[DATASET] Remote sensing imagery of threatened Vereda wetlands in the Brazilian Midwest: a segmentation-oriented model as a backbone for domain-specific applications

This repository provides the annotated dataset and pretrained U-Net model for segmenting Vereda wetlands (VED.SGF) in the Brazilian Cerrado. The data/model/ directory contains the final architecture (model.keras) and weights (model.weights.h5). The data/train/ folder includes the RGB mosaic (box_train_4m_mosaic.tif), annotation masks (box_train_4m_labels.tif), and training/validation files (train.h5, validation.h5). The data/predict/ folder provides an example input tile and its corresponding prediction. Shapefiles defining training/test regions (shapefiles/LearningBox) and the study area (shapefiles/StudyArea) are also included. Downloading this dataset complements the source code available on our GitHub (https://github.com/wetlands-vision/wetlands-segmentation-unet), enabling full reproducibility of training, evaluation, and inference workflows.