Bilingual
En/Fa
Geographical Research is Published in both Persian and English Full-text.
Volume 38, Issue 1 (2023)
GeoRes 2023, 38(1): 35-43 |
Back to browse issues page
Article Type:
Subject:
History
Received: 2022/12/16 | Accepted: 2023/03/12 | Published: 2023/03/19
Received: 2022/12/16 | Accepted: 2023/03/12 | Published: 2023/03/19
How to cite this article
Varameshc S, Mohtaram Anbaran S. Land Use Mapping by Classification of Landsat Images Using Synthetic Method. GeoRes 2023; 38 (1) :35-43
URL: http://georesearch.ir/article-1-1411-en.html
URL: http://georesearch.ir/article-1-1411-en.html
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Rights and permissions
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Authors
S. Varameshc *1, S. Mohtaram Anbaran1
1- Department of Forest Science and Engineering, Faculty of Agriculture and Natural Resources, University of Mohaghegh Ardabili, Ardabil, Iran
Full-Text (HTML) (82 Views)
Introduction
Land use refers to a combination of soil and living organisms that includes natural cover, agricultural areas, and human-made structures, serving as a crucial interface between human activities and the natural environment. The most significant global factors influencing biodiversity are the availability of suitable water resources and climatic conditions [Mohammady, 2015], which together act as a balancing element among economic, hydrological, and ecological factors [Varamesh, 2017]. Changes in land use and land cover, through interactions between humans and the environment, are closely related to ecosystem services and can affect various functions ranging from biodiversity to climate regulation. Moreover, alterations in land use and cover modify both the functioning and structure of ecosystems. In this regard, Tolessa et al. [2017] have concluded that diverse human activities and climate change negatively affect ecosystem services.
Land use diversity is a consequence of population growth, economic development, and technological advancement, which lead to landscape transformation and fragmentation into natural and man-made patches [Alemayehu et al., 2009; Guo, 2006]. Currently, both societies and governments express increasing concern over the problems arising from the excessive exploitation of land resources [Hasan et al., 2020]. Therefore, land use is a critical subject within the domain of natural resources, and the first step in its management is the preparation of an accurate land use map, for which various classification methods have been proposed.
Remote sensing technology and the use of satellite imagery with a wide range of spatial, temporal, spectral, and radiometric resolutions constitute powerful tools for land use studies. Landsat imagery has been extensively applied in surface studies, land use mapping, and land use change detection [Varamesh et al., 2017]. The advantages of Landsat images include low cost, easy accessibility, moderate spatial and temporal resolution, and a long-term imaging archive [Rozenstein & Karnieli, 2011]. Numerous studies have been conducted on land use mapping using Landsat imagery, and many of these have reported acceptable results (Kappa coefficient ≈ 0.85) [Manandhar et al., 2009; Varamesh et al., 2017; Schulz et al., 2010; Varamesh et al., 2022; Shih et al., 2022; Kafy et al., 2021; Hashim et al., 2022].
There are various approaches for preparing land use maps based on satellite imagery. For small areas, visual interpretation is often employed as the simplest method. However, in larger regions, automated methods such as supervised and unsupervised classifications, spectral analysis, and object-oriented approaches are preferred [Rozenstein & Karnieli, 2011]. Each of these methods has its own advantages and limitations, and the choice of classification approach depends on several factors, including available image-processing algorithms, data accessibility, and the analyst’s familiarity and experience with different techniques. Many studies have shown that the object-oriented method, which relies on image segmentation, generally provides more accurate and reliable results than pixel-based approaches for land use classification [Zhou et al., 2008].
Kolios and Stylios [2013], in their investigation of land use and land cover changes in Greece using Landsat 8 imagery, have compared the Maximum Likelihood, Minimum Distance, Parallelepiped, Mahalanobis Distance, Support Vector Machine (SVM), and Artificial Neural Network (ANN) methods. They have concluded that SVM and ANN yielded higher accuracy than the other techniques.
In heterogeneous regions characterized by variations in vegetation cover, climate, and topography, conventional methods may not achieve sufficient accuracy, and the use of hybrid approaches is often more effective [Lillesand et al., 2015]. In hybrid methods, auxiliary maps such as elevation, road networks, and slope, as well as vegetation and soil indices, are integrated with satellite data to improve accuracy [Mohammady et al., 2014]. Mohammady et al. (2014), in a comparative study of supervised, unsupervised, and hybrid methods for land use mapping in northern Iran, have found that the hybrid method provided higher accuracy, as the integration of remote sensing data, auxiliary data, and decision rules resulted in better classification performance compared to conventional methods.
Knowledge of land use status is valuable for defining precise management units, and evaluating classification techniques can support regional planning and policy-making aimed at optimizing ecosystem management. In another study, Abd El-Kawy et al. [2011] have enhanced land use classification accuracy by combining Maximum Likelihood classification with visual interpretation, achieving overall accuracy and Kappa coefficient values of 95% and 0.92, respectively.
