PaperlayerView exact revision on arXiv

Source-linked AI summary

AquaCubeAI-Powered Monitoring Turbidity on-board {\Phi}sat-2

Pietro Di Stasio, Francesca Razzano, Elisa Liparulo, Gabriele Meoni, Nicolas Longépé, Deodato Tapete, Paolo Gamba, Gilda Schirinzi, Silvia Liberata Ullo

arXiv:2609.12744v1cs.CV

TL;DR

Ground-based satellite workflows can delay coastal water-quality monitoring, while few onboard approaches directly regress water-quality indicators. AquaCubeAI pairs simulated Φsat-2 imagery with CMEMS turbidity references and applies a compact onboard regressor with dense spatial prediction; the authors report accurate estimation across heterogeneous coastal environments, strongest in the low-to-moderate regime.

  • Problem

    Ground-based satellite workflows introduce latency, and onboard frameworks for direct regression of coastal water-quality parameters from multispectral observations are scarce.

    Source: §I · pp. 1–2

  • Method

    AquaCubeAI trains a compact turbidity regressor on simulated Φsat-2 multispectral patches aligned with CMEMS labels and reformulates it for dense spatial prediction.

    Source: §I · p. 2, §IV · p. 7, A. Data Representation · p. 7

  • Results

    Accurate estimation is reported across heterogeneous coastal environments, strongest in the low-to-moderate regime and reduced in the high-turbidity tail; an external hold-out suggests informative predictions on an unseen date.

    Source: A. Architecture Comparison and… · p. 9, §VI · p. 13

  • Takeaways & Limitations

    The spatial-prediction formulation generates turbidity maps and threshold-based anomaly masks to support onboard decision logic and selective data downlink.

    Source: §VI · p. 12

    Training and evaluation use only simulated mission-like Φsat-2 data, leaving a possible residual gap with real in-orbit observations.

    Source: §VI · p. 13

Abstract

from arXiv · show

Timely monitoring of coastal water quality is critical for environmental protection, yet conventional satellite workflows rely on downlink and ground processing, introducing latency that can limit responsiveness to rapidly evolving turbidity events. To address this limitation, we propose AquaCubeAI, a lightweight machine-learning approach for onboard estimation of coastal water turbidity from {\Phi}sat-2 multispectral imagery. By shifting inference from the ground segment to the satellite, AquaCubeAI aims to enable lower-latency, more responsive, and more operationally useful turbidity monitoring under the strict compute and bandwidth constraints of spaceborne platforms. The model is trained on simulated {\Phi}sat-2 acquisitions spatially aligned with Copernicus Marine Service (CMEMS) High-Resolution Ocean Color (HR-OC) turbidity products over selected localized coastal sites spanning four European marine macro-regions. To provide a realistic evaluation of generalization in the presence of spatial correlation, we adopt a spatial block splitting protocol that mitigates data leakage between training and evaluation subsets. The main contributions of this work are: (i) a scalable dataset generation pipeline pairing simulated {\Phi}sat-2 multispectral patches with CMEMS HR-OC turbidity labels across selected localized European coastal sites; (ii) a compact Multi-Layer Perceptron (MLP)-based turbidity regressor trained under a leakage-aware geospatial split and tailored to embedded constraints; and (iii) a reformulation for dense spatial prediction via parameter sharing, enabling turbidity mapping and simple threshold-based anomaly masks for onboard decision logic. Embedded deployment on an Intel Myriad Vision Processing Unit (VPU) further confirms the feasibility of low-power hardware and supports low-latency inference from multispectral inputs.

I. INTRODUCTION

Satellite observations and AI offer scalable water-quality monitoring, but ground-based processing and limited in situ coverage constrain responsiveness. AquaCubeAI targets the gap in onboard regression of coastal water-quality indicators with a compact turbidity model and spatial prediction framework.

I. INTRODUCTION

  • Point-based in situ monitoring is spatially constrained, while conventional satellite AI workflows rely on ground processing that can delay responses to changing coastal conditions.Onboard processing can prioritize relevant information for downlink, but few works have addressed onboard marine water-quality monitoring.Source: §I · pp. 1–2
  • AquaCubeAI estimates turbidity from simulated Φsat-2 multispectral imagery and generates spatial anomaly masks within an onboard-processing framework.Its contributions include a CMEMS HR-OC paired dataset, a compact MLP regressor trained with a leakage-aware geospatial split, dense prediction through parameter sharing, and Intel Myriad 2 VPU deployment.Source: §I · p. 2

II. RELATED WORKS ON WATER QUALITY APPLICATIONS WITH RS AND AI TECHNIQUES

  • In situ systems remain a water-quality reference, while remote sensing and geospatial AI support repeated wide-area observations and operational decision-support tools.These approaches still face atmospheric, surface-condition, sensor, and calibration challenges, and many rely on ground-based processing.Source: §II · pp. 2–3
  • Existing satellite and AI water-quality approaches generally support mapping but remain oriented toward offline or near-offline processing, while onboard marine work has focused largely on detection.Prior onboard studies address harmful algal blooms, marine anomalies, and contaminants, rather than direct regression of coastal water-quality indicators.Source: §II · p. 3
  • This work extends prior onboard regression research with a scalable CMEMS HR-OC dataset, multi-region pilot sites, spatial block splitting, and a dense turbidity-prediction formulation.The study adds geographic diversity beyond a single-site configuration without claiming full representativeness of the marine domains.Source: §II · p. 3

III. DATASET

The dataset pairs simulated Φsat-2 multispectral imagery with CMEMS HR-OC turbidity references for localized coastal areas. Its four-region pilot spans selected sites and months, with patch-based preprocessing designed for onboard constraints.

