Document Type : Original Research Paper
Author
Department of Geomatics Engineering, University of Zanjan, Zanjan, Iran.
Abstract
Background and Objectives: Coastal–mountain landscapes are shaped by the simultaneous pressures of settlement growth, agriculture, tourism, accessibility, ecological protection, and pronounced topographic constraints. Under these conditions, net change alone provides an incomplete account of land-use/land-cover dynamics: it cannot distinguish persistence from gross exchange, identify the source classes contributing to new built-up land, describe the spatial configuration of expansion, or determine whether statistical associations remain informative when observations are separated in space. This study quantified the amount, direction, intensity, spatial configuration, clustering, and conditional spatial associations of mapped land-use/land-cover transitions in Ramsar County, northern Iran, between 2016 and 2025. The analysis addressed five related questions within a single auditable framework: how much mapped change occurred; which classes experienced the greatest gains, losses, and exchanges and at what intensities; whether new built-up patches occurred as infill, edge-expansion, or outlying growth; where change formed statistically significant spatial clusters; and how effectively linear and nonlinear models discriminated four dominant mapped transitions.
Methods: Two co-registered land use/land cover rasters for 2016 and 2025, with a spatial resolution of 10 m, were analyzed over a common valid extent of 72,603.52 ha. Both maps comprised seven classes: water, built-up, cropland, orchard, forest, rangeland, and bare land. Cell-by-cell comparison of the two maps was used to construct a transition matrix and calculate persistence, gross gains, gross losses, net change, exchanges, and total change. Intensity analysis at the interval, category, and transition levels compared observed rates with a uniform annual reference rate. Newly built-up cells were converted into patches using eight-neighbor connectivity, and the Landscape Expansion Index with an original 10-m buffer was used to classify patches into edge-expansion, infilling, and outlying growth. Change rates were also calculated across gradients of elevation, slope, and Euclidean distance from the coast, roads, baseline built-up areas, and rivers. To examine spatial concentration, total change density was aggregated within an approximately 500-m grid, and the local Getis–Ord Gi* statistic was calculated to identify hot and cold spots at the 95% and 99% significance levels. Four transitions—rangeland to bare land, orchard to forest, forest to orchard, and orchard to rangeland—were selected for modeling based on mapped area and a minimum number of occurrences. For each transition, logistic regression with standardized variables and heteroskedasticity-consistent errors was used to estimate conditional odds ratios, and model transferability was assessed using five spatial folds based on 2-km blocks. Extreme Gradient Boosting was employed as a nonlinear complement, and SHAP values were used to summarize the contribution of variables to model predictions. Finally, accounting checks were performed to verify the common mask, class coverage, transition totals, model sample sizes, and consistency of tabulated outputs.
Findings: Of the common valid area of 72,603.52 ha, 53,450.72 ha remained on the diagonal of the transition matrix, while 19,152.80 ha underwent a class transition; thus, 73.62% of the landscape was stable in terms of mapped labels, whereas 26.38% experienced mapped change. The uniform annual change intensity was 2.931%. Bare land showed the largest net gain, with an increase of 6,935.97 ha, while the rangeland-to-bare-land mapped transition, covering 6,707.51 ha, represented the largest off-diagonal flow associated with this pattern. In contrast, rangeland recorded the largest net loss, amounting to 5,544.90 ha. Orchard and forest exhibited substantial two-way exchanges, with their gross flows considerably exceeding their net changes. The built-up class gained 2,299.71 ha and lost 907.36 ha, resulting in a net increase of 1,392.35 ha. Among the 29,968 newly built-up patches, edge expansion accounted for 67.19% of the added area, whereas outlying patches represented 31.62% of the total number of patches but only 14.93% of the added area. The change rate decreased continuously with increasing distance from baseline built-up areas, declining from 37.53% at 0–100 m to 8.05% at 2,000–5,000 m. Along the elevation gradient, the highest change rate, 46.13%, was observed at 100–300 m. Of the 3,067 analytical grid cells, 18.75% were classified as hot spots and 20.22% as cold spots, and the correlation between change density and the standardized Gi* score was 0.838. The area under the receiver operating characteristic curve (AUC) for logistic regression under block cross-validation ranged from 0.722 to 0.774. For Extreme Gradient Boosting, AUC values obtained from a random holdout ranged from 0.826 to 0.893; however, because the evaluation designs were not equivalent, this difference cannot be attributed solely to the superiority of one algorithm over the other. Elevation had the highest mean absolute SHAP contribution across all four models, although the direction of its conditional odds ratio in logistic regression was not consistent across transitions.
Conclusions: The mapped land use/land cover change pattern in Ramsar cannot be explained solely by net changes, because a substantial proportion of the observed dynamics was associated with gross exchanges among classes. The major patterns included the mapped rangeland-to-bare-land transition, reciprocal flows between orchard and forest, net built-up expansion dominated in area by edge growth, a considerable number of outlying patches, and multi-centered clustering of change. The observed relationships with topography and accessibility also varied among transitions and were transition-specific. The accounting figures represent direct descriptions of differences between the paired maps, whereas the relationships derived from the models represent conditional, non-causal statistical associations. Furthermore, the absence of an independent accuracy assessment for the final 2025 raster means that the numerical accuracy of the reported areas should not be interpreted as equivalent to thematic certainty about the mapped changes at the parcel scale.
Keywords
- Land-use Change
- Intensity Analysis
- Urban Expansion
- Spatial Clustering
- Logistic Regression
- Explainable Machine Learning
- Ramsar
Main Subjects
COPYRIGHTS
© 2026 The Author(s). This is an open-access article distributed under the terms and conditions of the Creative Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
)