Modelo de segmentación semántica de imágenes satelitales basado en redes neuronales convolucionales para la clasificación de cobertura de la tierra en páramos
[EN] Páramo ecosystems are essential for water regulation and biodiversity conservation in mountainous areas. However, they face significant threats due to climate change and human activities such as agriculture, livestock farming, and mining. The absence of clear boundaries and continuous monitorin...
| Autores: | , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universitat Politècnica de València (UPV) |
| Repositorio: | RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia |
| Idioma: | español |
| OAI Identifier: | oai:riunet.upv.es:10251/213934 |
| Acceso en línea: | https://riunet.upv.es/handle/10251/213934 |
| Access Level: | acceso abierto |
| Palabra clave: | Remote sensing Páramo Convolutional neural networks Semantic segmentation Parámo Teledetección Redes neuronales convolucionales Segmentación semántica |
| Sumario: | [EN] Páramo ecosystems are essential for water regulation and biodiversity conservation in mountainous areas. However, they face significant threats due to climate change and human activities such as agriculture, livestock farming, and mining. The absence of clear boundaries and continuous monitoring systems for their land cover hinders effective protection highlighting the need to employ advanced digital techniques that provide highly accurate, up-to-date information. Convolutional neural networks (CNNs) have emerged as promising tools for semantic segmentation of satellite images. This research aimed to evaluate the performance of two CNNs architectures U-Net++ and DeepLabV3+ for land cover classification in the Tota-Bijagual-Mamapacha (TBM) páramo complex in Colombia, using Landsat 8 imagery from 2017 to 2019 and land cover labels from 1:100.000, national coverage map produced by IDEAM in 2018. The results showed U-Net++ achieved a kappa of 0.60, while DeepLabV3+ obtained a kappa of 0.59. In páramo covers, U-Net++ achieved an F1 of 78.43% for Herbazal and 79.22% for Forests, while DeepLabV3+ achieved F1 of 75% and 74.27%, respectively, confirming the potential of CNNs for land cover classification in these ecosystems. Although both models presented similar processing times, class imbalance and reliance on consistent labels affected their performance in heterogeneous covers. This research establishes a methodological foundation for future studies and suggests addressing these limitations toimprove efficiency and thematic accuracy in the classification and monitoring in páramo ecosystems. |
|---|