Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; the references it cites are not checked yet; none of the 142 sources its claims cite is checked for support yet.

An AI approach to operationalise global daily PlanetScope satellite imagery for river water masking

Open access

Abstract

Monitoring rivers is vital to manage the invaluable ecosystem services they provide, and also to mitigate the risks they pose to property and life through flooding and drought. Due to the vast extent and dynamic nature of river systems, Earth Observation (EO) is one of the best ways to measure river characteristics. As a first step, EO-based river monitoring often requires extraction of accurate pixel-level water masks, but satellite images traditionally used for this purpose suffer from limited spatial and/or temporal resolution. We address this problem by applying a novel Convolutional Neural Network (CNN)-based model to automate water mask extraction from daily 3 m resolution PlanetScope satellite imagery. Notably, this approach overcomes radiometric issues that frequently present limitations when working with CubeSat data. We test our classification model on 36 rivers across 12 global terrestrial biomes (as proxies for the environmental and physical characteristics that lead to the variability in catchments around the globe). Using a relatively shallow CNN classification model, our approach produced a median F1 accuracy score of 0.93, suggesting that a compact and efficient CNN-based model can work as well as, if not better than, the very deep neural networks conventionally used in similar studies, whilst requiring less training data and computational power. We further show that our model, specialised to the task at hand, performs better than a state-of-the-art Fully Convolutional Neural Network (FCN) that struggles with the highly variable image quality from PlanetScope. Although classifying rivers that were narrower than 60 m, anastomosed or highly urbanised was slightly less successful than our other test images, we showed that fine tuning could circumvent these limitations to some degree. Indeed, fine tuning carried out on the Ottawa River, Canada, by including just 5 additional site-specific training images significantly improved classification accuracy (F1 increased from 0.81 to 0.90, p < 0.01). Overall, our results show that CNN-based classification applied to PlanetScope imagery is a viable tool for producing accurate, temporally dynamic river water masks, opening up possibilities for river monitoring investigations where high temporal variability data is essential.

Overview

CoChat Summary

TL;DR

The paper develops a CNN-supervised classification (CSC) approach to extract river water masks from daily 3 m PlanetScope satellite imagery, testing it on 36 rivers across 12 global biomes. The M20 model achieved a median F1 of 0.93 and could be fine-tuned with 5 additional site-specific images to improve Ottawa River mean F1 from 0.81 to 0.90.

Key findings

  • The 20×20-tile CSC model (M20) produced the highest median F1 score of 0.93 and achieved F1 ≥ 0.9 for over 60% of hold-out images; M10, M32, and a very deep CSC model were slightly lower (0.90, 0.92, 0.92).
  • Otsu thresholding had a similar median F1 (0.90) but a much lower mean and longer-tailed error distribution; the state-of-the-art Tiramisu FCN predicted only 4 of 36 images with sufficient accuracy to record F1 scores (0.52–0.87).
  • Fine tuning M20 on the Ottawa River with 5 additional training images significantly improved mean F1 from 0.81 to 0.90 (n = 15, t = 8.4, p < 0.01), with three of the worst images improving from <0.7 to >0.95.
  • Log-transformed river width was moderately correlated with F1 score (r² = 0.34, p < 0.01); no significant relationships were found with spectral band ranges, PlanetScope metadata, or other visible physical river characteristics (p > 0.05).

Methods

  • Applied CSC to 4-band, 3 m PlanetScope surface reflectance imagery; selected 36 rivers from 12 WWF terrestrial biomes (3 rivers per biome), with 72 training scenes, 20×20-pixel tiles, and 393,000 training tiles.
  • Compared shallow CNN models M10, M20, and M32 against a VGG16-based very deep CSC control, a Tiramisu FCN, and Otsu segmentation; CSC uses CNN tile predictions to train a localised MLP for pixel-level segmentation, followed by a morphological operator to remove speckle.
  • Fine-tuning used 20 Ottawa River images (15 test, 5 additional training images labelled with the QGIS magic wand tool); statistical analysis used F1 scores, linear regression, and paired samples t-tests after Box-Cox transformation.

Limitations

  • Stated: lack of field data prevented formal ground validation, and the global scope precluded such an approach; manually intensive validation labelling also carries its own error, which inherently prevents perfect F1 scores.
  • Stated: performance was lower for rivers narrower than 60 m, anastomosed or braided channels, highly urbanised areas, and images with bridges, instream features, cloud, ice, or deep shadow; fine tuning improved Ottawa River results but slightly reduced overall mean F1 from 0.83 to 0.81, indicating reduced generalisability.
  • Evident from the text: the method relies on proprietary PlanetScope imagery and 4-band data without a shortwave infrared band, and training labels were simple polygons while test labels were semi-automatic, both of which the authors note can introduce error or limit transferability.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.rse.2023.113932

Cites

Statistics

Nothing to recompute.

