Land cover classification from Remote Sensing data

Land cover classification from Remote Sensing data


Manoj K Arora
Department of Electrical Engineering and Computer Sciences,
Syracuse University
[email protected]
Availability of accurate and up-to-date land use land cover information is central to many resource management, planning and monitoring programmes. For example, the land cover is a desired input parameter for a number of agricultural, hydrological and ecological models. Traditional methods of land cover mapping have been limited to field surveys that are time-consuming and uneconomical with data collected over long time intervals. The methods are particularly inefficient and impractical for real-time global and regional land cover mapping (Defries and Townshend, 1994).

Satellite sensor remote sensing images due to their synoptic view, map like format and repetitive coverage are a viable source of gathering effective land cover information. Figure 1 shows an IRS 1C image from PAN sensor. Typically, the pixels of the remote sensing image are grouped into meaningful and homogeneous land cover classes using digital image classification. Though remote sensing has long been championed for the provision of actual land cover information, there are a range of factors that restrict the ability of remote sensing to accurately extract the information at various scales (Arora and Foody, 1997). Failure to understand these factors may result into inappropriate land cover classifications. This paper addresses some of the issues related to land cover mapping from remote sensing data.

Fig. 1: IRS 1C PAN image acquired on 3.4.2000
(A portion of area in Jalpaiguri district, West Bengal)

Digital image classification
Digital image classification is the process of assigning a pixel (or groups of pixels) of remote sensing image to a land cover class. The objective is to classify each pixel into only one class (crisp or hard classification) or to associate the pixel with many classes (fuzzy or soft classification). The classification techniques may be categorized either on the basis of training process (supervised and unsupervised) or on the basis of theoretical model (parametric and non-parametric). Several classification algorithms (classifiers) have been developed under this categorization. For instance, Maximum Likelihood Classifier (MLC) is a supervised parametric algorithm whereas k-means clustering is an unsupervised parametric algorithm.

Generally, supervised classification is performed that has three distinct stages namely training, allocation and testing. Training is the identification of a sample of pixels of known class membership gathered from reference data such as ground truth, existing maps and aerial photographs. These training pixels are used to derive various statistics (e.g. mean and standard deviation) for each land cover class and are input to the second stage of the classification. In this stage, the pixels of the image are allocated to the class with which they show the greatest similarity based on the statistics. After classification in the allocation stage, the accuracy of classification is determined in the final stage. A sample or group of testing pixels is selected and their class identities are compared on both the classified image and the reference data. The pixels of agreement and disagreement for each testing sample are represented in the form of an error matrix, which can then be used to evaluate the classification accuracy.

Fig. 2: Crisp Classification of IRC 1C PAN image using MLC
Issue related to land cover classification from Remote Sensing
The digital image classification as such seems to be a simple process but in reality there are complications that limit the accuracy of land cover classification. These may arise partly due to the characteristics of the remote sensing images and the assumptions underlying the techniques employed in the classification process (Mather, 1990). Some thoughts on these issues are now presented.

Geometric registration of the remote sensing image
Generally, a remote sensing image is not used directly in raw form due to the presence of certain geometric errors in it. These errors may occur due to the movements of the satellite platforms that carry sensors. The image therefore may have to be geometrically rectified (registered) to extract useful information particularly for change detection studies, and for its input as a data layer to GIS. If accurate geometric registration is not achieved, then spurious differences will be detected, arising just because different locations are compared.

Often a polynomial transformation coupled with a resampling procedure is utilized to geometrically register the image to another geographical data sets through the knowledge of the actual positions of Ground Control Points (GCP) in both the data sets. However, the process of geometric registration may also have its own limitations due to the possibility of errors in GCP location, the number and distribution of GCP, the accuracy of polynomial transformation and the resampling procedure (Mather, 1995). Resampling is particularly important in classification as it alters the digital numbers of the pixels. Therefore, these things must be given due consideration to achieve sub-pixel registration accuracy that is often desired to obtain reliable estimates of land cover particularly for change detection applications.

Fig. 3: Fraction images produced from fuzzy classification of IRS 1C LISS III image of Jalpaiguri district, West Bengal. (MLC: maximum likelihood classifier, FCM: Fuzzy c-means classifier). The variation in grey shades indicate the class proportions (black: 0%, white: 100%)
Choice of land use land cover classification scheme
A classification scheme such as USGS land use land cover classification (Anderson, 1976) provides the number and definition of land cover classes to be mapped. Poorly defined or ambiguous class definition may have a considerable effect on the classification accuracy. Moreover, the choice of an appropriate classification scheme may also direct the user to select appropriate spectral, spatial and temporal resolutions of the data.

In general, the classification scheme should be internally consistent and applicable over regional to global scales. Often the choice of a classification scheme may be application-dependent, yet sometimes two or more classes may be merged to meet the desired accuracy standards. Failure to understand a classification scheme in the beginning may result in great loss of time and accuracy of classification afterwards. Therefore, the definitions of the classes should be given proper attention.

