Discussion

The performance of each classifier is discussed in this section, with reference to IKONOS MS, IRS LISS-III MS, Landsat ETM+ MS and MODIS data.

From visual interpretation and comparison of the IKONOS classified image with the Google Earth and ground truth data, it was noticed that among the seven techniques used for classifying IKONOS images, SMAP performed best with 86.92% overall accuracy followed by RF with 85.25% accuracy (table 10). SMAP takes into account the intra class spectral variations and exploits spatial information among neighbouring pixels to improve classification results (Magnussen et al., 2004). NN was difficult to train before it reached convergence as evident from the training RMS plot in figure 6. It has wrongly classified asbestos roof to concrete built-up and was unable to detect the blue plastic roof (table 2). SMAP has aggregated some pixels of similar class types, creating blocky appearance as evident from the disappearance of the linear tarred road (flyover), merging with the adjacent concrete built-up. Inside the race course area, asbestos roof is mixed with open area in MLC classified image, blue plastic roof has been overestimated in KNN and SVM (Poly), open area is overestimated among the concrete buildings especially in race course in SVM (RBF) as seen in figure 7, bringing down its user’s accuracy. SVM (Poly) and SVM (RBF) have similar accuracies.

LISS-III data belongs to a region where forest and plantation are mixed because of degradation of natural forest and afforestation as plantation. Forest have exposed soil between trees and is therefore not dense to have spectrally significant different signature compared to plantation. As a result, the two classes often mix and create confusion in interpreting classification result. The area has dominant wasteland which has similar reflectance as builtup or barren land. Wasteland has often been misclassified as builtup and vice-versa. KNN has classified LISS-III imagery with maximum accuracy (table 5). Agriculture is overestimated in SMAP and much wasteland is misclassified as builtup in DT and RF (figure 8). Plantation class has increased in DT due to mislabeled forest pixels. SMAP and KNN have problems in classifying water bodies and have underestimated the same. The second best classifier followed by KNN (89.02%) in first position is SVM (RBF) with 88% accuracy. A NN with 4 hidden layers, 0.1 learning rate and 0.1 momentum with 5000 epochs was used for training the network in 14 seconds which was fourth best classifier with 83.66% overall accuracy after MLC (86.59%) in third position.

Landsat data having a spatial resolution of 30 m were classified most accurately using SMAP algorithm (89.03% overall accuracy, table 7). The area covered by this image has forested landscape, dominated by evergreen and semi-evergreen flora. KNN, SVM (Poly) and MLC gave higher overall accuracies of more than 85%. Plantation was overestimated in DT and NN which had 4 hidden layers with 0.2 learning rate and 0.2 momentum with 20000 epochs, took 77 seconds to train. A second degree polynomial function with gamma as 0.167 was used in SVM (Poly), which gave lower accuracies in detecting water bodies in comparison to other seven techniques. DT, NN, RF and SVM (RBF) showed abnormal trends and have classified mountain ridges as narrow water channels (figure 9). Wasteland are often mixed with fallow land due to seasonal differences in crop practices, and it also reflects similar to sand on the sea shores/sea beaches (Arabian sea on the west portion in the image), which were prominent in DT classification as depicted in figure 9.

Due to coarse spatial resolution of MODIS pixels (250 m), misclassification seems to be obvious because of mixed pixel problem that was unavoidable. With pixilated and patchy appearance of classified images, overall classification accuracy ranged from 59.5% (DT) to 76.21% (RF). When compared with the ground truth and resampled LISS-III MS classified image at 25 m spatial resolution, (figure 10), RF has performed best in classifying LU categories, followed by SVM (RBF), MLC and SMAP. DT has underestimated plantation, while NN has classified many wasteland pixels as builtup in the northern part of study area. Whatever be the performance of each technique, the classified MODIS images are not useful in obtaining accurate LULC information at regional level. Table 10 provides the best four classification techniques for various resolution data depending on their performance and overall accuracies.  

Table 10 illustrates that SMAP, KNN, SVM and RF are applicable at regional level LULC mapping using medium to high spatial resolution data. RF has outperformed DT and is also reported by Gislason et al., (2006). RF are increasingly being merged with other techniques such as object oriented approaches for achieving better classification accuracy (Watts and Lawrence, 2008), for hyperspectral data classification (Ham et al., 2005; Joelsson et al., 2005), for ecological prediction (Prasad et al., 2006), etc.

Table 10: Best performing classification algorithms for different sensors

However, the results obtained here are contrary to Otukei and Blaschke, (2010), who indicated that DT was better than SVM and MLC. SVM has provided promising results in LULC classification (Watanachaturaporn, et al., 2004; He et al., 2005; Kavzoglu and Colkesen, 2009; Watanachaturaporn, et al., 2007, 2008). In spite of many advanced classifiers that do not depend on the underlying data distribution, MLC, which works on the assumption of normality, has performed better than many other techniques.

