN-LVQ classifier finds its first use in this research to be implemented as soft classifier. The use of the soft classifiers however partially resolves the problem of mixed pixels as these have been accounted for in the allocation stage only. For images dominated with mixed pixels, their incorporation in other stages may also be necessary. Suitable modifications have been made in the training stage of all supervised classifiers in order to accommodate mixed pixels in this stage, which is one of the main objectives of this research. In the testing stage, mixed pixels have been incorporated via the use of distance, entropy and fuzzy set based measures. The hypothesis of fuzzy error matrix has been promoted to assess the accuracy of soft classification. As the formulations of majority of these classifiers and accuracy measures in the existing commercial image processing software are not available, an indigenous interactive software in Visual C++, abbreviated here as Soft Classification Methods and Accuracy assessment Package (SCMAP), has also been developed. This software consists of all the four basic modules of classification - display, training, classification and accuracy assessment modules. A published artificial dataset has been used to validate the various modules of the software. Thereafter, the soft classifiers have been evaluated via a number of classifications produced from remote sensing data acquired by Indian Remote Sensing (IRS) IB satellite LISS II sensor (spatial resolution: 36.25 m). The area consist of a mixture of five dominant land cover classes - built-up land, grass land, trees, agriculture and barren land. The IRS 1C PAN image derived reference map registered to LISS II image to an accuracy of 1/5 of a ———__^ Abstract pixel, has been used to create hard and soft reference data for accuracy assessment ofhard and soft classifications. Several experiments were conducted to evaluate the soft classifiers and accuracy measures investigated in this study. Initially, Shanon's entropy and the index of fuzziness were used to quantify the uncertainty in class allocation. These measures showed that the data considered inthis study was highly uncertain, which was a challenge for the evaluation ofsoft classifiers investigated. Subsequently, the appropriate accuracy measures for the assessment of both hard and soft classifications were identified. Since overall accuracy has been widely accepted, therefore this accuracy measure was used for the assessment of hard classification in this study. Similarly, from a wide choice of measures for the accuracy assessment of soft classification, the fuzzy error matrix (FERM) based measures were used. In fact, formulation ofFERM is such that it can also be used for the evaluation of hard classification and has been recommended. The results showed that the distribution free classifiers based on fuzzy set and neural network produced more accurate classification than the statistical classifiers. An improvement in accuracy of 8% to 12% was observed. It was shown how PCM classifier was robust to the existence of noise (pixels with high degree of mixtures) in the data. From among various unsupervised classifiers, CLNN-LVQ, introduced here for soft classification, produced the highest classification accuracy of 53.89% and showed an improvement of more than 5%over the FCM classifier. A comparison of the accuracy of hard classification produced from supervised classifiers however showed an improvement in the accuracy of the order of 6% to 15%over the unsupervised classifiers. Thus, supervised classifiers may be preferred over unsupervised in Evaluation of Soft Classifiers for Remote Sensing Data classifiers. The accuracy of hard classifications was further increased by including a priori probabilities in BPNN and MLC classifiers. Anew approach to include a. priori probabilities by way of replicating the training data of a class in accordance with the proportional area covered bythat class on ground was suggested. The accuracy of BPNN classifier increased by 20% whereas the accuracy of MLC increased by7%on the inclusion of apriori probabilities. Evaluation of soft classification through FERM based measures led to an improvement of the order of 20% in the accuracy of the classification over the accuracy determined from traditional error matrix based measures for the same classification. Thus, it is recommended that soft classification outputs from any classifier should not be hardened for evaluation purposes, as this may results into loss of information. LMM as soft classifier produced the lowest accuracy whereas BPNN and PCM as soft classifiers produced the highest map accuracy of about 73%, which was an improvement of 20% over the highest accuracy achieved by the unsupervised classifiers. When the images are dominated by mixed pixels, their incorporation not just in allocation stage through generation of soft outputs, but also in training and testing stages were also assessed. The results showed that by properly accounting for mixed pixels in all stages, same level of accuracy could be achieved as would have been obtained by using pure pixels in all stages. Thus, the findings of this research illustrate that by utilising appropriate classification strategies, accurate and meaningful land use land cover classifications can be produced from remote sensing images fraught with mixed pixels. IV