Figures

Image 1
CC-BY-NC · From the PDF
Fig. 1 The Global terrestrial biomes presented by (Olson et al., 2001) overlayed with the rivers used in this study for testing and training. 12 biomes were selected for this study with 3 rivers selected from each biome. Additional sites used for testing the ability of the algorithm to be fine-tuned were in the area here.
Fig. 1 The Global terrestrial biomes presented by (Olson et al., 2001) overlayed with the rivers used in this study for testing and training. 12 biomes were selected for this study with 3 rivers selected from each biome. Additional sites used for testing the ability of the algorithm to be fine-tuned were in the area here.CC-BY-NC · From the PDF
Image 3
CC-BY-NC · From the PDF
Fig. 3 Comparison of the ability of tested models to predict hold-out images. All models shown here are skewed towards high F1 scores. There are different dis­ tributions of the scores for the weakest predictions made by each model. M20 in particular never predicts as poorly as the other models and has a thinner tail suggesting no clustering of poorly predicted images.
Fig. 3 Comparison of the ability of tested models to predict hold-out images. All models shown here are skewed towards high F1 scores. There are different dis­ tributions of the scores for the weakest predictions made by each model. M20 in particular never predicts as poorly as the other models and has a thinner tail suggesting no clustering of poorly predicted images.CC-BY-NC · From the PDF
Fig. 4 Selection of water masks produced using M20. The majority of these rivers scored an F1 accuracy > 0.97. However, the River Thames, through central London, is also included with a score of 0.85 to display model results in the most complex urban environments with multiple watercraft and bridges in a small spatial area (i). In this case, it has still produced a fairly accurate water mask in demanding circumstances. Rivers shown here a: Niger (Mali), Tropical Grasslands . b: Amazon (Brazil), Tropical Moist Broadleaf . c: Volga (Russia), Temperate Grasslands . d: Nile (Egypt), Flooded Grasslands . e: Betsiboka (Madagascar), Tropical Moist Broadleaf . f: Xi (China), Tropical Moist Broadleaf . g: Victoria (Australia), Tropical Grasslands . h: Pyasina (Russia), Tundra . i: Thames (UK), Temperate Broadleaf, j: Mackenzie (Canada), Taiga. K: Slave (Canada), Taiga . l: Congo (Democratic Republic of the Congo), Tropical Grasslands.
Fig. 4 Selection of water masks produced using M20. The majority of these rivers scored an F1 accuracy > 0.97. However, the River Thames, through central London, is also included with a score of 0.85 to display model results in the most complex urban environments with multiple watercraft and bridges in a small spatial area (i). In this case, it has still produced a fairly accurate water mask in demanding circumstances. Rivers shown here a: Niger (Mali), Tropical Grasslands . b: Amazon (Brazil), Tropical Moist Broadleaf . c: Volga (Russia), Temperate Grasslands . d: Nile (Egypt), Flooded Grasslands . e: Betsiboka (Madagascar), Tropical Moist Broadleaf . f: Xi (China), Tropical Moist Broadleaf . g: Victoria (Australia), Tropical Grasslands . h: Pyasina (Russia), Tundra . i: Thames (UK), Temperate Broadleaf, j: Mackenzie (Canada), Taiga. K: Slave (Canada), Taiga . l: Congo (Democratic Republic of the Congo), Tropical Grasslands.CC-BY-NC · From the PDF
Fig. 5 Within region F1 scores for biomes predicted by M20. Based on these results, biome predictability was classified as good, varied, or poor.
Fig. 5 Within region F1 scores for biomes predicted by M20. Based on these results, biome predictability was classified as good, varied, or poor.CC-BY-NC · From the PDF
Fig. 6 Comparison of CSC predictive workflows for well and poorly predicted images in the poor performing biomes. Despite narrow sinuous channels, the Vaal River in South Africa, from the Tropical Dry Broadleaf biome, scored an F1 score of 0.86 (a). Other images in the same biome struggled with issues such as sand bars and urban areas, including the Narmada River in India (b) with an F1 score of 0.72. It is clear that if the CNN prediction is particularly poor, then despite the predictive improvements made by the ANN section, the final prediction will suffer.
Fig. 6 Comparison of CSC predictive workflows for well and poorly predicted images in the poor performing biomes. Despite narrow sinuous channels, the Vaal River in South Africa, from the Tropical Dry Broadleaf biome, scored an F1 score of 0.86 (a). Other images in the same biome struggled with issues such as sand bars and urban areas, including the Narmada River in India (b) with an F1 score of 0.72. It is clear that if the CNN prediction is particularly poor, then despite the predictive improvements made by the ANN section, the final prediction will suffer.CC-BY-NC · From the PDF
Fig. 7 F1 score as function of variability in external (e.g., physical river characteristics) and internal (image metadata) properties. Log transformed width correlates significantly to F1 score ( r 2 = 0.34, p < 0.01), which is to be expected due to the greater importance of differences in smaller river widths, as opposed to differences in the widths of larger rivers. No other variables were found to be significantly correlated.
Fig. 7 F1 score as function of variability in external (e.g., physical river characteristics) and internal (image metadata) properties. Log transformed width correlates significantly to F1 score ( r 2 = 0.34, p < 0.01), which is to be expected due to the greater importance of differences in smaller river widths, as opposed to differences in the widths of larger rivers. No other variables were found to be significantly correlated.CC-BY-NC · From the PDF
Fig. 8 Examples highlighting where the M20 CSC model coped or struggled with more difficult images. The first column displays the original satellite image, the second displays the prediction, and the third shows correct predictions in blue and incorrect predictions in red. a: Amazon, Brazil with unprocessed cloud. b: Popigay, Russia with snow covered frozen pools . c: Ahuri, New Zealand, deep shade which resulted in a poor prediction. However, further research is needed because the model still separated the river in both shaded and unshaded sections of the image. Example c was not included in the analysis when selecting hold out images but was added as an additional test image to see how the model coped with more extreme examples. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 8 Examples highlighting where the M20 CSC model coped or struggled with more difficult images. The first column displays the original satellite image, the second displays the prediction, and the third shows correct predictions in blue and incorrect predictions in red. a: Amazon, Brazil with unprocessed cloud. b: Popigay, Russia with snow covered frozen pools . c: Ahuri, New Zealand, deep shade which resulted in a poor prediction. However, further research is needed because the model still separated the river in both shaded and unshaded sections of the image. Example c was not included in the analysis when selecting hold out images but was added as an additional test image to see how the model coped with more extreme examples. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)CC-BY-NC · From the PDF
Fig. 9 Difference in F1 scores between images predicted by the M20 model and after 5 images from the same river were included in the training data to fine tune it. Improvements are not universal but are dramatically better for the worst performing images.
Fig. 9 Difference in F1 scores between images predicted by the M20 model and after 5 images from the same river were included in the training data to fine tune it. Improvements are not universal but are dramatically better for the worst performing images.CC-BY-NC · From the PDF

