A prototype algorithm for daily water hyacinth monitoring at Hartbeespoortdam, South Africa, from Sentinel-3 OLCI data

Penning de Vries, Marloes; Shoko, Cletah; Dube, Timothy; Thamaga, Kgabo Humphrey; Salama, Suhyb

doi:10.5194/egusphere-2026-403

Preprints

https://doi.org/10.5194/egusphere-2026-403

Preprints

10 Feb 2026

| 10 Feb 2026

A prototype algorithm for daily water hyacinth monitoring at Hartbeespoortdam, South Africa, from Sentinel-3 OLCI data

Marloes Penning de Vries, Cletah Shoko, Timothy Dube, Kgabo Humphrey Thamaga, and Suhyb Salama

Abstract. Water HYacinth (WHY) is one of the world’s most disturbing invasive aquatic plant species, characterised by a high spatial and temporal variability. Remote sensing is a valuable approach to monitor WHY, as the dense, floating mats of vegetation can be detected using various satellite instruments, for example, OLCI on Sentinel-3. The multi-spectral instrument features only moderate spatial resolution (300 m), however, it achieves global coverage in two days, and with two instruments currently in orbit, it provides an opportunity to monitor WHY at near-daily resolution. This is crucial, considering that WHY cover patterns are highly variable due to the plants’ rapid reproduction and the influences of wind and hydrodynamics. We present the development of an algorithm for the creation of daily WHY maps by: (1) deriving WHY cover patterns from both OLCI instruments using the Normalized Difference Vegetation Index, NDVI; (2) merging the data sets into one with near-daily resolution; and (3) filling the gaps (due to missing observations or cloud interference) using a spatial-temporal interpolation scheme. We show that the gap-filling strategy leads to a consistent daily time series of WHY cover for the study region and increases the number of days with observations by 55%. A leave-one-out analysis showed that the interpolation algorithm performs well even for longer periods without observations, unless WHY cover patterns change abruptly. The presented algorithm is computationally light-weight and, although developed for Hartbeespoortdam Reservoir (South Africa), is easily adaptable for use with other water bodies. The prototype WHYmapping algorithm was used to analyse eighteen months of data (July 2022 – December 2023). In the future, long time series of daily WHY maps may provide an evaluation tool for WHY management and benefit water management strategies directly by allowing continuous monitoring of WHY.

Received: 23 Jan 2026 – Discussion started: 10 Feb 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Marloes Penning de Vries, Cletah Shoko, Timothy Dube, Kgabo Humphrey Thamaga, and Suhyb Salama

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-403', Anonymous Referee #1, 03 Jul 2026
General comments
The authors present a prototype algorithm that exploits time series of spectral indices derived from Sentinel-3 OLCI data to map floating aquatic vegetation cover. The approach incorporates multi-sensor data merging and spatio-temporal gap-filling capabilities. As a demonstration, the algorithm is applied to an 18-month dataset covering Hartbeespoortdam, a small artificial reservoir in South Africa that is heavily affected by the proliferation of water hyacinth, an invasive floating macrophyte.
The proposed framework has potential advantages over existing approaches, particularly regarding the gap-filling strategy, which appears straightforward and effective, although further evidence would be required to fully demonstrate its robustness and performance. Nevertheless, the manuscript presents several critical shortcomings that are especially relevant for a study introducing a new Earth Observation (EO)-based methodology.
First, there appears to be a mismatch between the stated target of the algorithm and the physical basis of the input variables. While water hyacinth is presented as the target species, the spectral features used as inputs are not specific to water hyacinth and are unable to discriminate among different types of floating materials characterized by pronounced red-edge signatures, such as floating macrophytes, algal scums, or similar vegetation aggregates. Second, the demonstration of the algorithm is limited both spatially and temporally. As a result, claims regarding transferability and broader applicability remain largely speculative and are not supported by empirical evidence. Third, several methodological aspects would benefit from further refinement and a more in-depth assessment, particularly concerning use of TOA radiance as input and potential anisotropy distortions.
In light of these considerations, I suggest that the authors further refine and optimize the algorithm for the specific conditions of Hartbeespoortdam and reposition the study as a comprehensive case study of water hyacinth—or, more broadly, floating vegetation—monitoring in this system. Such an approach could be significantly strengthened through the integration of historical EO datasets, for example by combining MERIS, OLCI, and potentially MODIS observations, thereby extending the analysis over a much longer temporal period and providing a more robust assessment of long-term vegetation dynamics.

