The rapid expansion of freely available Earth observation data, combined with advancements in forest monitoring capabilities, has led to the production of a variety of datasets on forest cover and change at the national and global scales. These datasets are essential for informing science and policy. Yet, this surge in available data products comes with the challenge of reconciling differences in the mapping of forest, non-forest, and forest loss across products. Differences in forest cover mapping between products result in divergent estimates of forest area change, potentially constraining the utility of these products to inform science and policy, consistently, across scales. We use the case of Colombia to demonstrate that these inconsistencies can be largely resolved by advancing operational forest monitoring capabilities toward the representation of forest cover and change as continuous rather than discrete variables. The analysis consisted of a comparison between the global forest change product (GFC) and the official national datasets on forest cover change for Colombia (IDEAM) in terms of overall and per-class classification agreement and estimates of forest-area loss. The comparison was performed after harmonizing the GFC dataset (HGFC) to maximize its overall classification agreement with the IDEAM map, based on the adoption of optimum percent tree cover thresholds, applied to GFC for subnational units. Based on these results, we evaluate whether classification agreement and disagreement between mapped forest and non-forest classes by HGFC and IDEAM can be explained by differences in continuous physical forest attributes, represented by forest Canopy Height (CH) and Above-Ground Biomass Density (AGBD). We produced the analysis for the entire country and also for mountainous and non-mountainous areas separately. We show that the optimal threshold in tree cover that maximizes map agreement, varies largely between subnational units, with values ranging between 20% and 100%. Results show a high overall agreement and also high agreement for stable forest and non-forest classes. Agreement between pixels classified as forest loss was much lower, regardless of the map used as reference. Classification agreement was lower for mountainous areas than for non-mountainous areas, particularly for forest loss and stable forest classes. Pixels representing classification disagreement were characterized by intermediate values of CH and AGBD that are significantly different from pixels classified as non-forest and forest for both maps, with significantly lower and higher CH and AGBD values, respectively. Differences were more notable in non-mountainous areas than in mountainous ones. We conclude that solving these large existing discrepancies in forest classification and forest loss estimates between datasets requires advancing operational capabilities toward the representation of forest cover and change as continuous rather than as categorical variables, with particular emphasis on mountainous areas. We illustrate how moving in such a direction can (1) promote consistency between forest cover definitions and map representation, (2) facilitate comparability among products and estimates of forest area loss at different scales of analysis, and (3) promote accountability in forest area loss assessments. © 2024 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.