| Sumario: | Errors in thematically detailed land-cover maps have large consequences for downstream applications. Moreover, simulation-based studies suggest that land-cover classifiers are sensitive to errors in reference data. We (1) quantified the expected error from field interpretation of land-cover types; (2) the sensitivity of classifiers to reference data errors; and (3) the error transferred from reference data to classifiers. Lastly, we (4) recommended strategies to reduce errors. The study area was mapped by 12 field interpreters divided into three equal-sized experience-level groups. The field-based land-cover maps were aggregated to three thematic resolutions and used to train 6804 land-cover classifiers by varying inputs, algorithm, and hyperparameter values. Separately from the first field campaign, four field interpreters classified validation data points, which were used to quantify error for each field interpreter and land-cover classifier, as the proportion of incorrectly classified validation points. We observed (1) generally high and varying levels of interpreter error; (2) a strong relationship between interpreter and classifier error; and (3) a net positive transfer of errors from reference data to classifiers. Because classifier error seems largely driven by interpreter error at the levels commonly observed in thematically detailed land-cover mapping, we (4) recommend strategies to reduce interpreter error before modelling.
|
|---|