Multi-Modal Unsupervised Change Detection of Urban Vegetation in Birmingham, UK: A Cross-Method Comparison under 2022 Drought Conditions

Naya Desai, Emma J.S. Ferranti, Sarah V. Greenham, Joshua D. Vande Hey, Andrew D. Quinn, Stef De Sabbata

As the climate changes, cities are increasingly exposed to extreme weather events such as droughts, further amplified by the urban heat island effect. Urban vegetation, a key Nature-based Solution for cooling and climate resilience, is vulnerable to water stress. Therefore, it is increasingly important to understand how urban vegetation responds during drought years, particularly where ground-truth data are limited. This study investigates vegetation change across Birmingham, UK, using three unsupervised change detection techniques that require neither labeled nor ground-truth data. A multi-modal dataset was leveraged by combining optical (Sentinel-2) and C-band radar (Sentinel-1) observations through decision-level fusion. Three approaches were applied to the bitemporal composites from July 2021 (non-drought year) and July 2022 (drought year): (i) baseline image differencing with k-means clustering; (ii) floating references; and (iii) pre-trained representation learning via the Clay foundation autoencoder. Three methods demonstrated complementary strengths. Baseline differencing provided simplicity, speed, and reliable performance for fine-scale urban analysis at 10 m, although single-sensor detections were more susceptible to misclassifications of vegetation change due to rooftop-driven artefacts. The floating references method, operating at 10 m, reduced bias from inter-scene drift and illumination differences, proving particularly effective for detecting agricultural sensitivity; however, it required greater computational demand. Meanwhile, the Clay-derived binary change map, generated through principal component analysis and Otsu’s thresholding, produced spatially coherent and semantically rich vegetation change patterns, though its coarser effective resolution (~80 m) likely smoothed fine-scale artefacts and contributed to higher spatial coherence, limiting direct comparison with the 10 m methods. Overall, the resulting change maps offer valuable and complementary insights for urban vegetation management under a changing climate, thereby supporting the use of unsupervised multi-sensor approaches for vegetation monitoring in data-scarce urban environments. Future work should explore the incorporation of L-band radar, the generation of a consensus agreement map for species-level change analysis, the examination of soft clustering algorithms, and an expansion of the regional scope to quantify drought-associated vegetation change across broader regions.