Selection of appropriate number of waveband
The remote sensing data are usually obtained in a number of wavebands, which are often significantly correlated. Therefore, it may be lavish to utilize the data from all the wavebands thereby increasing the computational time and cost of classification. The situation is more complex with hyper-spectral sensors such as Airborne Visible and Infra-Red Imaging Spectrometer (AVIRIS) providing data in 224 wavebands. With this hyper-spectral data, certain classifiers may also not perform adequately as it would be difficult to get enough training samples to provide reliable estimates of training data statistics (Roger, 1996). Hence, it may be advisable to select the most effective bands which really are sufficient in discriminating various land cover classes and efficient in reducing the computational time. The approaches such as divergence analysis and principal component transformation may be used.

Training data characteristics
Training is a critical step in supervised image classification. As the training samples should be representative of the land cover classes, these are collected from relatively homogeneous areas on the ground. Therefore, these are chosen subjectively and deliberately away from mixed pixels containing two or more classes. Mixed pixels are a major problem in land cover classification and is one of the key issues addressed here subsequently. Nevertheless, the collection of training data is generally a costly affair and, thus, the size of the training data should be kept small (Arora et al., 1998). Conversely, it should be large enough to accurately characterize the classes. Typically, the size of the training samples is related with the number of wavebands (Swain and Davis, 1978).

The timing of training data collection may be crucial to some studies. Ideally, the training data should be collected at or near the time of the satellite pass to accurately characterize the land cover classes on the image. Practically, however, this may not be the case due to financial, time and resource constraints. Nonetheless, the temporal requirements differ from application to application.

Type of classifier
Numerous classifiers have been developed each with their own merits and demerits. MLC is the most commonly used classifier than any other parametric classifiers despite its longer computational time and inherent assumptions. It assumes that the input data (i.e. training data) are normally distributed and independent. Once these criteria are met, the MLC is well suited for accurate classifications. Many a times, the spectral properties of the land cover classes are far from the assumed distribution (e.g. in complex and heterogeneous environments) where the use of MLC may not be appropriate. Further, the MLC is based on one pixel one class allocation phenomenon thereby forces each pixel to one land cover class thus producing crisp classification output (Figure 2). In contrast, the pixel (e.g. mixed pixel) may correspond to two or more land cover classes. Other approaches are, therefore, required to classify mixed pixels (see later).

In the last decade, neural network classifier has been viewed as a replacement for the most widely used MLC. Its major appeal is mainly based on its distribution free assumption as well as its capability of handling the data from other sources besides remote sensing. Typically, a neural network consists of an input layer, a hidden layer and an output layer. The units in the input layer equal the number of variables (e.g. number of wavebands) used in the classification. The output layer produces the network’s results that denote the various land cover classes. Introducing hidden layer(s) between input and output layers increases the network’s ability to model complex functions. Selection of appropriate number of hidden layers and their units is, however, critical for successful implementation of the neural network (Arora et al., 2000) and must be given due consideration. Some other classifiers such as evidential reasoning (Gong, 1996) and support vector machines (Brown et al., 1999) have also been used recently for land cover classifications. However, their applications have been limited due complex heuristics that may be difficult to understand.
Inclusion of ancillary data in image classification
Many a times, digital classification solely on the basis of statistical analysis of spectral reflectance values (i.e. digital numbers) may not truly represent the ground actuality. Therefore, some form of additional information in the classification process is quite useful. Data from other sources such as soil type, Digital Elevation Models (DEM), and geophysics etc. may be incorporated. For example, remote sensing image is an ideal source for land cover mapping in hilly areas as it minimizes the accessibility problems. To reduce the effect of shadows due to hills in the remote sensing image, the inclusion of information from DEM can be advantageous. MLC has the limitations to handle multi-source data but evidential reasoning and neural network classifiers may be useful as they permit the information to be gathered consistently and objectively from data of varying type, format and scale of measurement (Peddle et al., 1994).

However, the multi-source data may contain different types of errors, which may be difficult to model when assessing the accuracy of the end products. Therefore, utmost care must be taken if these approaches are utilized. Nevertheless, once the ancillary information has been incorporated effectively, the accuracy of classification may be increased (Arora and Mathur, 2001).

Problem of mixed pixels
Another source of error encountered in image classification is the presence of mixed pixels. These usually occur near the boundaries of two or more classes and where the land cover classes are continuous and inter-grade gradually (Foody and Arora, 1996). The mixed pixels are a common phenomenon in coarse resolution images from sensors such as Wifs, AVHRR, MODIS and MERIS providing data at resolutions ranging from 188 m to 1.1 km, which have an enormous potential for global land cover mapping. Using coarse resolution data provides efficiency but reduces accuracy as these may be contaminated by mixed pixels. Conversely, the fine resolution data may be used but the volume of data increases enormously. Therefore, a trade-off may have to be maintained. Nevertheless, the mixed pixels may exist in fine resolution data also as additional details may be resolved. Thus, a particular pixel may be treated as a mixed pixel within the context of a detailed classification scheme, while the same pixel may be called pure in the context of a more generalized classification scheme.

