Dataset for manuscript: Adapting global land-cover products to map fluvial sediment with incremental learning: an Amazon Basin case study

This repository contains the data and supporting materials used to develop MLFluv, an incremental-learning workflow for adapting global land-use and land-cover labels to fluvial-sediment segmentation from Sentinel-1 and Sentinel-2 imagery. The workflow first learns broad land-cover classes, then introduces fluvial sediment through targeted incremental training, followed by fine-tuning with manually interpreted reference labels. It distinguishes exposed fluvial sediment from other bare surfaces while retaining surrounding land-cover information. The dataset was prepared for an Amazon Basin case study. The experiments evaluate the workflow using U-Net, DeepLabV3+ and SegFormer-B2. This release contains manually interpreted reference dataset with 220 scenes, each covering 512 × 512 pixels at 10-m resolution. The repository supports model development, validation and reproducibility of the reported experiments. The manually interpreted masks represent the study’s application-specific sediment definition and should not be treated as an independently annotated external test benchmark. Detailed descriptions of dataset construction, label sources, training procedures and evaluation metrics are provided in the manuscript “Adapting global land-cover products to map fluvial sediment with incremental learning: an Amazon Basin case study”.