The Fandoghloo region, located on the border between Ardabil and Gilan provinces, exhibits high heterogeneity in terms of climate, topography, and land use types. In such a heterogeneous area, with elevations ranging from 1,346 to 2,018 meters and diverse land covers, conventional land use classification methods do not provide sufficient accuracy. Therefore, hybrid approaches are expected to yield better results. Accordingly, the objective of this study was to prepare a land use map through the integration of supervised, unsupervised, object-oriented, and spectral methods, along with vegetation and soil indices and auxiliary elevation and slope maps.
Methodology
This study employed a descriptive–analytical approach and was conducted in the Fandoghloo region, located on the border of Ardabil and Gilan provinces. The dataset used in this research consisted of a Landsat 8 (OLI) satellite image captured in mid-July 2019, covering the entire study area. Seven spectral bands with a spatial resolution of 30 meters were utilized for the analysis.
During fieldwork, a total of 240 training samples representing different land use types were collected using GPS devices. Approximately one-third of these samples were reserved for validation purposes and were not included in the classification process. In addition, a Digital Elevation Model (DEM) derived from ASTER satellite data with a 30-meter spatial resolution was used to prepare elevation and slope maps. All stages of image processing and map generation were carried out using ENVI, eCognition, and ArcGIS software packages.
Prior to the main analysis, preprocessing of satellite data was performed to enhance image quality. Because similar land covers located on opposite sides of topographic features often showed different spectral values, topographic correction was applied to reduce terrain-induced variations. This step was particularly important due to the diverse relief of the study area. The digital numbers of the image pixels were converted into reflectance values, and atmospheric effects were corrected to obtain more accurate surface reflectance data.
In the next stage, several spectral indices were calculated to improve the classification process. These indices, which are mathematical combinations of image bands, are widely used for distinguishing vegetation cover and land types. Among them, vegetation and soil indices were calculated and incorporated as additional layers into the classification process to enhance the accuracy of results.
Subsequently, the main classification process was carried out. Both supervised and unsupervised methods were tested. The unsupervised ISODATA algorithm, which groups spectrally similar pixels without requiring training samples, did not produce satisfactory separation among classes and was therefore excluded. A supervised classification using the Maximum Likelihood method was then applied, utilizing the field training samples. This method generally provides higher accuracy due to its reliance on prior knowledge of the area.
The Maximum Likelihood classification successfully identified classes such as water bodies, rangelands, fallow lands, and agricultural areas, but it showed considerable overlap among residential areas, forests, orchards, and barren lands. To resolve this issue, additional techniques were employed. Because residential areas and orchards have relatively regular geometric patterns, an object-oriented approach was adopted using eCognition software. This method considers multiple parameters, including shape, texture, and spectral characteristics, allowing for more precise delineation of land use classes.
For forest classification, a threshold-based supervised method was applied. The classification boundaries were determined based on the spectral characteristics of the features. Since forests in the study area are generally found below 2000 meters above sea level, elevation data from the DEM were used to refine the classification by reassigning higher-elevation areas to the rangeland class.
The spectral characteristics of residential and barren land classes were quite similar, making them difficult to distinguish with standard methods. Therefore, a spectral analysis technique known as matched filtering was applied to isolate barren lands more accurately. Furthermore, slope data were used to improve the distinction between agricultural and rangeland areas. Agricultural areas located on steep slopes were reassigned to the rangeland class to ensure greater classification reliability.
In the final stage, all classified layers derived from different techniques were combined to produce the final land use map.
For the accuracy assessment, approximately one-third of the training samples were used as validation data. A confusion matrix was generated to determine the overall accuracy of the classification, and the Kappa coefficient was calculated as a standard measure for evaluating classification performance. According to international standards, a Kappa coefficient value of 0.85 or higher is considered acceptable for land use maps
Findings
The overall accuracy and Kappa coefficient for the land-use map produced using the supervised Maximum Likelihood method were 94.39% and 0.88, respectively. In this classification, water bodies, agricultural lands, fallow areas, and rangelands were well distinguished. The mountainous parts of the region were mainly covered by rangeland and forest, while agricultural, fallow, and residential areas were concentrated in the western parts of the study area. Small water bodies were also observed in the northeastern and southwestern portions of the region.
Accuracy assessment of the map produced by the combined method indicated an overall accuracy of approximately 97% and a Kappa coefficient close to 0.96, suggesting a high level of classification reliability.
Among the classified categories, the highest omission error (representing pixels belonging to a specific class that were incorrectly assigned to other classes) was found in the rangeland class, while the lowest omission error occurred in the forest class. The highest commission error (pixels incorrectly assigned to a given class that do not truly belong to it) was associated with the fallow land class, whereas the lowest commission error was related to orchards.
