Automatic segmentation of river plumes from satellite data using deep learning methods

A.N. Yakusheva, Nikita A. Knyazev, M.V. Vrublevsky, Olga Yu. Lavrova, Polina D. Zhadanova, Andrey G. Kostianoy, Emil Bayramov, Saida Aliyeva

2026

Closed access Metadata (CSL-JSON)
DOI 10.1080/01431161.2026.2726572
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

River plumes play an important role in the formation of coastal marine ecosystems; however, their automatic detection from satellite data is complicated by blurred boundaries, high shape variability, and weak contrasts with the surrounding water area. This paper considers the application of deep learning methods for the segmentation of river plumes in satellite images. Optical data from MSI sensors of Sentinel-2 satellites, as well as OLI/OLI-2 of Landsat-8/9 satellites with spatial resolution of 10–30 m were used. The study was performed for the areas of Sulak, Terek, Ural, and Kura rivers in the Caspian Sea, the Mzymta River in the Black Sea, and the Kaliningrad Bay outflow and the Vistula River in the Baltic Sea. The first contribution of this work is a dedicated expert-annotated dataset of 744 satellite images covering seven estuarine systems in three seas, combined with an adaptive, brightness-aware photometric augmentation pipeline; on this basis, convolutional and transformer segmentation architectures were compared, including U-Net, DeepLabv3+, FCN, YOLO11seg, SegFormer, and Mask2Former. Quality assessment was conducted using the Dice coefficient, Precision/Recall metrics, and Hausdorff distances. Second, we propose a weighted ensemble segmentation framework in which the predictions of four complementary architectures (U-Net, DeepLabv3+, YOLO11seg, and Mask2Former) are combined with weights proportional to their validation performance; the resulting ensemble outperforms every individual model in overall mean per-scene Dice on the test set, and its lowest regional score remains above the lowest regional score of any individual model.