Global vegetation maps are essential tools for ecosystem monitoring, yet they often fail to accurately represent vegetation in arid regions. This study evaluates the performance of four global products-GFCC, GFW, CGLSLC100, and DW-in the xerophytic shrublands of southwestern Madagascar, a region undergoing rapid deforestation. Satellite-derived estimates were compared with field data collected in 2022 from 41plots (900 m2 each) using the point-intercept method. Results revealed substantial discrepancies among products, with a coefficient of variation reaching 78.94 % for tree cover, primarily due to differences in vegetation definitions and classification methods. Aggregating tree and shrub cover into a single woody cover layer significantly reduced this variability by 58.19 %. Among the products tested, DW provided the most accurate estimates of woody cover (R2 = 0.70), while CGLS-LC100 was the most responsive to floristic composition. Product biases varied by vegetation type: DW and CGLS-LC100 tended to overestimate tree cover in structurally developed plots, while GFCC and GFW underestimated it in multi-stemmed thickets. Shrub cover was consistently underestimated across all products. In this context of active forest loss, overestimations may reflect past landscape conditions, while underestimations expose structural limitations. These findings highlight the need for mapping strategies that better capture the complexity of dryland vegetation.