The producer’s accuracy, which reflects the spectral separability of each class, was highest for forests, indicating a strong ability to discriminate this land cover type from others. The highest user’s accuracy was recorded for orchards, with all pixels correctly classified in this category, while the lowest value was observed for fallow lands. Overall, the orchard class exhibited the highest classification accuracy among all categories.
Analysis of the areal extent of each land-use class showed that rangelands covered the largest portion of the study area, whereas water bodies accounted for the smallest area
Discussion
Landsat 8 imagery was used to prepare the land-use map, as Landsat data with suitable spatial resolution and temporal frequency are a key source for land-use mapping and have been widely utilized in previous research. Despite using a considerable number of training samples in the Maximum Likelihood method, some confusion remained between certain land-use classes, and the classification accuracy was not fully satisfactory for all categories. Considering the topographic, climatic, and land-use diversity of the study area, such results were expected.
The extraction of the water body class was straightforward using the supervised method, given its distinct spectral characteristics and strong separability from other land-use types. Rangeland, fallow, and agricultural classes were also well discriminated, although some overlap occurred in steep areas. This was corrected based on expert knowledge indicating that cultivation rarely occurs on slopes greater than 20%. It should be noted, however, that the 20% slope threshold is not universal and can vary depending on local land-use practices.
The forest class was accurately extracted using the supervised density-slicing approach. Nevertheless, because of spectral overlap among land-cover types and the limited range of defined classes, this method cannot be regarded as a stand-alone or comprehensive classification technique. It is most suitable for distinguishing a limited number of classes or for assisting in the identification of specific features.
Due to the high spectral similarity between residential areas and bare lands, as well as between orchards and forests, these classes were extracted using the object-based method. In object-based classification, the attributes and quality of features used for classification depend directly on the objects and their segmentation quality. This approach has proven effective in various applications in earth and life sciences. It is also an iterative process, meaning that classification is repeated several times until the highest membership value for each class is achieved.
For barren lands, an adaptive filtering technique was applied, where each pixel value in the output image corresponded to the proportion of target material within that pixel. Pixels with values below zero were treated as background. This approach allows for spectral unmixing and sub-pixel estimation, which has become increasingly important for improving classification precision.
In the forest class, corrections were applied for elevations above 2000 meters, where forest cover is absent, and these areas were reclassified as rangelands. This adjustment was made based on prior knowledge of the local ecosystem. The use of auxiliary geographic data, such as slope and elevation maps, to improve classification accuracy has also been validated by other researchers.
The final map achieved an overall accuracy of about 97% and a Kappa coefficient of approximately 0.96, which are both significantly higher than the minimum acceptable values proposed by the U.S. Geological Survey. These high accuracy levels confirm the effectiveness of integrating multiple classification techniques and incorporating ancillary data such as slope and elevation, supported by expert knowledge of the study area.
Compared with previous studies, the accuracy obtained in this research was considerably higher. Earlier investigations have reported a wide range of accuracy values for land-use classifications, depending on the number and type of classes used, both of which strongly influence the resulting accuracy. Studies that employed combined or hybrid approaches and auxiliary data have generally achieved improved accuracy compared with conventional classification methods.
For instance, researchers have demonstrated that the use of auxiliary factors and hybrid techniques can substantially enhance classification results. Other studies have confirmed that combining visual interpretation with conventional supervised methods significantly improves overall accuracy and the Kappa coefficient. Similarly, the integration of multi-seasonal images and topographic maps has been shown to further increase classification performance.
It is important to emphasize that the selection of classification methods and ancillary data should be tailored to the specific conditions of each study area. For example, in this research, forests were not present above 2000 meters, and agricultural activities were rare on slopes exceeding 20%. Such criteria may not be applicable elsewhere; therefore, local knowledge is essential for effectively utilizing auxiliary information.
In regions with similar conditions, such as the northern parts of the country, the approach used in this study can also be effectively applied. However, in large or heterogeneous areas where multiple images are required or where classes exhibit high spectral similarity (e.g., residential and bare lands, agriculture and rangeland), conventional pixel-based algorithms, such as Maximum Likelihood or IsoData tend to yield lower accuracy.
Furthermore, because natural phenomena are often nonlinear in nature, the inherently linear assumptions of traditional methods may introduce substantial errors. Therefore, advanced classification techniques such as Support Vector Machines, Artificial Neural Networks, and object-based approaches should be considered. Depending on the analyst’s expertise and familiarity with the study area, other hybrid or knowledge-based techniques can also be implemented. It should finally be noted that different classification methods may yield varying results across regions, highlighting the need for careful method selection and contextual adaptation.
Conclusion
The combined classification method achieved higher accuracy compared with conventional approaches, with an overall accuracy of approximately 97% and a Kappa coefficient close to 0.96. This level of precision demonstrates the suitability of the hybrid approach for effective land-use planning and management. Owing to the spectral, shape, and color similarities among certain land-use classes, the integration of multiple classification techniques and the use of auxiliary maps significantly improve the reliability of results compared to traditional single-method approaches. Therefore, depending on the specific characteristics of each region—such as its degree of homogeneity and the type of satellite imagery employed—different combinations of methods and indices can be applied to enhance the accuracy of land-use classification
Acknowledgments: None declared.