Specific comments
Intro and rationale
The use of an acronym for the target species name is unnecessary and, in my opinion, does not improve readability. Referring to the species by its common name, or alternatively using generic expressions such as "the target species" where appropriate, would be sufficient to avoid excessive repetition.

The rationale presented in the Introduction appears to underestimate the extent of the existing literature on water hyacinth ecology, management, and control. Research on the biology of the species and on management approaches has been substantial, particularly over the last 10–15 years. A quick Scopus search using the query “TITLE-ABS-KEY ("water hyacinth" AND crassipes AND management approach*)” returns more than one hundred publications, including more than 30 review papers. This seems to contradict the statement suggesting a lack of systematic assessments of management strategies. Indeed, the subsequent paragraphs of the Introduction already cite a few relevant studies, indicating that a considerable body of literature is available and should be discussed more comprehensively.

The manuscript describes the proposed algorithm as computationally lightweight multiple times; however, this claim is not substantiated by quantitative evidence. No computational benchmarks or comparisons with alternative approaches are provided. Ideally, the authors should provide objective performance metrics demonstrating the computational efficiency of the method to support such claims.

Methods
OLCI is a wide-swath instrument characterized by substantial variability in viewing zenith angle (VZA) across the scene. For targets with pronounced anisotropic reflectance properties, such as aquatic environments and floating vegetation, VZA effects cannot be assumed negligible. The manuscript does not appear to account for directional effects, but the potential impact on the temporal consistency of the spectral indices used is relevant to understanding and interpreting (apparent) temporal changes in vegetation cover.

The proposed approach relies primarily on NDVI (based on red and near-infrared reflectance). However simple and seen as generally effective, this spectral information is insufficient to distinguish among different types of photosynthetically active floating materials. Similar NDVI values may be exhibited by different floating or floating-leaved macrophyte species, as well as by dense accumulations of phytoplankton scums. The latter are especially common in eutrophic, cyanobacteria-dominated systems such as Hartbeespoortdam. Consequently, the method appears more suitable for detecting generic floating vegetation (or floating phytomass) rather than specifically targeting water hyacinth.

The only discussion of direct field observations (L379–383) effectively reinforces the concerns raised above. Rather than supporting the specificity of the proposed approach for water hyacinth detection, these observations highlight the presence of multiple floating materials that may contribute to the observed signal. As a result, the field evidence appears to undermine, rather than strengthen, the central assumption that the algorithm is detecting water hyacinth specifically.

The use of TOA radiometric quantities as input data represents a major limitation. TOA observations do not provide standardized surface reflectance measurements and are therefore intrinsically affected by changes in atmospheric conditions, illumination and viewing geometry. This severely limits the transferability of the methodology across seasons, years, sensors, and sites.

Related to the previous point, the manuscript attributes part of the observed discrepancies between Sentinel-3A and Sentinel-3B to calibration differences between the two OLCI instruments. This explanation does not appear convincing. Instrument calibration differences are unlikely to account for NDVI discrepancies on the order of 0.1. Variations in viewing geometry (particularly VZA), atmospheric conditions, and the use of TOA rather than atmospherically corrected reflectance are more plausible explanations and should be investigated before drawing conclusions regarding sensor-related biases.

More generally, the manuscript does not adequately disentangle the respective contributions of instrument characteristics, observation geometry, atmospheric variability, and actual changes in floating vegetation cover. Without such analysis, there is a risk that part of the outcomes of the algorithm may reflect biases in input data and quantities more than water hyacinth variability.

Outcome/discussion
The assessment of the gap-filling procedure is very limited in scope, only five days, and does not incorporate actual data to be used as reference for assessing errors. While the results are qualitatively encouraging, the evidence provided is insufficient to support broader conclusions regarding the reliability, or applicability of the proposed approach.

