Automatic Extraction of Oxbow Lakes from Landsat Imagery Using Deep Learning Across the Amazon Basin
TL;DR
This study provides a new image-based semantic segmentation approach for basin-scale oxbow lake mapping and offers an expanded spatial inventory to support floodplain geomorphology, river dynamics, and river–floodplain ecosystem studies.
Abstract
Oxbow lakes are key geomorphic and ecological components of river floodplains, but basin-scale inventories remain limited because these features are narrow, fragmented, and spectrally similar to active channels and other floodplain water bodies. This study developed a deep-learning-based workflow for oxbow lake extraction in the Amazon Basin using Landsat imagery. A reference label dataset was constructed by combining automated cutoff extraction, surface-water transition information, and manual interpretation. An enhanced DeepLabV3+ semantic segmentation model was trained with nine-band Landsat image patches and applied to basin-wide Landsat 8/9 composite images through sliding-window inference and post-processing. Independent validation against manually digitized PlanetScope imagery showed strong detection performance, with an overall F1-score of 0.819 and recall of 0.966. A spatially independent sub-basin validation further indicated partial transferability in an unseen Amazon sub-basin, with an F1-score/Dice coefficient of 0.532, precision of 0.498, recall of 0.571, and IoU of 0.362. Basin-scale inference detected 1836 oxbow lake objects, approximately 2.34 times the number of reference labels. Our study provides a new image-based semantic segmentation approach for basin-scale oxbow lake mapping and offers an expanded spatial inventory to support floodplain geomorphology, river dynamics, and river–floodplain ecosystem studies.