Extended Abstract

Introduction

Land use and land cover change (LULCC) in arid and semi-arid regions has emerged as a critical challenge for sustainable land management, particularly in areas experiencing rapid urbanization and socio-economic transformation. In Iran, accelerated population growth, industrial expansion, and infrastructure development have intensified land conversion processes, often in environmentally sensitive landscapes. Reliable detection of such changes is essential for spatial planning, environmental governance, and risk mitigation. However, conventional LULCC detection approaches, including supervised classification and spectral index–based methods such as the Normalized Difference Vegetation Index (NDVI), suffer from fundamental limitations. These methods typically depend on extensive local training data, are prone to cumulative classification errors, and primarily capture vegetation dynamics rather than structural or functional land-use transformations.

Recent advances in foundation models for Earth observation provide a new paradigm for land change analysis. Pre-trained on massive, globally representative datasets, these models encode high-level semantic and structural information in latent feature spaces, enabling transferable and scalable analyses. Among these, the AlphaEarth foundation model offers dense, pixel-wise latent embeddings that summarize multi-spectral, spatial, and contextual characteristics of the land surface. This study leverages AlphaEarth embeddings to propose a pixel-based, multitemporal similarity framework for analyzing patterns of land use and land cover change without reliance on local training samples. The rapidly urbanizing semi-arid province of Alborz, Iran, is selected as a representative case study for the period 2017–2024..

Methodology

The methodological framework is based on the extraction and analysis of 64-dimensional AlphaEarth latent embeddings at a native spatial resolution of 10 m for annual composites spanning 2017 to 2024. To ensure spatial consistency with auxiliary datasets, all outputs were resampled to 30 m resolution. Change analysis was conducted using a pixel-wise similarity approach rather than categorical classification, thereby reducing the risk of error propagation associated with multi-class labeling.

Four complementary similarity metrics were computed for each pair of temporal embeddings: cosine similarity, Pearson correlation coefficient, dissimilarity (1 − cosine similarity), and the Structural Similarity Index Measure (SSIM). Each metric captures a distinct aspect of change, including directional similarity in latent space, linear association, magnitude of divergence, and spatial–textural variation. This multi-metric strategy improves the interpretability and consistency of latent-space change analysis and supports the identification of both subtle and abrupt changes.

Spatial analyses included hotspot and coldspot identification based on statistically derived similarity thresholds, as well as three-class and five-class stability classifications to represent varying intensities of change. To support decision-making at administrative levels, similarity values were aggregated at the county scale, enabling comparative stability ranking across Alborz Province. Temporal trend analysis was conducted for seven consecutive annual intervals, from 2017–2018 to 2023–2024, using linear regression and the non-parametric Mann–Kendall test to identify long-term tendencies and anomalous periods. Finally, results were conceptually and operationally compared with NDVI-based change detection and with findings from previous studies employing traditional machine learning and deep learning approaches.

Results and discussion

The results reveal pronounced overall stability in land use and land cover across Alborz Province during the study period. Approximately 77.78% of the provincial area, equivalent to 4,030.40 km², exhibited high similarity values, indicating persistent land-use conditions. In contrast, only 2.36%, equivalent to 122.30 km², experienced intense changes, forming spatially concentrated hotspots rather than diffuse patterns. These hotspots were primarily located in the counties of Nazarabad, Eshtehard, Fardis, and parts of Charbagh, closely aligned with industrial corridors, transportation axes, and zones of suburban expansion.

The provincial mean cosine similarity of 0.9569 falls within the upper range of values reported in related change detection studies, suggesting that latent-space similarity analysis provides a scalable framework for representing broad patterns of land-use stability and potential structural transformation. Among the four metrics, cosine similarity and Pearson correlation showed strong spatial and statistical convergence, indicating the consistency of the detected patterns across vector-based similarity measures. SSIM exhibited higher variability, reflecting its sensitivity to local structural and textural changes, particularly in heterogeneous urban environments.

Temporal analysis identified a statistically significant increase in overall stability over the 2017–2024 period, interrupted by a distinct decline in 2020–2021. This decline represents an anomalous interval within the multiyear similarity pattern and indicates a temporary increase in land-use volatility. The subsequent recovery in similarity values during 2021–2022 suggests a return to relative stability, although identifying the specific drivers of this temporal fluctuation requires further investigation using ancillary environmental and socio-economic data.

Comparison with NDVI-based change maps revealed important conceptual differences. NDVI primarily captured vegetation and biomass fluctuations, particularly in mountainous and vegetated northern areas, which were identified as relatively stable by the AlphaEarth-based framework. Conversely, several urban and industrial transformation zones exhibited limited NDVI change but appeared as hotspots in latent-space analysis. This contrast suggests that vegetation-based spectral indices and latent embeddings capture different but complementary dimensions of land-surface change. NDVI is more sensitive to vegetation phenology and biomass variability, whereas AlphaEarth embeddings can provide complementary evidence of patterns consistent with non-vegetative and structural–functional land-use transformations.

Conclusion

This study shows that multitemporal similarity analysis of foundation model embeddings can provide a scalable, training-free, and complementary framework for land use and land cover change analysis in semi-arid regions. By operating directly in latent feature space, the proposed approach reduces dependence on local training data and avoids the error propagation commonly associated with repeated categorical classification. Its main contribution lies in representing patterns of land-use stability and identifying areas potentially consistent with structural and functional land-use transformations beyond vegetation dynamics.

The application to Alborz Province reveals a land-use system characterized by broad stability, interrupted by localized and policy-relevant hotspots of transformation. These findings highlight the potential of foundation models such as AlphaEarth to support continuous land monitoring, early-warning applications, and evidence-based spatial planning, particularly in data-scarce environments. Nevertheless, because latent embeddings are not directly equivalent to ground-truth land-use classes, the results should be interpreted as complementary analytical evidence rather than definitive proof of land-use conversion. Future research should integrate socio-economic drivers, in situ validation, official land-use maps, predictive modeling, and explainable artificial intelligence approaches to further enhance the interpretability, validation, and policy relevance of latent-space change detection.