Classification

Definition[edit]

Classification in GIS, more specifically in remote sensing, refers to the systematic process of categorizing spatial data into discrete and ideally meaningful and robust classes. It is one of the most widely used methods to interpret and analyze remotely sensed imagery, enabling the transformation of raw sensor measurements into thematic information about the Earth’s surface. There are diverse ways of classification, which can range from cluster analysis to supervised classifications, all of which can be lump-summed as part of the nexus of machine learning. The primary purpose of classification is to simplify complexity. Millions of pixels, each containing reflectance values or other attributes, are reduced to a manageable number of categories, for example as land cover classes (e.g. forest, cropland, urban, or water), eroded environments (e.g. by identifying bare soil signatures) or by time series analysis (e.g. the shrinking of a glacier over time). This process makes it possible to investigate spatial patterns from complex large data sets as well as produce land use maps, monitor environmental change, assess natural resources, or guide spatial planning and policy decisions. In essence, classification translates continuous spatial data into a tangible form that can be used for analysis, communication, and decision-making across disciplines and into society as pathway towards actionable measures.

Background[edit]

First Classifications[edit]

Before the digital era, classification in GIS and remote sensing was largely based on visual interpretation. Analysts manually inspected aerial photographs, identifying land cover classes using texture, tone, shape, spatial context and local ground knowledge. This approach was subjective and labor-intensive but laid the foundation for systematic categorization. The launch of Landsat 1 (at that time known as the Earth Resources Technology Satellite (ERTS)) on July 23rd, 1972, the first civilian Earth observation satellite, marked a turning point. For the first time, large-scale, multispectral data were continuously available. This spurred the development of digital classification techniques, moving from human judgment to algorithmic processing.

Emergence and diversification of Satellite Missions[edit]

As the volume of large-scale Earth observation data grew, classification became more established, and were typically automated approaches. Missions such as NOAA’s AVHRR, SPOT, and later MODIS and Sentinel satellites produced vast datasets -particularly when considering the computer resources at the time- at varying spatial, spectral and temporal resolutions. These advances expanded the scope of classification from local land cover mapping to global monitoring, supporting studies of deforestation, urban expansion, biodiversity, and climate change impacts. Notable studies are …

The Method[edit]

At its core, classification is about reducing real-world complexity. The Earth’s surface and its nuances are dynamic, heterogeneous, and difficult to capture in full. Classification translates this complexity into tangible, structured spatial information that can be stored, analyzed, and compared. For example, a pixel in a satellite image may record reflectance in several spectral bands. Alone and without a relative context these values are abstract. Through classification, values become an interpretable unit of information within a system that can be aggregated and analyzed at scale. These classes can be predefined, such as “agricultural land,” or they can be inductively generated by grouping similar spectral signatures that are then interpreted by the analyst. The end result is a thematic dataset that represents the spatial distribution of often predefined categories.

Types of Classification[edit]

Different approaches to classification exist, each with its own logic, requirements, benefits and applications. The two main types widely used in GIS are unsupervised and supervised classification, which differ in how classes are determined and applied. In addition, there are other classification approaches that do not rely solely on raw spectral clustering or training data but instead use thresholds, indices, or object-based methods.

Unsupervised Classification[edit]

Unsupervised classification is a method in which the algorithm automatically groups pixels based on their similarity in spectral or statistical space. This approach does not require prior knowledge of the landscape or labeled training data. Instead, algorithms are applied similar to clustering methods in statistics, which are then interpreted and labeled after the fact. It begins by specifying the number of clusters, or by allowing the algorithm to determine them dynamically. Pixels with similar reflectance values into these prechosen clusters. Noteworthy examples are k-means clustering and ISODATA clustering. Both require the analyst to specify an initial number of classes. K-means produces exactly the number of classes requested, while ISODATA can iteratively split or merge clusters based on statistical criteria. This makes ISODATA more adaptive to the underlying data distribution, though it may not result in the exact number of classes originally specified. Once the algorithm has produced its clusters, the analyst examines them in relation to reference information, such as field knowledge or high-resolution imagery, and manually assigns them to context dependent specific classes. However, sometimes it is not obvious which clusters belong to which classes. In such cases, it may be advisable to re-execute the algorithm or clustering method. Hence, the process can be tedious and desired cluster-to-class matches may not be achieved at all. Ultimately, unsupervised classification is particularly useful when little or no ground truth data are available. Hereby, this method serves as an exploratory tool for investigating spatial data patterns. However, the clusters generated by the algorithm may not perfectly correspond to real-world categories, and considerable effort may be required to interpret and label them meaningfully. Subsequently, rigorous accuracy assessment needs to be performed to test for representing real world features.