Ethical Approval: Not applicable.
Conflict of Interest: The authors declare no conflict of interest.
Authors’ Contributions: Varameshc S (first author), Main Researcher/Introduction Writer/Methodologist/Discussion Writer (60%); Mohtaram Anbaran S (second author), Assistant Researcher/Discussion Writer/Statistical Analyst (40%).
Funding: No financial support was received for this study.
Land use refers to a combination of soil and living organisms that includes natural cover, agricultural areas, and human-made structures, serving as a crucial interface between human activities and the natural environment. The most significant global factors influencing biodiversity are the availability of suitable water resources and climatic conditions [Mohammady, 2015], which together act as a balancing element among economic, hydrological, and ecological factors [Varamesh, 2017]. Changes in land use and land cover, through interactions between humans and the environment, are closely related to ecosystem services and can affect various functions ranging from biodiversity to climate regulation. Moreover, alterations in land use and cover modify both the functioning and structure of ecosystems. In this regard, Tolessa et al. [2017] have concluded that diverse human activities and climate change negatively affect ecosystem services.
Land use diversity is a consequence of population growth, economic development, and technological advancement, which lead to landscape transformation and fragmentation into natural and man-made patches [Alemayehu et al., 2009; Guo, 2006]. Currently, both societies and governments express increasing concern over the problems arising from the excessive exploitation of land resources [Hasan et al., 2020]. Therefore, land use is a critical subject within the domain of natural resources, and the first step in its management is the preparation of an accurate land use map, for which various classification methods have been proposed.
Remote sensing technology and the use of satellite imagery with a wide range of spatial, temporal, spectral, and radiometric resolutions constitute powerful tools for land use studies. Landsat imagery has been extensively applied in surface studies, land use mapping, and land use change detection [Varamesh et al., 2017]. The advantages of Landsat images include low cost, easy accessibility, moderate spatial and temporal resolution, and a long-term imaging archive [Rozenstein & Karnieli, 2011]. Numerous studies have been conducted on land use mapping using Landsat imagery, and many of these have reported acceptable results (Kappa coefficient ≈ 0.85) [Manandhar et al., 2009; Varamesh et al., 2017; Schulz et al., 2010; Varamesh et al., 2022; Shih et al., 2022; Kafy et al., 2021; Hashim et al., 2022].
There are various approaches for preparing land use maps based on satellite imagery. For small areas, visual interpretation is often employed as the simplest method. However, in larger regions, automated methods such as supervised and unsupervised classifications, spectral analysis, and object-oriented approaches are preferred [Rozenstein & Karnieli, 2011]. Each of these methods has its own advantages and limitations, and the choice of classification approach depends on several factors, including available image-processing algorithms, data accessibility, and the analyst’s familiarity and experience with different techniques. Many studies have shown that the object-oriented method, which relies on image segmentation, generally provides more accurate and reliable results than pixel-based approaches for land use classification [Zhou et al., 2008].
Kolios and Stylios [2013], in their investigation of land use and land cover changes in Greece using Landsat 8 imagery, have compared the Maximum Likelihood, Minimum Distance, Parallelepiped, Mahalanobis Distance, Support Vector Machine (SVM), and Artificial Neural Network (ANN) methods. They have concluded that SVM and ANN yielded higher accuracy than the other techniques.
In heterogeneous regions characterized by variations in vegetation cover, climate, and topography, conventional methods may not achieve sufficient accuracy, and the use of hybrid approaches is often more effective [Lillesand et al., 2015]. In hybrid methods, auxiliary maps such as elevation, road networks, and slope, as well as vegetation and soil indices, are integrated with satellite data to improve accuracy [Mohammady et al., 2014]. Mohammady et al. (2014), in a comparative study of supervised, unsupervised, and hybrid methods for land use mapping in northern Iran, have found that the hybrid method provided higher accuracy, as the integration of remote sensing data, auxiliary data, and decision rules resulted in better classification performance compared to conventional methods.
Knowledge of land use status is valuable for defining precise management units, and evaluating classification techniques can support regional planning and policy-making aimed at optimizing ecosystem management. In another study, Abd El-Kawy et al. [2011] have enhanced land use classification accuracy by combining Maximum Likelihood classification with visual interpretation, achieving overall accuracy and Kappa coefficient values of 95% and 0.92, respectively.
The Fandoghloo region, located on the border between Ardabil and Gilan provinces, exhibits high heterogeneity in terms of climate, topography, and land use types. In such a heterogeneous area, with elevations ranging from 1,346 to 2,018 meters and diverse land covers, conventional land use classification methods do not provide sufficient accuracy. Therefore, hybrid approaches are expected to yield better results. Accordingly, the objective of this study was to prepare a land use map through the integration of supervised, unsupervised, object-oriented, and spectral methods, along with vegetation and soil indices and auxiliary elevation and slope maps.