Whatever is the origin of mixed pixels, these create problems in the image classification. Since the mixed pixel displays a composite spectral response of the component classes, the pixel may not be allocated to any of the component classes by the crisp classifiers and thus alternatives are sought.

Although, MLC has generally been used as a crisp classifier, its a posteriori probabilities may indicate the partial and multiple class membership within each mixed pixel (Foody et al., 1992). Another statistical technique namely Linear Mixture Modeling (LMM) may be used to unmix the land cover classes within a mixed pixel (Settle and Drake, 1993; Tiwari et al., 1999). It is based on the theory that the spectral response of an individual pixel is the linear sum of the spectral responses of the component classes of the pixels weighted by their relative proportions. However, to adopt either MLC or LMM, certain assumptions regarding the data must be satisfied which often are untenable. Alternative classifiers such as fuzzy c-means (Bezdek et al., 1984) and neural network (Foody, 1996), which do not depend upon any assumption, may be more appropriate for fuzzy classification. Unlike the crisp classifications, the outputs from fuzzy classification are a set of fraction images equal to the number of classes (Figure 3). The fractions images illustrate the percentage of class proportions within a pixel from 0 to 100%.

Reference data characteristics for accuracy assessment
Image classification is incomplete until an accuracy assessment has been performed. This serves as the basis for analyzing the errors that may creep in during the classification process due to causes mentioned above. To determine the classification accuracy, sample testing data are desired which are usually collected from the reference data. The testing pixels should be independent of the training pixels. Yet in many studies, the same data sets have been used for training and testing resulting in an optimistic bias in classification accuracy (Hammond and Verbyla, 1996). To get an unbiased estimate of accuracy, the testing samples should be mutually exclusive and collected randomly. Hence, the choice of sampling schemes and sample size are important considerations to obtain a statistically valid testing sample. Arguments for and against a sampling scheme may be made but the decision to choose a scheme is often application and data dependent. Still, the analyst must understand the statistical properties and assumptions underlying the sampling scheme used (Stehman, 1997). The sample size depends on many other factors such as number of classes to be mapped and their relative importance, percentage of area covered by each class, and cost of field data collection. Therefore, often the practical considerations dictate the sample size. However, the objective should be to balance the statistical recommendations and practical considerations.

Another important matter related to reference data is the errors in the reference data themselves as these are derived from different sources at varied accuracy such as existing maps, aerial photographs, GPS surveys or a combination of all. The maps may have inherent errors arose while compiling them and may be out-dated. The aerial photography has the potential advantages as it may be acquired during the same time as the satellite pass but may have errors due to tilt and relief displacements resulting into positional inaccuracies. GPS surveys may be seen as an alternative solution but again are expansive and bring in other errors (Tiwari et al., 2000). Thus, it may be seen that whatever be the source of the reference data collection, the data are often not error free. Any error in the reference data may result into a conservative bias in the classification accuracy (Verbyla and Hammond, 1995).

It is clear from above that there are a number of issues that affect the land cover mapping from remote sensing data. Failure to understand these may result into an inappropriate land cover representation. It is essential to state the requirements as precisely as possible in terms of accuracy and precision for effective land use land cover information from remote sensing data.


  • Anderson, J. M., Hardy, E. E., Roach, J. T., and Witmert, R. E.,1976, A land use classification system for use with remote sensing data, Geological Survey Professional Paper 964, Washington D.C., 28p
  • Arora, M. K. and Foody, G. M. (1997). Log-linear modelling for the evaluation of variables affecting the accuracy of probabilistic, fuzzy and neural network classifications. Int. J. Remote Sensing, 18: 785-798.
  • Arora, M. K., Narsimham, G., and Mohanty, B. (1998). Neural network approach for the classification of remotely sensed data. Asian Pacific Remote Sensing and GIS Journal. 10: 11–18.
  • Arora, M. K., Tiwari, K. C. and Mohanty, B. (2000). Effect of neural network variables on image classification. Asian Pacific Remote Sensing and GIS Journal. 13: 1 – 11.
  • Arora, M. K. and Mathur, S. (2001). Multi-source image classification using neural network in a rugged terrain. Geo Carto International. 16: 37 – 44.
  • Bezdek, J. C., Ehrlich, R., and Full, W., 1984, FCM: The fuzzy c-means clustering algorithm, Computers and Geosciences, 10:191-203.
  • Brown, M., Gunn, S. R. and Lewis, H. G., 1999, Support vector machines for optimal classification and spectral unmixing, Ecological Modelling, 120: 167-179.
  • DeFries, R. S. and Townshend, J. R. G. (1994). Global land cover: comparison of ground-based data sets to classification with AVHRR data. In Environmental Remote Sensing from Regional to Global Scales. Edited by G. M. Foody and P. J. Curran (Chichester: Wiley). 84-104.
  • Foody, G. M., Campbell, N. A., Trodd, N. M. and Wood, T. F., 1992, Derivation and applications of probabilistic measures of class membership from maximum likelihood classification, Photogrammetric Engineering & Remote Sensing, 58: 1335-1341.