The demonstration use case (section 3.3) is similarly limited in temporal extent. It is not clear why the analysis covers only approximately 1.5 years of observations. Given the authors' repeated emphasis on the computational efficiency and lightweight nature of the algorithm, a longer time series analysis would appear both feasible and highly beneficial for demonstrating robustness, transferability, and practical value.

In several instances, the Discussion appears to overstate the actual findings of the study, such as the following:

L335–338: Based on the spectral information used, the prototype is capable of detecting floating photosynthetically active material more generally, including floating aquatic vegetation and potentially algal scums, not only water hyacinth. As this, the statement is not fully supported by the actual results presented.

L344–346: The text refers to agreement with observations, yet no actual observations are presented in this section. What is shown is a comparison between map products. Without corresponding field observations or independent validation data, the statement should be rephrased more cautiously.

L365–366: While I understand the authors' concern that cloud-screening procedures applied to Level-2 products may be overly conservative for their intended application, the advantages of using bottom-of-atmosphere (BOA) reflectance generally outweigh the disadvantages associated with TOA spectral quantities. I therefore encourage the authors to perform an exploratory comparison using Sentinel-3 Level-2 products to evaluate the sensitivity of the approach to standard atmospheric correction.

L367–369: The discussion of spectral ratios is not accurate: band ratios – and NDVI is not strictly a simple ratio - do not eliminate the effects of viewing geometry or atmospheric variability. At best, they may partially attenuate such effects under certain conditions. Consequently, VZA- and atmosphere-induced variability cannot be assumed negligible solely because an index-based approach is employed.

L371–374: I do not see a convincing reason why the properties attributed here to TOA radiance/reflectance would not also be met, and generally met more consistently, by BOA reflectance products. Since BOA retrievals aim to remove atmospheric effects and provide physically comparable surface reflectance measurements, the argument presented in favour of TOA data requires further justification.

More generally, a recurring issue throughout the manuscript is that the limitations of the input data and spectral approach are frequently mentioned and, in some cases, explicitly acknowledged, yet their implications are subsequently downplayed when interpreting the results. Given the lack of species-specific discriminatory capability, the use of TOA radiometry, the potential influence of viewing geometry effects, and the limited validation presented, the evidence supporting the performance of the proposed algorithm remains relatively weak.

Technical issues
In addition, I list here some punctual issues of various relevance that, in my opinion, need the authors' attention. These are referenced by manuscript line:
L31: The scientific name of water hyacinth has been changed to Pontederia crassipes few years ago. Better use this accepted name.
L73-74: is this based on microwaves (SAR) fitting the rationale of this work?
L121: what does it mean that NDVI can be biased by very still waters? I cannot understand it.
L141-142: do you mean that L1C data are not gridded originally?
Table 2: class taxonomy is not consistent here: class 1 relates to a density attribute, while class 2 on a confidence attribute. Either use consistent dichotomies (sparse/dense, uncertain/confident or something similar).
Fig. 3: colour scale for panels E and F should be discrete, not continuous.
Fig. 7-8: To provide a clearer and more synthetic overview, they should be merge into a unique figure.
L394-397: The concluding remarks would also benefit from a more balanced perspective. I would suggest avoiding conclusions that may be perceived as overly Eurocentric or technology-centric, particularly when discussing environmental management challenges that are highly dependent on local conditions, resources, and governance frameworks.
Citation: https://doi.org/10.5194/egusphere-2026-403-RC1
- RC2:
  'Reply on RC1', Anonymous Referee #2, 22 Jul 2026
  
  1. The manuscript lacks validation against independent ground observations or
  
  high-resolution UAV/Sentinel-2 derived reference maps. Although the leave-
  
  one-out analysis evaluates interpolation performance, it does not quantify the
  
  accuracy of the water hyacinth detection itself. Please include quantitative
  
  validation metrics (e.g., Overall Accuracy, F1-score, IoU, or Kappa
  
  coefficient) using independent reference data.
  
