Enhanced ephemeral gully mapping through multi-classifier integration and spectral feature analysis

• A new approach was developed for ephemeral gully mapping. • Dempster-Shafer theory was used to integrate the different classifiers’ results. • The ephemeral gullies were mapped based on spectral features. • The proposed approach accuracy was higher than the common classifiers. The mapping of ephemeral gullies (EGs) is essential for improving and managing agriculture, but it poses challenges in terms of their identification, monitoring, and measurement. The primary objective of this study was to devise a novel approach that integrates multiple classifiers to map EGs. This was achieved by utilizing spectral features extracted from Pleiades-1 satellite imagery of the Niagara region in Canada, as a case study site, alongside a ground dataset collected during field visits, to train and validate the classifiers. Initially, maps were generated with spectral features deemed effective for EG identification, encompassing four spectral bands and eight spectral indices that reveal surface characteristics. Subsequently, four distinct classifiers, namely artificial neural network (ANN), logistic regression (LR), support vector machine (SVM), and random forest (RF), were employed to produce EG maps. In the third phase, the Dempster-Shafer (D-S) theory was employed to amalgamate the results from all classifiers, thereby enhancing the accuracy of the EGs map. Lastly, the performance of the various classifiers was evaluated using diverse metrics, including user accuracy, producer accuracy, overall accuracy, prediction rate, and receiver operating characteristics (ROC) analysis. The most influential variables in identifying EGs were determined to be Norm NIR (18%), Soil line (15%), NDVI (12%), and NDWI (10%). The average producer (user) accuracy for EGs and non-EGs classes across all four classifiers was 0.53 (0.67) and 0.97 (0.95), respectively. Incorporating the D-S theory improved these accuracy values to 0.68 (0.86) for EGs and 0.99 (0.97) for non-EGs. Furthermore, the overall accuracy (prediction rate) for EGs mapping, based on ANN, LR, SVM, RF classifiers, and D-S, was 0.94 (8.2), 0.94 (9.7), 0.93 (7.7), 0.95 (10.1), and 0.97 (12.5), respectively. ROC analysis revealed that the D-S classifier exhibited the highest accuracy in EG identification, while LR performed the least effectively. In summary, this research underscores that the proposed ensemble modeling approach for mapping EGs surpasses traditional classifiers in meeting accuracy criteria, showcasing its promising potential for guiding future informed decision-making processes.