Document Type : Research Article
Authors
Soil Science Department, Faculty of Agriculture, University of Tabriz
Abstract
Introduction
High-resolution satellite imagery data is widely utilized for Land Use/Land Cover (LULC) mapping. Analyzing the patterns of LULC and the data derived from changes in land use caters to the increasing societal demands, improving convenience, and fostering a deeper comprehension of the interaction between human activities and environmental factors. Although numerous studies have focused on remote sensing for LULC mapping, there is a pressing need to improve the quality of LULC maps to achieve sustainable land management, especially in light of recent advancements made. This study was carried out in an area covering approximately 8000 hectares, characterized by diverse conditions in LULC, geomorphology and pedology. The objective was to investigate the potential for achieving maximum differentiation and accurate mapping of land features related to LULC. Additionally, the study assessed the impact of various spectral indices on enhancing the results from the classification of Landsat 8 imagery, while also evaluating the efficacy of support vector machine (SVM) and maximum likelihood algorithms in producing maps with satisfactory accuracy and precision.
Materials and Methods
As an initial step, LULC features were identified through fieldwork, and their geographic coordinates were recorded using GPS. These features included various types of LULC, soil surface characteristics, and landform types. Following the fieldwork, 12 types of LULC units were identified. Subsequently, the LULC pattern in the study area was classified using the RGB+NIR+SWIR1 bands of Landsat 8, employing both SVM and maximum likelihood classifiers. To assess the impact of various spectral indices on improving the accuracy of the LULC maps, a set of vegetation indices (NDVI, SAVI, LAI, EVI, and EVI2), bare soil indices (BSI, BSI3, MNDSI, NBLI, DBSI, and MBI), and integrated indices (TLIVI, ATLIVI, and LST), and digital elevation model of study area were successively incorporated into the classification algorithms. Finally, the outcomes from the two classification algorithms were compared, taking into account the influence of the applied indexes. The classification process continued with the selected classifier and indices until reaching the maximum overall accuracy and kappa coefficient.
Results and Discussion
Field observations revealed that the study area could be categorized into 12 primary LULC units, including irrigated farms, flow farming, dry farming, traditional gardens (with no evident order observed among planted trees), modern gardens (featuring regular rows where soil reflectance is visible between tree rows), grasslands, degraded grasslands, highland pastures (covered by Astragalus spp., dominantly), lowland pastures (covered by halophyte plants), salt domes (with no or very poor vegetation), outwash areas (River channel with many waterways), and resistant areas. The results of image classification indicated that the performance of the SVM algorithm across different band combinations is superior to that of the maximum likelihood method. Using SVM resulted in an increase in overall accuracy and Kappa coefficient by 3-8% and 0.03-0.08, respectively. For the map generated using RGB+NIR+SWIR1 bands and employing SVM, overall accuracy and Kappa coefficient were determined to be 76.6% and 0.72, respectively. Among the vegetation indices used in the SVM algorithm, LAI had the most significant impact, increasing the classification accuracy by 2.64%. Among the soil indices, BSI and MBI indices demonstrated the best performance; with BSI increasing the classification accuracy by 1.95% and MBI by 1.64%. Among the integrated indices, LST and ALTIVI enhanced the classification accuracy by 2.75% and 2.35%, respectively. It should be noted that the inclusion of the digital elevation model did not significantly improve the classification accuracy when using the support vector machine algorithm; in fact, it led to a decrease in accuracy when applied to the maximum likelihood classification. The probable reason for this issue is the different nature of DEM data compared to the other input data, as well as the limitations of parametric statistical approaches to effectively integrating data from diverse sources. Finally, the classification process was executed using the three visible bands, NIR, and SWIR1, in conjunction with selected indices (LAI, BSI, MBI, LST, and ALTIVI). Results indicated that using these spectral indices significantly improved classification accuracy, particularly for the DF, DGL, MG, O, and IF land cover/use classes. The calculated accuracies for these classes increased by 11.62%, 18.57%, 20.06%, 29.39%, and 33.19% respectively. Consequently, the accuracy of the classification and the Kappa coefficient (using support vector machine algorithm) increased to 85.24% and 0.82, respectively.
Conclusion
In this research, we aimed to accurately map various land use/land covers by utilizing Landsat 8 imagery and incorporating three group of spectral indexes. Despite spectral interferences and overlaps among various phenomena related to LULC, the utilization of different spectral indices resulted in significant differentiation among LULC classes. Finally, considering the limitations of modelling in ENVI software, it is recommended to investigate the effectiveness of other models for classification in more specialized software, such as R.
Keywords
Main Subjects
)
Send comment about this article