Methodology
This study employed a descriptive–analytical approach and was conducted in the Fandoghloo region, located on the border of Ardabil and Gilan provinces. The dataset used in this research consisted of a Landsat 8 (OLI) satellite image captured in mid-July 2019, covering the entire study area. Seven spectral bands with a spatial resolution of 30 meters were utilized for the analysis.
During fieldwork, a total of 240 training samples representing different land use types were collected using GPS devices. Approximately one-third of these samples were reserved for validation purposes and were not included in the classification process. In addition, a Digital Elevation Model (DEM) derived from ASTER satellite data with a 30-meter spatial resolution was used to prepare elevation and slope maps. All stages of image processing and map generation were carried out using ENVI, eCognition, and ArcGIS software packages.
Prior to the main analysis, preprocessing of satellite data was performed to enhance image quality. Because similar land covers located on opposite sides of topographic features often showed different spectral values, topographic correction was applied to reduce terrain-induced variations. This step was particularly important due to the diverse relief of the study area. The digital numbers of the image pixels were converted into reflectance values, and atmospheric effects were corrected to obtain more accurate surface reflectance data.
In the next stage, several spectral indices were calculated to improve the classification process. These indices, which are mathematical combinations of image bands, are widely used for distinguishing vegetation cover and land types. Among them, vegetation and soil indices were calculated and incorporated as additional layers into the classification process to enhance the accuracy of results.
Subsequently, the main classification process was carried out. Both supervised and unsupervised methods were tested. The unsupervised ISODATA algorithm, which groups spectrally similar pixels without requiring training samples, did not produce satisfactory separation among classes and was therefore excluded. A supervised classification using the Maximum Likelihood method was then applied, utilizing the field training samples. This method generally provides higher accuracy due to its reliance on prior knowledge of the area.
The Maximum Likelihood classification successfully identified classes such as water bodies, rangelands, fallow lands, and agricultural areas, but it showed considerable overlap among residential areas, forests, orchards, and barren lands. To resolve this issue, additional techniques were employed. Because residential areas and orchards have relatively regular geometric patterns, an object-oriented approach was adopted using eCognition software. This method considers multiple parameters, including shape, texture, and spectral characteristics, allowing for more precise delineation of land use classes.
For forest classification, a threshold-based supervised method was applied. The classification boundaries were determined based on the spectral characteristics of the features. Since forests in the study area are generally found below 2000 meters above sea level, elevation data from the DEM were used to refine the classification by reassigning higher-elevation areas to the rangeland class.
The spectral characteristics of residential and barren land classes were quite similar, making them difficult to distinguish with standard methods. Therefore, a spectral analysis technique known as matched filtering was applied to isolate barren lands more accurately. Furthermore, slope data were used to improve the distinction between agricultural and rangeland areas. Agricultural areas located on steep slopes were reassigned to the rangeland class to ensure greater classification reliability.
In the final stage, all classified layers derived from different techniques were combined to produce the final land use map.
For the accuracy assessment, approximately one-third of the training samples were used as validation data. A confusion matrix was generated to determine the overall accuracy of the classification, and the Kappa coefficient was calculated as a standard measure for evaluating classification performance. According to international standards, a Kappa coefficient value of 0.85 or higher is considered acceptable for land use maps
Findings
The overall accuracy and Kappa coefficient for the land-use map produced using the supervised Maximum Likelihood method were 94.39% and 0.88, respectively. In this classification, water bodies, agricultural lands, fallow areas, and rangelands were well distinguished. The mountainous parts of the region were mainly covered by rangeland and forest, while agricultural, fallow, and residential areas were concentrated in the western parts of the study area. Small water bodies were also observed in the northeastern and southwestern portions of the region.
Accuracy assessment of the map produced by the combined method indicated an overall accuracy of approximately 97% and a Kappa coefficient close to 0.96, suggesting a high level of classification reliability.
Among the classified categories, the highest omission error (representing pixels belonging to a specific class that were incorrectly assigned to other classes) was found in the rangeland class, while the lowest omission error occurred in the forest class. The highest commission error (pixels incorrectly assigned to a given class that do not truly belong to it) was associated with the fallow land class, whereas the lowest commission error was related to orchards.
The producer’s accuracy, which reflects the spectral separability of each class, was highest for forests, indicating a strong ability to discriminate this land cover type from others. The highest user’s accuracy was recorded for orchards, with all pixels correctly classified in this category, while the lowest value was observed for fallow lands. Overall, the orchard class exhibited the highest classification accuracy among all categories.