Supervised Classification[edit]

Supervised classification relies on the knowledge of the analyst to guide the classification process. In this approach, the analyst provides training samples—sets of pixels that are known to belong to specific classes such as forest, water, or urban areas. These training samples can be vector data, where each feature represents a training data point. These samples form the basis for the algorithm to classify the remainder of the dataset. The process generally follows a structured sequence. First representative training data for each class of interest are gathered, either from field surveys or by identifying and selecting features in (high-resolution) imagery. An algorithm then analyzes the spectral or statistical characteristics of these training samples. Using methods such as maximum likelihood, minimum distance, support vector machine, or random forest, such algorithms assign every other pixel to the class it most closely resembles. The quality of a classification output is expressed through its accuracy, which indicates the degree to which the classified results correspond to real-world features observed in the imagery. Hence, an incremental follow-up step is validating the results against independent reference data to assess their accuracy. Here is an article regarding accuracy assessment. In this context, the classification accuracy depends on the training data set. To improve classification results, a feasible way is to increase the number of training data points, either across all classes, for specific classes with higher error rates, or in geographic areas where inaccuracies are more pronounced. This way the used algorithm has more information to work with and correctly assign pixels to classes. As a result, supervised classification can be an iterative and lengthy process, where one might need to perform classifications and reevaluate the training data in multiple rounds until satisfactory results are achieved. Nevertheless, supervised classification often achieves higher accuracy than unsupervised methods, particularly when training data are representative and abundant. However, it is also sensitive to bias: poorly chosen training samples may lead to systematic misclassification.

Other Types of Classification[edit]

Beyond supervised and unsupervised approaches, other methods of classification exist. A common example is the classification of indices (e.g. Normalized Difference Vegetation Index (NDVI), Normalized Difference Snow Index (NDSI), Normalized Difference Moisture Index (NDMI), Normalized Difference Building Index (NDBI)). In this approach, values from -1 to 1 are calculated for each pixel and then divided into categories, such as high vegetation density, medium vegetation density, low vegetation density, and bare soil for NDVI. While classification usually involves some form of algorithm, this approach is a simple matter of grouping or segmenting values defined by the researcher but still results in interpretable classes. Pixels are assigned to categories based on their index values, creating a thematic map that expresses relative conditions rather than strict land cover types. Another approach is object-based image analysis (OBIA), in which imagery is segmented into meaningful objects or shapes, rather than individual pixels. These objects are then classified based on spectral, spatial, or contextual information. Similarly, rule-based systems can be used to define classes through explicit conditions, for example, by assigning a pixel to “wetland” only if it has both a high vegetation index and proximity to water. These alternative approaches can be used in combination with supervised or unsupervised methods, providing flexibility and improving classification outcomes in complex landscapes.

Applications in Science[edit]

Classification has become a cornerstone of scientific research. One of its most widespread applications is in land cover and land use mapping (LULC). Such analysis provide a foundation for studies across disciplines, such as biodiversity assessments, management of natural resources, and development of human-made structures. In agricultural research, classification techniques are used to differentiate crop types, monitor growth stages, and assess stress caused by drought, pests, or nutrient deficiencies. This information is essential for yield estimation, food security monitoring, and the planning of agricultural policy. In forestry, classification methods support the detection of forest cover changes, illegal logging activities, and degradation of habitats. A well-known example is the monitoring of Amazon deforestation, where classified Landsat and MODIS imagery has been central to annual reports on forest loss. Urban and regional studies also rely heavily on classification. By distinguishing between built-up areas, green spaces, and agricultural land, researchers can monitor and model urban expansion and sprawl, providing valuable insights for sustainable development and planning. Another important application lies in the monitoring of water bodies and wetlands, which are classified to support hydrological modeling, water resource management, and conservation efforts. Seasonal and long-term changes in wetlands or river systems can be identified and quantified, allowing scientists to study flood dynamics, drought conditions, and ecosystem changes. In climate and environmental research, classification is used to generate global datasets of vegetation, snow cover, and desertification. These classifications serve as inputs for climate models and help researchers understand land–atmosphere interactions. Finally, in the field of disaster management, classification provides rapid assessments of affected areas. After and even during events such as wildfires, floods, or hurricanes, classified images can quickly reveal the extent of burned, inundated, or destroyed land, thereby supporting emergency response and recovery planning. All these application examples are just the tip of the iceberg.

