Guardado en:
| Autores principales: | , , , , |
|---|---|
| Formato: | Recurso digital |
| Lenguaje: | |
| Publicado: |
Zenodo
2025
|
| Acceso en línea: | https://doi.org/10.5281/zenodo.15013314 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
Tabla de Contenidos:
- <p>Ensuring sufficient food for the rising global population is a critical focus, especially given the<br>current turbulent political circumstances. The first step for policymakers in addressing this issue<br>is acquiring accurate crop maps. Modern Earth observation missions and advancements in data<br>processing, particularly in machine learning, have enabled the successful generation of these<br>maps. Although considered a traditional method nowadays, the Random Forest algorithm has<br>consistently proven effective for crop classification tasks. Conversely, while synthetic aperture<br>radar (SAR) satellite data lags behind optical satellite missions in terms of practical usage, its<br>contribution remains significant. In this study, we combined Sentinel-1 data with Random<br>Forest to classify nine crop types in Vojvodina, Serbia. Specifically, we assessed how varying<br>amounts of information from radar satellites impact crop classification accuracy by analyzing<br>three data availability scenarios: 1) single monthly acquisitions, 2) bi-monthly acquisitions<br>(every 15 days), and 3) all available acquisitions. For the first two scenarios, bicubic<br>interpolation was used. Out of 2399 parcels, 80% were used for training and validation (5-fold<br>cross-validation), with the remaining 20% for testing. We calculated the mean value for each<br>parcel and considered data acquisitions between April 1st and September 30th, 2021. The<br>scenarios achieved overall accuracies of 90%, 94%, and 96%, respectively, with high precision<br>and recall across nearly all classes, despite class imbalance. These findings underscore the<br>potential of SAR data for classifying diverse crop types and highlight the possibility of reducing<br>inputs for regional or global mapping typically characterized with high computational demands,<br>while still achieving remarkable results.</p>