  2. The NDVI thresholds (NV_low and NV_high) were derived from a single
  
  reference date (1 September 2022). It is unclear whether these thresholds
  
  remain valid under varying seasonal, atmospheric, and illumination
  
  conditions. A sensitivity analysis or multi-season threshold calibration would
  
  improve the robustness of the proposed algorithm.
  3. The methodology relies on Level-1C Top-of-Atmosphere (TOA) radiances
  
  instead of atmospherically corrected surface reflectance. The authors should
  
  justify why atmospheric correction was omitted and discuss the possible
  
  influence of aerosols, haze, and changing atmospheric conditions on NDVI-
  
  derived classifications.
  4. The study would be strengthened by comparing the proposed NDVI-based
  
  approach with alternative vegetation indices (e.g., FAI, NDAVI, WAVI) or
  
  machine learning classifiers reported in previous literature. Such comparisons
  
  would better demonstrate the advantages and limitations of the proposed
  
  algorithm.
  5. Although the gap-filling algorithm successfully increases temporal coverage,
  
  interpolation may smooth rapid spatial changes caused by wind-driven
  
  movement of floating vegetation. The manuscript should quantify the
  
  uncertainty introduced by interpolation and discuss its implications for
  
  operational monitoring during abrupt water hyacinth redistribution events.
  6. The authors state that the algorithm can easily be adapted to other water
  
  bodies; however, no demonstration or discussion is provided for reservoirs
  
  with different morphologies, trophic conditions, or vegetation densities. A
  
  discussion of algorithm transferability and required parameter adjustments
  
  would enhance the paper.
  
  7. Since the algorithm is described as computationally lightweight, the
  
  manuscript should report processing time, computational requirements,
  
  software environment, and scalability for larger study areas. These details are
  
  important for assessing operational implementation.
  8. The moderate spatial resolution (300 m) of Sentinel-3 OLCI may fail to detect
  
  small or fragmented water hyacinth patches and may introduce mixed-pixel
  
  effects near shorelines. The manuscript should include a more detailed
  
  discussion of these limitations and their influence on monitoring accuracy,
  
  particularly for smaller reservoirs.
  9. The authors should justify why Sentinel-3 OLCI was selected instead of
  
  MODIS (MOD09GQ/MYD09GQ) products. MODIS provides near-daily
  
  observations and 250 m NDVI products, which offer slightly finer spatial
  
  resolution than the 300 m OLCI data and have been widely used for vegetation
  
  monitoring. A discussion comparing the advantages and limitations of OLCI
  
  and MODIS (e.g., spatial resolution, revisit frequency, spectral characteristics,
  
  radiometric performance, cloud sensitivity, and suitability for floating water
  
  hyacinth detection) would strengthen the manuscript and better justify the
  
  choice of sensor
  
  Citation: https://doi.org/10.5194/egusphere-2026-403-RC2
  - AC2: 'Authors' reply to referee 2', Marloes Penning de Vries, 21 Aug 2026
    