This analysis showed that overall classification accuracy of the IKONOS images were between 69.84 and 86.92%. The highest overall accuracy obtained from SMAP (86.92%) with IKONOS bands is still lower than highest overall accuracy from IRS LISS-III (89.02%) using KNN, and 89.03% from SMAP with ETM+ image classification. One reason is that urban areas contain highly contrasting features that exhibit similar reflectance’s, creating spectral confusion between classes, such as building roofs of different types. Even with very high resolution data such as IKONOS MS, classification is not always easy and leads to poor accuracy in areas with mixture of land uses. ETM+ image classification ranged between 74% and 89% which are comparable to LISS-III image classification with reasonable good accuracies for mapping LU dynamics at a regional scale. These results are parameter specific and may not give similar output when the parameters are altered or adjusted. Classification errors in MODIS have occurred since the signal of the pixel is ambiguous or perhaps as a result of spectral mixing in pixels. Also, as the pixel resolution becomes coarser (in this case 250 m), the chance of high accuracy as the product of random assignment of values also declines. On similar lines, Quattrochi and Goodchild (1997) suggest that a fine-scale classification system is needed for a classification at local level. Thus high spatial resolution data such as IKONOS and SPOT 5 HRG are helpful. At a regional scale, medium spatial resolutions such as Landsat TM/ETM+, Terra ASTER are most frequently used data. At a continental or global scale, coarse spatial resolution data such as AVHRR, MODIS, and SPOT Vegetation are preferable. However, a recent study by Eva et al., (2010) demonstrates the usage of medium spatial resolution satellite imagery (Landsat-5 TM and SPOT-HRV) for monitoring forest areas from continental to territorial levels. Inter-sensor comparison between Resourcesat, LISS-III, LISS-IV and AWiFS with reference to coastal LULC (mudflat, mangroves, vegetated dune, coastal water, etc.) were carried out on the basis of DN values, converting radiance and reflectance values of each sensor (Chauhan and Dwivedi, 2008). The study revealed that LISS-IV can be used in place of LISS-III or it can be merged (LISS-III MS + PAN). Comparison of AWiFS and LISS-IV did not show good correlation as with LISS-III, because of the large difference in spatial resolutions.

Spatial resolution is an important factor that affects classification details and accuracy (Chen et al., 2004; Velpuri et al., 2009) and influences the selection of classification approaches (Atkinson and Aplin, 2004), since the size of ground objects relative to the spatial resolution of a sensor is directly related to image variance (Woodcock and Strahler, 1987). When the object in the scene becomes increasingly smaller relative to the pixel resolution, they are no longer regarded as individual objects. In such cases, reflectance is treated as a sum of interactions among various classes as weighted by their relative proportions (Strahler et al., 1986). Therefore, medium spatial resolution data such as IRS LISS-III and Landsat are useful at a regional scale but not appropriate at a local level.

For fine spatial resolution data, although mixed pixels are reduced, the spectral variations within classes decrease the classification accuracy (Lu and Weng, 2007). Therefore, selection of a suitable data mining algorithm must consider important factors such as aim of classification, classification accuracy, algorithm performance, computational resources and effective separation of classes. Lu et al., (2004) and South et al., (2004) argue that in many cases, contextual-based classifiers, per-field approaches, and machine-learning approaches provide a better classification result than MLC, although some tradeoffs exist in classification accuracy, time consumption, and computing resources. A combination of multi-sensor data with varying image characteristics, considering the economic conditions, are also important factors that affect the selection of RS data, classification technique and time affecting the quality of classification results.

Uncertainties involved in different stages of classification procedures influence classification accuracy, as well as the area estimation under different LULC classes (Dungan, 2002). Understanding relationships between classification stages, identifying the probable factors influencing the accuracy and improving them are essential for successful image classification. For example, spatial or radiometric resolutions constraints, geometric rectification of multi-sensor data and the atmospheric condition during image acquisition time causes uncertainty. On the other hand, the algorithm used for calibrating atmospheric or topographic effects may cause radiometric errors (Lu and Weng, 2007). Dungan (2002) listed five types of uncertainties in RS data processing such as positional, support, parametric, structural (model) and variables. Friedl et al., (2001), emphasises on three sources of errors, namely, image acquisition process, during data processing and the interaction between resolution and the scale of ecological processes. Yu et al., (2008) has discussed the factors affecting spatial variation of classification uncertainty in an image object-based vegetation mapping. Such errors have to be considered while handling coarse resolution data such as MODIS, due to existence of many mixtures among LULC classes. In such cases, geocomputation (Beekhuizen and Clarke, 2010), geo-visualisation and interactive visualisation techniques (Lucieer and Kraak, 2004) have proved to be useful.

Spectral characteristic of RS data needs to be considered during classification using data mining algorithms. As spatial resolution increases, texture or contextual information play vital role and turn out to be decisive factors. Classification approaches may vary with varying RS data. For instance, with high spatial resolution data such as IKONOS, the impact of shadow resulting from topography and trees and wide spectral variation within the LU classes may outweigh the advantages of high spatial resolution during classification (Lu and Weng, 2007), where contextual based classification such as SMAP may be better as evident from the present study (table 3). In this case, a combinatorial approach of spectral and textural information can reduce the problem. For medium to coarse resolution data, however, spectral information is more important than spatial where per-pixel classification may not be appropriate. In such cases, image fusion or sub-pixel classification would be more suitable. For a particular study, it is often difficult to identify the best classifier due to lack of a guideline for selection, and the availability of suitable classification algorithm. Comparative study of different classifiers such as the one conducted here, would certainly add to the already existing information on spatio-spectral image classification. Next, we discuss a new Hybrid Bayesian Classifier.