Verification

Claims the paper makes

0 Valid · 5 Not valid · 96 Not checked · 156 Own results and interpretations

115 more claims

Create a free account to see all 257 claims and to verify the 125 references still open.

See all claims — free account
  • Probably not a substantial claim. 76
  • It may state two separate claims as one. 88
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

The citation lists haven’t been fetched yet

References 103 · Cited by 70

Research works printed but not linked: 1

  1. Bradski, 2000, OpenCV, Dr. Dobb’s J. Softw. Tools, 3, 120

Metadata

Authors
Valman, S. J., Boyd, D. S., Carbonneau, P. E., Johnson, M. F., & Dugdale, S. J.
Affiliations
  • University of Nottingham (GB)
  • Durham University (GB)
Corresponding author
Valman, Samuel J.
Venue
Remote Sensing of Environment
Publisher
Elsevier BV
Published
2024-02
Volume
301
Pages
113932
ISSN
0034-4257 (ISSN-L)
OpenAlex
W4389469222
Cited by
70
References
103
Citations by year
2026: 24 · 2025: 33 · 2024: 13
FWCI
4.45
Type
journal article
Language
en
Open access
Licence
Locations
  • Remote Sensing of Environment (publishedVersion, cc-by-nc)
  • Repository@Nottingham (University of Nottingham) (publishedVersion, cc-by-nc)
  • Durham Research Online (Durham University) (publishedVersion, other-oa)
  • Repository@Nottingham (University of Nottingham) (publishedVersion, other-oa)
Funders
Engineering and Physical Sciences Research Council (EP/ S023577 /1)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Remote-Sensing Image Classification
  • Physical Sciences › Environmental Science › Global and Planetary Change › Flood Risk Assessment and Management
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
PlanetScope imagery, CubeSat, convolutional neural networks, water body classification, river monitoring, Earth observation, fine-tuning, Ottawa River, terrestrial biomes

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-18, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-18 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: An AI approach to operationalise global daily PlanetScope satellite imagery for river water masking. https://app.cochat.ai/papers/10.1016/j.rse.2023.113932
The paper
Valman, S. J., Boyd, D. S., Carbonneau, P. E., Johnson, M. F., & Dugdale, S. J. (2024). An AI approach to operationalise global daily PlanetScope satellite imagery for river water masking. Remote Sensing of Environment. https://doi.org/10.1016/j.rse.2023.113932

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.