Ecosystem condition assessments: A context-specific workflow to integrate local expert knowledge and remote sensing

Stephni van der Merwe , Vernon Visser, Colleen Seymour, M. Timm Hoffman, Glenn Moncrieff, Jasper A. Slingsby, Janine Steytler, Alastair J. Potts, Andrew L. Skowno

Despite decades of conservation science, we still struggle with a deceptively simple question: how do we know if an ecosystem is in good or poor condition? We developed a reproducible six-step workflow for assessing ecosystem condition across South Africa's terrestrial ecosystem types, and demonstrated it in Arid Thicket. Rather than prescribing universal indicators or analytical methods, the workflow standardizes the decision process for identifying ecosystem experts and assessment units, identifying key pressures, defining natural reference states, selecting ecological indicators and appropriate remote sensing metrics, and mapping and validating ecosystem condition. Expert knowledge guides decisions throughout. Arid Thicket is a dense, evergreen, woody-succulent vegetation where the key pressure, chronic livestock browsing, removes palatable plants, opens the canopy and exposes bare soil. However, this well-documented degradation pathway is poorly captured by national land cover products and remains classified as “natural”. Experts confirmed the natural reference state description, identified browsing-related ecological responses and verified 425 natural reference sites. AlphaEarth satellite embeddings were used, which summarizes remotely sensed surface characteristics of each pixel as a numerical signature. We averaged the signatures of natural reference sites and calculated cosine distance, with larger values indicating greater multivariate distance from the reference signature. We evaluated the distance metric against 425 independent expert-labelled sites representing “intact”, “moderately browsed” and “severely browsed” vegetation. Among reference sites, cosine distance was unrelated to rainfall or temperature and across expert-labelled sites, distance increased with browsing intensity (Spearman’s ρ = 0.634, p < 0.001) and reproduced fence-line contrasts. Thresholds selected by maximizing quadratic weighted kappa classified 75.3% of the remaining natural extent as severely browsed. The case study demonstrates the potential of reference-based condition mapping where natural reference sites are identifiable, one pressure dominates the degradation pathway and the ecological response produces a detectable signal. In more disturbance-driven ecosystems, or those affected by multiple interacting pressures, the workflow accommodates pressure-specific indicators, time-series metrics and alternative analytical approaches. The workflow complements existing ecosystem condition approaches by providing an explicit, expert-guided process for translating ecosystem-specific knowledge into spatial condition assessments. Its outputs could support biodiversity reporting, ecosystem risk assessments, ecosystem accounting and restoration prioritzation.