| Texto completo | |
| Autor(es): |
Miranda, Mateus de Souza
;
Hansch, Ronny
;
Junior, Valdivino Alexandre de Santiago
;
Korting, Thales Sehn
;
Monteiro, Erison Carlos dos Santos
Número total de Autores: 5
|
| Tipo de documento: | Artigo Científico |
| Fonte: | IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING; v. 18, p. 11-pg., 2025-01-01. |
| Resumo | |
The Cerrado faces increasing environmental pressures, necessitating accurate land use and land cover mapping despite challenges, such as class imbalance and visually similar categories. To address this, we present CerraData-4 MM, a multimodal dataset combining Sentinel-1 synthetic aperture radar and Sentinel-2 multispectral imagery with 10 m spatial resolution. The dataset includes two hierarchical classification levels with seven and 14 classes, respectively, focusing on the diverse Bico do Papagaio ecoregion. We benchmark two models trained on CerraData-4 MM, employing a visual transformer-based architecture and a convolutional-based architecture. The ViT achieves superior performance in multimodal scenarios, with the highest macro F1-score of 57.60% and a mean Intersection over Union of 49.05% at the first hierarchical level. Both models struggle with minority classes, particularly at the second hierarchical level, where U-Net's performance drops to an F1-score of 18.16%. Weighted loss improves representation for underrepresented classes but reduces overall accuracy, underscoring the trade-off in weighted training. CerraData-4 MM offers a challenging benchmark for advancing deep learning models to handle class imbalance and multimodal data fusion. (AU) | |
| Processo FAPESP: | 23/09118-6 - Descoberta de conteúdo em catálogos de imagens de sensoriamento remoto |
| Beneficiário: | Thales Sehn Körting |
| Modalidade de apoio: | Auxílio à Pesquisa - Regular |