Introduction

Classification of remote sensing (RS) data accurately is a requirement for many applications such as global change research (Nemani et al. 2011), geological research (Smith et al. 1985), wetlands mapping (Ramachandra and Kumar 2008), crop estimation (Pacheco and McNairn 2010), vegetation classification (Hostert et al. 2003), forest classification (Huang et al. 2008), urban studies (Pu et al. 2008), feature extraction (Do´pido et al. 2011), land cover (LC) change (Turner et al. 1994), Earth system modelling (Collins et al. 2011), etc. Satisfactory classification of RS data depends on many factors including
(a) the characteristics of study area,
(b) availability of suitable RS data,
(c) ancillary and ground reference data,
(d) proper use of variables and classification algorithms,
(e) user’s experience with reference to the application and
(f) time constraints (Lu and Weng 2005).
Furthermore, diverse landscapes and terrain types have a mixture of both homogeneous and heterogeneous LC classes and require supplemental environmental or geographical layers for improved classification accuracies. Increased spectral variation is common with high degree of spectral heterogeneity; for example, urban landscapes are composed of features having a complex mix of buildings, roads, trees, lakes, lawns, concrete, etc., often responsible for low classification accuracy. In landscapes with mountains and dense forests, problems arise due to changes in elevation, topographic differences and often shades (shadows) produced by hillocks and long trees due to altitudinal variations, which is a major challenge for selection of suitable image processing approach. Fine spatial resolution data such as IKONOS Multispectral (MS) and Panchromatic (PAN) often lead to high spectral variation within the same LC class resulting in poor classification performance (Lu et al. 2010). Therefore, reducing the spectral variation within the same LC and increasing the separability of different LC types are the keys for improving LC classification (Lu and Weng 2007). In practice, data acquired from medium spatial resolution sensors such as Landsat TM/ETM+ or IRS LISS-III, being readily available for multiple dates, are commonly used for most landscape analysis (urban and forested terrain at a regional scale). In this regard, different approaches such as multi-sensor data integration (Haack et al. 2002), full spectral image classification (Stuckens et al. 2000; Shaban and Dikshit 2001), expert classification (Hung and Ridd 2002), etc. have been used. Traditional per-pixel spectral-based supervised classification is based only on spectral signatures, and does not make use of rich spatial information inherent in the data (Lu et al. 2010). Therefore, making full use of RS information along with ancillary information (acquired or derived environmental layers) would be an efficient way to improve classification accuracy. For example, Na et al. (2010) used 103 geographical layers to show improvement in LC mapping using Landsat TM bands 1–5 and 7, NDVI (Normalised Difference Vegetation Index), EVI (Enhanced Vegetation Index), a data fusion transformation combining the six bands information from the Landsat TM image (first principal component—PC1) as additional predictors, image texture measures (variance, homogeneity, contrast, dissimilarity and entropy) with window size of 3 * 3 pixels and 11 * 11 pixels, DEM, slope and soil type with Random Forest (RF), Classification and Regression Tree (CART) and Maximum Likelihood Classifier (MLC) based classification. Among these, RF yielded accurate classification with an overall accuracy of 91% and kappa 0.8943. They also quantified the effect of training set size on the performance of classification algorithms. Xiaodong et al. (2009) integrated TM data with NDVI, EVI, PC1, slope, soil types and five texture measures (variance, homogeneity, contrast, dissimilarity and entropy) for LC classification of Marsh Area using Classification Trees and MLC. They concluded that image spectral, textural, terrain data and ancillary Geographical Information System (GIS) improved the land use land cover (LULC) classification accuracy significantly. Fahsi et al. (2000) evaluated the contribution and quantified the effectiveness of DEM (digital elevation model) in improving LC classification using Landsat TM data over a rugged area in the Atlas Mountains, Morocco which considerably improved the classification accuracy by reducing the effect of relief on satellite images, increasing the individual accuracies of the different classes by up to 60%. Recio et al. (2011) used historical land use (LU) and ancillary data and showed improvement in overall classification accuracy considered case-by-case for each class. Masocha and Skidmore (2011) used DEM along with ASTER imagery and georeferenced point data obtained from field to increase the accuracy of invasive species (Lantana camera) mapping using Neural Network and SVM (support vector machine) hybrid classifiers. The overall accuracy increased from 71% (kappa 0.61) to 83% (kappa 0.77) with Neural Network and from 64% (kappa 0.52) to 76% (kappa 0.67) with SVM hybrid classifiers. Dorren et al. (2003) studied the effect of topographic correction and the role of DEM as additional band using per-pixel and object based classification to classify forest stand type maps using Landsat TM data in a steep mountainous terrain. They concluded that both topographic correction and classification with DEM as additional band increased the overall accuracy. Xian et al. (2008) quantified multi-temporal urban development characteristics in Las Vegas from Landsat and ASTER Data. Apart from the satellite imageries, NDVI, slope, aspect and temperature were used for classification. Lu and Weng (2005) demonstrated urban classification using full spectral information of Landsat ETM+ imagery in Marion County, Indiana using PC’s of ETM+ MS bands, texture, temperature and data fusion of MS and PAN. They concluded that texture and temperature may improve classification accuracy for some classes, but may degrade for other classes. Data fusion of MS and PAN are useful but high spatial resolution also increases spectral variation within the classes, decreasing the classification accuracy. Data fusion combined with texture significantly improves classification accuracy.

In this study, three different terrain types with varying characteristics in the Indian context are considered for classification with Landsat ETM+ MS bands and several other ancillary and derived layers using Random Forest (RF) classifier. The role of vegetation indices such as NDVI and EVI, elevation and derived layers (slope and aspect), temperature, texture (Haralick et al. 1973; Haralick 1979; GRASS GIS 2017), and addition of PAN band in addition to MS bands are examined in the process of image classification.