Analysis of the areal extent of each land-use class showed that rangelands covered the largest portion of the study area, whereas water bodies accounted for the smallest area
Discussion
Landsat 8 imagery was used to prepare the land-use map, as Landsat data with suitable spatial resolution and temporal frequency are a key source for land-use mapping and have been widely utilized in previous research. Despite using a considerable number of training samples in the Maximum Likelihood method, some confusion remained between certain land-use classes, and the classification accuracy was not fully satisfactory for all categories. Considering the topographic, climatic, and land-use diversity of the study area, such results were expected.
The extraction of the water body class was straightforward using the supervised method, given its distinct spectral characteristics and strong separability from other land-use types. Rangeland, fallow, and agricultural classes were also well discriminated, although some overlap occurred in steep areas. This was corrected based on expert knowledge indicating that cultivation rarely occurs on slopes greater than 20%. It should be noted, however, that the 20% slope threshold is not universal and can vary depending on local land-use practices.
The forest class was accurately extracted using the supervised density-slicing approach. Nevertheless, because of spectral overlap among land-cover types and the limited range of defined classes, this method cannot be regarded as a stand-alone or comprehensive classification technique. It is most suitable for distinguishing a limited number of classes or for assisting in the identification of specific features.
Due to the high spectral similarity between residential areas and bare lands, as well as between orchards and forests, these classes were extracted using the object-based method. In object-based classification, the attributes and quality of features used for classification depend directly on the objects and their segmentation quality. This approach has proven effective in various applications in earth and life sciences. It is also an iterative process, meaning that classification is repeated several times until the highest membership value for each class is achieved.
For barren lands, an adaptive filtering technique was applied, where each pixel value in the output image corresponded to the proportion of target material within that pixel. Pixels with values below zero were treated as background. This approach allows for spectral unmixing and sub-pixel estimation, which has become increasingly important for improving classification precision.
In the forest class, corrections were applied for elevations above 2000 meters, where forest cover is absent, and these areas were reclassified as rangelands. This adjustment was made based on prior knowledge of the local ecosystem. The use of auxiliary geographic data, such as slope and elevation maps, to improve classification accuracy has also been validated by other researchers.
The final map achieved an overall accuracy of about 97% and a Kappa coefficient of approximately 0.96, which are both significantly higher than the minimum acceptable values proposed by the U.S. Geological Survey. These high accuracy levels confirm the effectiveness of integrating multiple classification techniques and incorporating ancillary data such as slope and elevation, supported by expert knowledge of the study area.
Compared with previous studies, the accuracy obtained in this research was considerably higher. Earlier investigations have reported a wide range of accuracy values for land-use classifications, depending on the number and type of classes used, both of which strongly influence the resulting accuracy. Studies that employed combined or hybrid approaches and auxiliary data have generally achieved improved accuracy compared with conventional classification methods.
For instance, researchers have demonstrated that the use of auxiliary factors and hybrid techniques can substantially enhance classification results. Other studies have confirmed that combining visual interpretation with conventional supervised methods significantly improves overall accuracy and the Kappa coefficient. Similarly, the integration of multi-seasonal images and topographic maps has been shown to further increase classification performance.
It is important to emphasize that the selection of classification methods and ancillary data should be tailored to the specific conditions of each study area. For example, in this research, forests were not present above 2000 meters, and agricultural activities were rare on slopes exceeding 20%. Such criteria may not be applicable elsewhere; therefore, local knowledge is essential for effectively utilizing auxiliary information.
In regions with similar conditions, such as the northern parts of the country, the approach used in this study can also be effectively applied. However, in large or heterogeneous areas where multiple images are required or where classes exhibit high spectral similarity (e.g., residential and bare lands, agriculture and rangeland), conventional pixel-based algorithms, such as Maximum Likelihood or IsoData tend to yield lower accuracy.
Furthermore, because natural phenomena are often nonlinear in nature, the inherently linear assumptions of traditional methods may introduce substantial errors. Therefore, advanced classification techniques such as Support Vector Machines, Artificial Neural Networks, and object-based approaches should be considered. Depending on the analyst’s expertise and familiarity with the study area, other hybrid or knowledge-based techniques can also be implemented. It should finally be noted that different classification methods may yield varying results across regions, highlighting the need for careful method selection and contextual adaptation.
Conclusion
The combined classification method achieved higher accuracy compared with conventional approaches, with an overall accuracy of approximately 97% and a Kappa coefficient close to 0.96. This level of precision demonstrates the suitability of the hybrid approach for effective land-use planning and management. Owing to the spectral, shape, and color similarities among certain land-use classes, the integration of multiple classification techniques and the use of auxiliary maps significantly improve the reliability of results compared to traditional single-method approaches. Therefore, depending on the specific characteristics of each region—such as its degree of homogeneity and the type of satellite imagery employed—different combinations of methods and indices can be applied to enhance the accuracy of land-use classification
Acknowledgments: None declared.
Ethical Approval: Not applicable.
Conflict of Interest: The authors declare no conflict of interest.
