A global 1‐km consensus land‐cover product for biodiversity and ecosystem modelling

Abstract

Aim For many applications in biodiversity and ecology, existing remote sensing‐derived land‐cover products have limitations due to among‐product inconsistency and their typically non‐continuous nature. Here we aim to help address these shortcomings by generating a 1‐km resolution global product that provides scale‐integrated and accuracy‐weighted consensus land‐cover information on an approximately continuous scale. Location Global. Methods Using a generalized classification scheme and an accuracy‐based integration approach, we integrated four global land‐cover products. We evaluated the performance of this product compared with inputs for estimating subpixel 30‐m resolution land cover. We also compared the accuracy of deductive and inductive species distribution models built with the different products for modelling the continental distributions of six avian habitat specialists. Results Our product offers accuracy‐weighted consensus information on the prevalence of 12 land‐cover classes within every nominal 1‐km pixel across the globe (except for Antarctica). Compared with the four base products, it better captures the land‐cover information contained in the fine‐grain validation data for all classes combined and for most individual classes. It also has the highest sensitivity and overall accuracy for detecting the presence of every fine‐grain land‐cover class. Both deductive and inductive models built with the consensus dataset have the highest or second highest accuracy for modelling bird species distributions. Main conclusions Our consensus product integrates the four base products and successfully maximizes accuracy and reduces errors of omission. Specifically, the consensus product reduces limitations caused by misclassifications, false absence rates and the categorical format of existing land‐cover products. Consequently, it surpasses single base products in the ability to capture subpixel land‐cover information and the utility for modelling species distributions. Both the presented methodology and the consensus product have multiple applications in biodiversity research and for understanding and modelling of global terrestrial ecosystems.

Plant functional type classification for earth system models: results from the European Space Agency's Land Cover Climate Change Initiative

“…Following the requests of the user survey, the land cover data set is available across multiple spatial domains, conforms to standard file formats used in numerical models, and includes information on classification confidence levels for the land cover classes and resulting PFT fractions. The standardized conversion tool provides users with a consistent documented approach for aggregating land cover classes and thus overcomes limitations associated with consensus approaches (e.g., Tuanmu and Jetz 2014). Of particular importance is that the multitemporal LC_CCI mapping approach facilitated more accurate mapping leading to improved remote sensing observations of deforested areas in the tropics, the tree line-tundra boundary in the high latitudes, and better distinctions between managed and non-managed grasslands in Africa.…”

Towards an Integrated Global Land Cover Monitoring and Mapping System

“…Moreover, the MODIS Collection 5 maps was developed with five different legends to address the thematic requirements of different users [5]. Map accuracy assessment, as well as map integration, has also been conducted by taking the perspectives and needs of specific users into account [37][38][39].…”

Biodiversity Modelling as Part of an Observation System

“…The second activity, which we here call 'predictive modelling' (Shmueli 2010), instead presumes that a model describing the relationship of interest is already known, as are observed or estimated values of the relevant predictor variables, and therefore combines these to predict previously unknown values of the biodiversity response variable. The model used to make such predictions can be either an 'inductive model' derived through data analysis, in which case the activities of data analysis and prediction are integrally linked, or a 'deductive model' built directly from existing expert knowledge of the relationship between response and predictor variables (Corsi et al 2000;Overmars et al 2007;Tuanmu and Jetz 2014).…”