Strengths and Challenges[edit]

A major strength of classification lies in its scalability. It enables the analysis of regions, countries, or even the entire globe. Such holistic efforts would be impossible to achieve using field surveys alone. Classification is also highly repeatable in terms of reclassifying based on new information, underlining the machine learning approach behind many classification approaches in GIS. Automated methods allow consistent monitoring across time and space, which is essential for long-term studies. Moreover, classified datasets integrate easily into broader models of climate, hydrology, biodiversity, or land use planning. They also provide clear and interpretable categories that can inform decision-making and policy processes. Regarding challenges, most fundamentally, classifications are abstractions. They are models of reality, not reality itself, and they inevitably simplify and exclude detail even though this is their purpose in the first place. The results of any classification also depend on the subjective analyst’s decisions, which introduces bias. For example, the choice of training data or algorithm strongly shapes the outcome. Resolution adds another layer of complexity: the same landscape may appear very different when classified using coarse or fine spatial data. Context further complicates matters, as certain land cover classes have very similar spectral properties, such as dry vegetation and bare soil, which can easily be confused. Hence, achieving reliable and accurate results can be quite challenging. However, classification results are only as reliable as their validation. Without robust accuracy assessments, classifications risk being misleading or misinterpreted.

Normativity[edit]

Classification is not only a technical exercise but also a normative one. The definition of class schemes is inherently subjective, as categories vary depending on institutional, political, or disciplinary perspectives. For example, the Food and Agriculture Organization defines forest in ways that may differ substantially from national definitions used in forest management or conservation, leading to divergent interpretations of the same landscape. Subjectivity also extends to the choice of methods. The selection of classification algorithms is not purely technical but often shaped by disciplinary traditions, available expertise, or personal preferences. Different communities may engage in ongoing debates over which algorithms are “better” or “worse,” reflecting broader methodological orientations rather than universally valid criteria. In practice, no algorithm is objectively superior across all contexts; each has strengths and weaknesses depending on the purpose of the study, the quality of the data, and the assumptions of the analyst. Ultimately, the interpretation of classification results is shaped by context and intention. A conservation organization may interpret a classification as evidence of alarming forest loss, while a government may emphasize the same data as proof of agricultural growth and development. In this way, classifications are not neutral outputs but rather abstract representations of reality that carry the values, goals, and assumptions of those who produce and use them.

References[edit]

• Campbell, J. B., & Wynne, R. H. (2011). Introduction to Remote Sensing. 5th ed. Guilford Press.
• Eman A. Alshari, Bharti W. Gawali (2021). Development of classification system for LULC using remote sensing and GIS. Global Transitions Proceedings, 2(1), 8-17.
• Lillesand, T., Kiefer, R., & Chipman, J. (2015). Remote Sensing and Image Interpretation. 7th ed. Wiley.
• Congalton, R.G., & Green, K. (2019). Assessing the Accuracy of Remotely Sensed Data: Principles and Practices, Third Edition (3rd ed.). CRC Press.
• Richards, J. A., & Jia, X. (2006). Remote Sensing Digital Image Analysis. 4th ed. Springer.
• Townshend, J. R. G., et al. (1991). “Global land cover classification by remote sensing: Present capabilities and future possibilities.” Remote Sensing of Environment, 35(2–3), 243–255.
• Wulder, M. A., Roy, D. P., Radeloff, V. C., et al. (2022). “Fifty years of Landsat science and impacts.” Remote Sensing of Environment, 280, 113195.
• Wulder, M. A., White, J. C., Goward, S. N., Masek, J. G., Irons, J. R., Herold, M., Cohen, W. B., Loveland, T. R., & Woodcock, C. E. (2008). “Landsat continuity: Issues and opportunities for land cover monitoring.” Remote Sensing of Environment, 112(3), 955–969.

---

The author of this entry is Christoph Schwenck.