    We appreciate the review provided by referee 2. We will take some time to revise the manuscript according to the referees’ comments. Please find our replies to each of the referee’s points below.
    1. Validation
    As mentioned in the manuscript, validation of floating vegetation classification is extremely challenging: we have tried to obtain “ground truth” observations by noting down the geo-locations of the outline of vegetation patches by boat; but even on a small water body such as Hartbeespoortdam this is an arduous task, particularly if one is restricted to the overpass time of the satellite. In addition, we were unable to obtain UAV observations for the study area, hence independent validation data were not available. Other satellite data, such as Sentinel-2 MSI data, cannot be considered “independent reference data”, as the detection of hyacinths would proceed via the same NDVI algorithm as for Sentinel-3 OLCI. For this reason, and because the main aim of the manuscript was the presentation of the gap-filling algorithm (rather than an algorithm for the detection of hyacinth), an inter-comparison was not presented. In the revised manuscript, we will add a short section in which we show how the spectral response of floating vegetation, calculated using a physics-based model, depends on fractional coverage of the satellite pixel.
    2. Selected NDVI thresholds
    The referee is right in pointing out that the choice of the selected thresholds is not well substantiated, and we admitted that this is a limitation of the algorithm on line 361 in the manuscript. We will reconsider the strategy of threshold selection for the revised manuscript, using the outcome of the systematic modelling study mentioned above to determine threshold values based on physical properties.
    3. Atmospheric correction
    Although we justified the use of Level-1 TOA radiances instead of Level-2 BOA radiances in lines 365-374 (mainly to avoid the conservative cloud filter applied to level-2 products), we realise that this has likely introduced appreciable errors. For the revised manuscript, we will make use of Level-2 products to determine vegetation cover; this is also mentioned in our reply to referee 1.
    4. Comparison with alternative vegetation indices
    We appreciate the referee’s suggestion to compare the performance of our algorithm with alternative vegetation indices, but believe this to be outside the scope of the current study, as the main aim of the manuscript is the presentation of the gap-filling strategy. Note that without independent reference data, a performance comparison of various vegetation indices is not possible.
    5. Uncertainty of interpolation by gap-filling algorithm
    We have addressed this issue qualitatively in lines 344-346, where we state that “we find that gap-filled maps are in good agreement with the actual observations if there are no sudden changes in WHY cover patterns – as may be expected for an algorithm that relies on continuity.”. It is difficult (and possibly not even very useful) to put a number to the uncertainty caused by sudden changes in wind, water flow, or human actions (e.g., eradication efforts). In lines 346-349 we add that: “A more accurate gap-filling may be achieved by explicitly modelling the evolution and propagation of floating vegetation, e.g. by taking into account drivers affecting the position of WHY (water currents, wind speed and direction) and parameters affecting growth (temperature, nutrient concentration)”, but this development is outside of the scope of the current study.
    6. Evidence for adaptability of algorithm to other regions
    The referee is right in saying that we should back up our claim that the algorithm can be adapted to other water bodies by empirical evidence. This will be included in the revised manuscript.
    7. Evidence for computational efficiency
    As mentioned in our replies to referee 1, we agree that the claimed efficiency should have been substantiated by quantitative evidence. We are not aware of alternative approaches to compare to, but we will make sure that any performance claims in the revised manuscript are backed up by evidence.
    8. Mixed-pixel effects
    The referee is right in addressing mixed-pixel artefacts. These were not discussed sufficiently in the submitted manuscript, but will be included in the revised version by referring to the results from the modeling study mentioned in reply to comments 1 and 2.
    9. Justification of choice for OLCI data
    We will include a rationale for our choice of sensor in the revised manuscript.
    
    Citation: https://doi.org/10.5194/egusphere-2026-403-AC2
- AC1: 'Authors' reply to referee 1', Marloes Penning de Vries, 21 Aug 2026
  
  We would like to start by expressing our deep appreciation of the time and effort that referee 1 has put into the extensive review of our manuscript. The feedback is detailed and specific, formulated in a positive and constructive tone of voice, and it will greatly aid us with the revision of our manuscript. Addressing the referee’s points will require appreciable changes to the algorithm and manuscript, which we have not managed to complete at this stage. For this reason, we here provide some high-level reactions to the main feedback points provided by referee 1, anticipating a more detailed reply upon submission of the revised manuscript.
  1. Mismatch between stated target (daily monitoring of water hyacinth) and choice of observation parameters
  
  The observation that NDVI is not suited to distinguish water hyacinth from other emergent and floating vegetation is astute, and we should not have implied that the algorithm was able to do so. The main aim of the manuscript is the presentation of the gap-filling algorithm, therefore we did not attempt to make the algorithm more selective for either floating/emergent vegetation or even water hyacinth. In the revised manuscript, we will not make the distinction and refer to the target species (e.g., “floating and emergent vegetation”) in a consistent manner; this will also be reflected in a new manuscript title.
  