Authors’ Contributions: Varameshc S (first author), Main Researcher/Introduction Writer/Methodologist/Discussion Writer (60%); Mohtaram Anbaran S (second author), Assistant Researcher/Discussion Writer/Statistical Analyst (40%).
Funding: No financial support was received for this study.
References
1. Abd El-Kawy OR, Rød JK, Ismail HA, Suliman AS (2011). Land use and land cover change detection in the western Nile delta of Egypt using remote sensing data. Applied Geography. 31(2):483-494. [Link] [DOI:10.1016/j.apgeog.2010.10.012]
2. Alemayehu F, Nurhussen T, Jan N, Atkilt G, Amanuel Z, Mintesinot B, et al (2009). The impacts of watershed management on land use and land cover dynamics in Eastern Tigray (Ethiopia). Resources, Conservation and Recycling. 53(4):192-198. [Link] [DOI:10.1016/j.resconrec.2008.11.007]
3. Anderson JR, Hardy EE, Roach JT, Witmer RE (1976). A land use and land cover classification system for use with remote sensor data. Washington DC: US Government Printing Office. [Link] [DOI:10.3133/pp964]
4. Baatz M, Benz U, Dehghani S, Heynen M, Höltje A, Hofmann P, Lingenfelder I, et al (2004). eCognition professional user guide 4. Munich: Definiens Imaging. [Link]
5. Bakr N, Weindorf DC, Bahnassy MH, Marei SM, El-Badawi MM (2010). Monitoring land cover changes in a newly reclaimed area of Egypt using multi-temporal Landsat data. Applied Geography. 30(4):592-605. [Link] [DOI:10.1016/j.apgeog.2009.10.008]
6. Brandt JS, Haynes MA, Kuemmerle T, Waller DM, Radeloff VC (2013). Regime shift on the roof of the world: Alpine meadows converting to shrublands in the southern Himalayas. Biological Conservation. 158:116-127. [Link] [DOI:10.1016/j.biocon.2012.07.026]
7. Fatahi MM, Nowrozi, AA, Abkar AA, Khalkhali A (2007). Comparison of methods of classification and preparation of land use map (Landuse) of arid regions using satellite images. Pajouhesh va Sazandgi. 20(30):129-135. [Persian] [Link]
8. Guo L (2006). Analysis of spatio-temporal changes in the landscape pattern of the Taishan mountain. Journal of Mountain Ecology. 8:1-6. [Link]
9. Guo Y, Wu Y, Ju Z, Wang J, Zhao L (2010). Remote sensing image classification by the Chaos genetic algorithm in monitoring land use changes. Mathematical and Computer Modelling. 51(11-12):1408-1416. [Link] [DOI:10.1016/j.mcm.2009.10.023]
10. Hantson S, Chuvieco E (2011). Evaluation of different topographic correction methods for landsat imagery. International Journal of Applied Earth Observation and Geoinformation. 13(5):691-700. [Link] [DOI:10.1016/j.jag.2011.05.001]
11. Hasan SS, Lin Z, Giashuddin M, Tofayel A, Abdus S (2020). Impact of land use change on ecosystem services: A review. Environmental Development. 34:100527. [Link] [DOI:10.1016/j.envdev.2020.100527]
12. Hashim BM, AL Maliki A, Sultan M, Shahid S (2022). Effect of land use land cover changes on land surface temperature during 1984-2020: A case study of Baghdad city using landsat image. Natural Hazards. 112(2):1223-1246. [Link] [DOI:10.1007/s11069-022-05224-y]
13. Kafy AA, Al Rakib A, Akter KS, Jahir DA, Sikdar S, Al Faisal A, et al (2021). Assessing and predicting land use/land cover, land surface temperature and urban thermal field variance index using landsat imagery for Dhaka metropolitan area. Environmental Challenges. 4:100192. [Link] [DOI:10.1016/j.envc.2021.100192]
14. Kolios S, Stylios CD (2013). Identification of land cover/land use changes in the greater area of the Preveza peninsula in Greece using Landsat satellite data. Applied Geography. 40:150-160. [Link] [DOI:10.1016/j.apgeog.2013.02.005]
15. Lillesand T, Kiefer RW, Chipman J (2015) Remote sensing and image interpretation. Toronto: John Wiley and Sons. [Link]
16. Manandhar R, Odeh IOA, Ancev T (2009). Improving the accuracy of land use and land cover classification of landsat data using post-classification enhancement. Remote Sensing. 1(3):330-344. [Link] [DOI:10.3390/rs1030330]
17. Mohammady M, Zeinivand H, Moradi HR, Pourghasemi HR, Farazjoo H (2015) Investigating the effects of land use on runoff generation using WetSpa model. Iranian journal of Ecohydrology. 2(4):357-369. [Persian] [Link]
18. Mohammady M, Moradi HR, Zeinivand H, Temme AJAM (2014). A comparison of supervised, unsupervised and synthetic land use classification methods in the north of Iran. International Journal of Environmental Science and Technology. 12(5):1515-1526. [Link] [DOI:10.1007/s13762-014-0728-3]
