Extended Abstract

Introduction

Landforms, as fundamental units of geomorphology, result from complex interactions between tectonic, climatic, erosional, and hydrological processes. Their identification and classification are crucial for understanding environmental changes, soil science, land use planning, and natural resource management. Traditional landform mapping methods, based on visual interpretation of topographic maps and expert judgment, are time-consuming, expensive, and lack reproducibility. Consequently, automated and semi-automated methods based on Digital Elevation Models (DEM) have gained prominence. Among various morphometric indices, the Topographic Position Index (TPI) has received special attention due to its conceptual simplicity and high efficiency in distinguishing low features (valleys) from high features (ridges and peaks). However, TPI performance is highly dependent on two key parameters: scale (moving window size) and neighborhood type (neighborhood geometry) . Scale determines whether a landform is identified at a local or regional level, while neighborhood type determines which surrounding cells influence the mean elevation calculation. Despite their importance, many previous Iranian studies have used default window sizes and neighborhood types without optimization. Therefore, the primary objective of this study is to classify landforms in northeastern Eqlid using TPI and determine the optimal neighborhood type (circular, rectangular, annular, and triangular) and optimal scale to produce the most accurate landform map. The study area, covering 51.27 km², is located in Fars Province, between longitudes 52°37' to 52°44' E and latitudes 30°45' to 30°51' N.





Methodology

This study utilized an ASTER DEM with 30-meter spatial resolution (2012). All analyses were performed in ArcGIS software. The Topographic Position Index (TPI) was calculated as \(TPI = Z_0 - \frac{1}{n}\sum_{i=1}^{n} Z_i\), where \(Z_0\) is the elevation of the central point, \(Z_i\) is the elevation of neighboring points, and \(n\) is the number of neighbors. Positive TPI values indicate that the central point is higher than its surroundings (ridges, peaks), negative values indicate lower positions (valleys, streams), and near-zero values indicate flat areas or constant slopes. To determine the optimal scale, square moving windows ranging from 3×3 to 45×45 pixels were evaluated. The optimization criterion was the minimum Root Mean Square Error (RMSE) derived from polynomial distribution fitting. Subsequently, four neighborhood geometries were compared: circular (all cells within a specified radius), annular (less weight to distant cells), triangular (directional pattern), and rectangular (full and symmetric coverage). After calculating TPI and classifying landforms according to the standard Weiss (2001) classification table, the extracted landforms were assigned to geomorphological units: mountain, pediment, and playa. These units were defined based on slope: slopes >25° (mountain), slopes between 1° and 25° (pediment), and slopes <1° (playa).

Results and discussion

The results are significant in three aspects. First, the strong scale dependency of TPI was confirmed. The TPI range at the 45×45 scale (range ≈452 units) was more than double that at the 3×3 scale (range ≈197 units), meaning a single point could be classified as a "local peak" at a small scale but as part of a "mountain slope" at a larger scale. This aligns with theoretical geomorphometry and studies by Weiss (2001) and Wilson & Gallant (2000). Second, and most importantly, the rectangular neighborhood was decisively superior. This superiority stems from its full and symmetric coverage of surrounding cells, enabling more accurate modeling of elevation changes in complex, heterogeneous topography. In contrast, annular and circular neighborhoods tend to smooth out intermediate and small landforms, while the triangular neighborhood's inherent directionality makes it unsuitable for regions with random, multi-directional topography. Third, the challenging finding of mountain peaks within the pediment unitrequires careful interpretation. Although seemingly contradictory, given that geomorphological units were classified solely based on slope with a 25° threshold, it is possible that some weathered, broad peaks with slope angles below 25° (due to long-term denudation or footslope deposition) were classified as pediment. This highlights that slope alone is insufficient for distinguishing large geomorphological units; complementary criteria such as absolute elevation, ruggedness index, or larger-scale TPI should be used. The main limitation of this study is the lack of field validation and dependence on 30-meter DEM resolution, which may not capture small-scale local landforms (<30 m).



Conclusion

This study performed automated landform classification in northeastern Eqlid using TPI, with a focus on optimizing scale and neighborhood type. Results conclusively demonstrate that the rectangular neighborhood significantly outperforms circular, annular, and triangular neighborhoods, being the only type capable of discriminating all nine landforms in the region. Optimal scales for TPI calculation were the 3×3 window (for local landforms) and 45×45 window (for regional, structural landforms). Stream channels occupy the largest area, indicating the dominant role of erosional and drainage processes in shaping the landscape. In geomorphological unit separation, no playa unit was present; mountain peaks dominate the mountain unit, while stream channels dominate the pediment unit. However, the presence of mountain peaks within the pediment unit indicates that slope-based classification alone is insufficient for geomorphological unit separation, and complementary criteria are recommended. The approach presented here—scale optimization using RMSE and systematic comparison of neighborhood types—represents a methodological advancement over previous Iranian studies that used fixed windows and default neighborhoods. Automated TPI-based landform classification, with careful, area-specific selection of scale and neighborhood type, can be an efficient tool for environmental studies, land-use planning, and natural resource management. Nevertheless, field validation and higher-resolution satellite imagery (≤10 m) are strongly recommended for future research