  2. Spatial and temporal limitations of demonstration data set
  
  The data shown in the manuscript were limited to the site and timing of a project of which the findings are described in the manuscript. In contrast to the referee, we find that the 18-month timeseries provides sufficient data for the purpose of the study, particularly because there is a lot of variation in spatial vegetation patterns visible on account of seasonal and other effects. Nevertheless, we will strive to include the complete OLCI timeseries, and possibly even include historical data sets (see next point). We agree that limiting the study to a single reservoir is at odds with the claim that the algorithm is applicable globally. We will take this into account while revising the manuscript.
  3. Use of TOA radiance as input for the algorithm
  
  The main point of critique is the use of Level-1 TOA radiances to calculate NDVI instead of using Level-2 BOA radiances. The main reason to do so, as mentioned in the manuscript, is the conservative cloud-screening applied to Level-2 products. In addition, OLCI Level-2 processing is performed by two different institutions, depending on if the scenes are classified as “land” (ESA) or “water” (EUMETSAT). For our complicated case of an inland waterbody, we find that whereas Level-2 data from parts of the reservoir covered by floating vegetation are available in the “land” product, they are not available in the “water” product. The reverse is the case for unvegetated parts of the reservoir. For this reason, we used Level-1 data, which is available over both land and water, but did not apply an atmospheric correction. We realise that the assumption that errors introduced by not applying an atmospheric correction would vanish by the use of a spectral difference index is not justified. The focus on the main aim of the manuscript, introduction of the gap-filling strategy, led us to ignore the inaccuracies of the chosen data set. For the revised manuscript, we will re-do the analysis with Level-2 data, applying an atmospheric correction to Level-1 data ourselves if necessary.
  
  In addition, we appreciate and will consider the suggestion to merge OLCI data with historical data sets from MODIS and/or MERIS.
  4. Introduction and review of literature
  
  For various reasons, a lot of time has elapsed since the inception of the manuscript, and the referee is probably right in pointing out that literature review is no longer up to date – an issue that will be amended in the next version of the manuscript. Nevertheless, we would like to point out that, although there are a number publications on water hyacinth monitoring from satellite, none of these were based on consistent, frequently sampled data sets. This is described on lines 78-83 of the submitted manuscript, where we argue that the time resolution of data used in previous studies is “not sufficient to track the rapid movements of [water hyacinth]” (lines 81-82).
  5. Computational efficiency of the algorithm
  
  The referee is right in saying that the claimed efficiency should have been substantiated by quantitative evidence. We are not aware of alternative approaches to compare to, but we will make sure that any performance claims in the revised manuscript are backed up by evidence.
  
  Citation: https://doi.org/10.5194/egusphere-2026-403-AC1

Marloes Penning de Vries, Cletah Shoko, Timothy Dube, Kgabo Humphrey Thamaga, and Suhyb Salama

Viewed

Total article views: 1,315 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
592	666	57	1,315	180	222

HTML: 592
PDF: 666
XML: 57
Total: 1,315
BibTeX: 180
EndNote: 222

Views and downloads (calculated since 10 Feb 2026)

Month	HTML	PDF	XML	Total
Feb 2026	313	122	36	471
Mar 2026	98	200	8	306
Apr 2026	46	94	4	144
May 2026	37	49	1	87
Jun 2026	7	6	1	14
Jul 2026	50	179	3	232
Aug 2026	24	11	2	37
Sep 2026	17	5	2	24

Cumulative views and downloads (calculated since 10 Feb 2026)

Month	HTML	PDF	XML	Total
Feb 2026	313	122	36	471
Mar 2026	98	200	8	306
Apr 2026	46	94	4	144
May 2026	37	49	1	87
Jun 2026	7	6	1	14
Jul 2026	50	179	3	232
Aug 2026	24	11	2	37
Sep 2026	17	5	2	24

Viewed (geographical distribution)

Total article views: 1,132 (including HTML, PDF, and XML) Thereof 1,132 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 06 Sep 2026

Short summary

Water hyacinth (WHY) are one of the most costly aquatic invasive species. Regular monitoring of WHY would the management of infested waterboddies. We present an algorithm to derive daily maps of WHY from satellites, featuring an interpolation scheme, which led to 55% more daily WHY maps of Hartbeespoortdam Reservoir. The algorithm is light-weight and can be run on a cloud-computing platform. This allows applications in (sub-)tropical regions with limited access to computing power and/or funding.


Total:	0
HTML:	0
PDF:	0
XML:	0