19. Pakhale GK, Gupta PK (2010). Comparison of advanced pixel based (ANN and SVM) and object-oriented classification approaches using landsat-7 Etm+ data. International Journal of Engineering and Technology. 2(4):245-251. [Link]
20. Rezaei Moghadam MH, Rezaei Banafsheh M, Faizizadeh B, Nazmfar H (2010). Classification of land cover/land use based on object-oriented technique and satellite images, case study: West Azerbaijan province. Pajouhesh va Sazandgi. 23(2):19-32. [Persian] [Link]
21. Rojas C, Pino J, Basnou C, Vivanco M (2013). Assessing land-use and-cover changes in relation to geographic factors and urban planning in the metropolitan area of Concepción (Chile),implications for biodiversity conservation. Applied Geography. 39:93-103. [Link] [DOI:10.1016/j.apgeog.2012.12.007]
22. Rosenfield GH, Fitzpatrick-Lins K (1986). A coefficient of agreement as a measure of thematic classification accuracy. Photogrammetric engineering and remote sensing. 52(2):223-227. [Link]
23. Rozenstein O, Karnieli A (2011). Comparison of methods for land-use classification incorporating remote sensing and GIS inputs. Applied Geography. 31(2):533-544. [Link] [DOI:10.1016/j.apgeog.2010.11.006]
24. Schulz JJ, Cayuela L, Echeverria C, Salas J, Benayas JMR (2010). Monitoring land cover change of the dryland forest landscape of central Chile (1975-2008). Applied Geography. 30(3):436-447. [Link] [DOI:10.1016/j.apgeog.2009.12.003]
25. Sexton JO, Urban DL, Donohue MJ, Song C (2013). Long-term land cover dynamics by multi-temporal classification across the landsat-5 record. Remote Sensing of Environment. 128:246-258. [Link] [DOI:10.1016/j.rse.2012.10.010]
26. Shih H, Stow DA, Chang KC, Roberts DA, Goulias KG (2022). From land cover to land use: Applying random forest classifier to Landsat imagery for urban land-use change mapping. Geocarto International.37(19):5523-5546. [Link] [DOI:10.1080/10106049.2021.1923827]
27. Stefanov WL, Ramsey MS, Christensen PR (2001). Monitoring urban land cover change: An expert system approach to land cover classification of semiarid to arid urban centers. Remote sensing of Environment. 77(2):173-185. [Link] [DOI:10.1016/S0034-4257(01)00204-8]
28. Sun J, Yang J, Zhang C, Yun W, Qu J (2013). Automatic remotely sensed image classification in a grid environment based on the maximum likelihood method. Mathematical and Computer Modelling. 58(3-4):573-581. [Link] [DOI:10.1016/j.mcm.2011.10.063]
29. Ekalagan JR (2007) Land use: The interaction of economy, ecology and hydrology. In: Tabibian M translator. Tehran: Tehran University Publications. [Persian] [Link]
30. Tokola T, Sarkeala J, Van der Linden M (2001). Use of topographic correction in landsat TM-based forest interpretation in Nepal. International Journal of Remote Sensing. 22(4):551-563. [Link] [DOI:10.1080/01431160050505856]
31. Tolessa T, Senbeta F, Kidane M (2017). The impact of land use/land cover change on ecosystem services in the central highlands of Ethiopia. Ecosystem services. 23:47-54. [Link] [DOI:10.1016/j.ecoser.2016.11.010]
32. USGS [Internet] Landsat 8 Using Product (2015). Virginia: United States Geological Survey [Cited 2022, November 20]. Available from: https://www.usgs.gov/search?keywords=Landsat8+ [Link]
33. Varamesh S, Mohtaram Anbaran S, Rouhnavaz Z (2022). Evaluation and monitoring of the thirty-year physical expansion process of Ardabil city using satellite images. Research Quarterly of Geographical Data (SEPEHR). 31(123):139-153. [Persian] [Link]
34. Varamesh S, Hosseini.SM, Rahimzadegan M (2017). Comparison of conventional and advanced classification approaches by landsat-8 imagery. Applied Ecology and Environmental Research. 15(3):1407-1416. [Link] [DOI:10.15666/aeer/1503_14071416]
35. Varamesh S, Hosseini SM, Rahimzadegan M (2017). Detection of land use changes in northeastern Iran by landsat satellite data. Applied Ecology and Environmental Research. 15(3):1443-1454. [Link] [DOI:10.15666/aeer/1503_14431454]
36. Zhou W, Troy A, Grove M (2008). Object-based land cover classification and change analysis in the Baltimore metropolitan area using multitemporal high resolution remote sensing data. Sensors. 8(3):1613-1636. [Link] [DOI:10.3390/s8031613]