A novel approach to field data augmentation with remote sensing and machine learning in rangelands
Rangelands provide essential ecosystem services, supporting biodiversity, sequestering carbon, and sustaining millions of livelihoods worldwide. However, these ecosystems are increasingly threatened by woody plant encroachment (WPE), which impacts forage availability, habitat quality, and other ecos...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Ecological Informatics |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1574954125003620 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Rangelands provide essential ecosystem services, supporting biodiversity, sequestering carbon, and sustaining millions of livelihoods worldwide. However, these ecosystems are increasingly threatened by woody plant encroachment (WPE), which impacts forage availability, habitat quality, and other ecosystem functions. Despite advances in remote sensing and machine learning (ML) for monitoring and managing WPE, a significant gap persists in providing effective, scalable, and high-quality training data to support early detection efforts. We applied an RF model to classify vegetation types in a savanna rangeland, evaluating four training data approaches: (1) visual interpretation, (2) subsampled visual interpretation, (3) field-collected data, and (4) a novel neighborhood-based augmentation of field data. A total of 47 spectral, textural, topographic, and edaphic predictors were derived from PlanetScope imagery and open-source datasets. Overall accuracies ranged from 51.14 to 75.66 % on the test set, with the neighborhood-based approach achieving the highest accuracy. Our findings demonstrate that augmenting field-collected data with visually verified neighboring pixels improves RF classification performance, offering a practical solution to address the scarcity of high-quality training data. This study introduces a framework for enhancing training data in RF-based vegetation classification and highlights the value of neighborhood-based augmentation for improving early detection of woody encroachment in rangelands. |
|---|---|
| ISSN: | 1574-9541 |