Mainali K., Evans M., Saavedra D., Mills E., Madsen B., Minnemeyer S. (2023). Convolutional neural network for high-resolution wetland mapping with open data: variable selection and the challenges of a generalizable model. Science of the Total Environment, 25/02/2023, vol. 861, p. 1-11.
https://doi.org/10.1016/j.scitotenv.2022.160622
https://doi.org/10.1016/j.scitotenv.2022.160622
Titre : | Convolutional neural network for high-resolution wetland mapping with open data: variable selection and the challenges of a generalizable model (2023) |
Auteurs : | K. Mainali ; M. Evans ; D. Saavedra ; E. Mills ; B. Madsen ; S. Minnemeyer |
Type de document : | Article |
Dans : | Science of the Total Environment (vol. 861, February 2023) |
Article en page(s) : | p. 1-11 |
Langues : | Anglais |
Langues du résumé : | Anglais |
Catégories : |
Catégories principales 07 - ENVIRONNEMENT ; 7.4 - Ressources Naturelles : Paysage, Biodiversité, Patrimoine naturelThésaurus IAMM CARTOGRAPHIE ; ZONE HUMIDE ; RESEAU DE NEURONES ; DONNEES OUVERTES ; TELEDETECTION ; MODELE |
Résumé : | Landscape scale wetland conservation requires accurate, up-to-date wetland maps. The most useful approaches to creating such maps are automated, spatially generalizable, temporally repeatable, and can be applied at large spatial scales. However, mapping wetlands with predictive models is challenging due to the highly variable characteristics of wetlands in both space and time. Currently, most approaches are limited by coarse resolution, commercial data, and geographic specificity. Here, we trained a deep learning model and evaluated its ability to automatically map wetlands at landscape scale in a variety of geographies. We trained a U-Net architecture to map wetlands at 1-meter spatial resolution with the following remotely sensed covariates: multispectral data from the National Agriculture Imagery Program and the Sentinel-2 satellite system, and two LiDAR-derived datasets, intensity and geomorphons. The full model mapped wetlands accurately (94 % accuracy, 96.5 % precision, 95.2 % AUC) at 1-meter resolution. Post hoc model evaluation showed that the model correctly predicted wetlands even in areas that had incorrect label/training data, which penalized the recall rate (90.2 %). Applying the model in a new geography resulted in poor performance (precision = ~80 %, recall = 48 %). However, limited retraining in this geography improved model performance substantially, demonstrating an effective means to create a spatially generalizable model. We demonstrate wetlands can be mapped at high-resolution (1 m) using free data and efficient deep-learning models that do not require manual feature engineering. Including LiDAR and geomorphons as input data improved model accuracy by 2 %, and where these data are unavailable a simpler model can efficiently map wetlands. Given the dynamic nature of wetlands and the important ecosystem services they provide, high-resolution mapping can be a game changer in terms of informing restoration and development decisions. |
Cote : | En ligne |
URL / DOI : | https://doi.org/10.1016/j.scitotenv.2022.160622 |