III. DATASET

  • The dataset focuses on coastal and near-shore waters, where turbidity has strong spatial heterogeneity and seasonal variability.Its pipeline pairs multispectral observations with consistent bio-geochemical reference information.Source: §III · p. 3
  • A single open-access CMEMS source provides the reference water-quality information, supporting reproducibility and extension to additional geographic areas.Fig. 1 locates the selected reference sites within CMEMS coverage; those locations define the Φsat-2 simulation study areas.Source: §III · p. 3, Figure 1 · p. 4
  • Simulated Φsat-2 imagery from the official mission simulator is paired with CMEMS HR-OC turbidity labels across four European marine macro-regions.Localized AoIs span the Mediterranean Sea, Black Sea, Iberian coastal region, and North-West Shelf; the pilot is not intended to represent each domain's full range of optical water types or environmental conditions.Source: §III · p. 3

A. Copernicus Marine Service

The study uses CMEMS HR-OC turbidity as reference labels for simulated Φsat-2 imagery over selected coastal sites in four European marine regions. The simulator approximates payload characteristics, but not Φsat-2 tasking or revisit constraints.

A. Copernicus Marine Service

  • CMEMS HR-OC provides 100 m coastal turbidity labels in FNU, derived from Sentinel-2 MSI observations and covering a coastal strip extending approximately 20 km offshore.Turbidity is the primary variable of interest in the product suite, alongside suspended particulate matter and chlorophyll-a.Source: A. Copernicus Marine Service · p. 4
  • The 2023 dataset samples selected dates across seven months and four European marine regions, while a 2024 acquisition set is reserved for interannual generalization assessment.The 2024 samples are excluded from training, validation, and model selection.Source: A. Copernicus Marine Service · p. 4
  • The study evaluates simulated rather than real in-orbit Φsat-2 observations because aligned acquisitions and turbidity references across the selected sites were not available within its scope.Evaluation on real in-orbit observations is left for future work.Source: B. Φsat-2 Data Collection · p. 5
  • The simulator uses Sentinel-2 bands corresponding to seven Φsat-2 multispectral channels, while the additional panchromatic band is synthesized from Sentinel-2 bands B2–B6.The Φsat-2 instrument has eight bands: seven multispectral channels and one panchromatic band.Source: B. Φsat-2 Data Collection · p. 5
  • Simulated scenes represent diverse coastal environments and optical conditions, while their acquisition dates follow selected Sentinel-2 observations rather than Φsat-2 tasking.The simulator does not reproduce Φsat-2 tasking or revisit constraints; its average global access period is approximately 15 days under quasi-nadir conditions.Source: B. Φsat-2 Data Collection · p. 5, Figure 2 · p. 6
  • The official simulator adapts Sentinel-2 observations to approximate Φsat-2 spectral, radiometric, spatial, and sensor characteristics, including resampling and sensor-effect modeling.The simulated Level-1C products preserve source acquisition content; these uncorrected inputs are selected for onboard feasibility, with prior comparative work reporting comparable performance to atmospherically corrected inputs in the considered coastal setting.Source: B. Φsat-2 Data Collection · pp. 5–6

C. Dataset Pre-Processing

The pre-processing pipeline aligns HR-OC turbidity observations with simulated Φsat-2 imagery and extracts compact, georeferenced multispectral patches for supervised learning.

C. Dataset Pre-Processing

  • The pipeline harmonizes HR-OC observations and simulated Φsat-2 imagery into a unified dataset for onboard regression.It is designed to preserve spatial correspondence and physical consistency across the data sources.Source: C. Dataset Pre-Processing · p. 6
  • After filtering invalid or out-of-range labels, the workflow retains HR-OC samples within the corresponding simulated-scene footprint that allow a valid patch to be extracted.Turbidity labels above 140 FNU are removed; preserved cloud and shadow masks are not used to reject individual patches.Source: C. Dataset Pre-Processing · p. 6
  • HR-OC coordinates are transformed from WGS84 into the image reference system and mapped to pixel indices when needed.Source: C. Dataset Pre-Processing · p. 6
  • Each retained observation is paired with a 20 × 20 multispectral patch covering approximately 100 × 100 m, close to the HR-OC product’s nominal 100 m resolution.Patches are selected only at locations with valid CMEMS turbidity values; no dedicated water/non-water segmentation is applied.Source: C. Dataset Pre-Processing · p. 6

D. Dataset Analysis and Statistical Characterization

The dataset analysis characterizes turbidity distribution across the selected coastal conditions, including its range, seasonal variability, and regional heterogeneity.

D. Dataset Analysis and Statistical Characterization

  • The dataset analysis is intended to assess whether samples represent heterogeneous coastal conditions for evaluating spatial and seasonal generalization.Source: D. Dataset Analysis and Statistical… · p. 6
  • The 0–40 FNU interval contains the most observations, while values above 100 FNU form a less frequent upper tail within the dataset’s 0–140 FNU range.The paper defines low-to-moderate turbidity as 0–100 FNU and high turbidity as greater than 100 through 140 FNU, explicitly as dataset-specific intervals rather than universal classes.Source: D. Dataset Analysis and Statistical… · pp. 6–7
  • Monthly turbidity distributions show seasonal differences in central tendency and spread, and regional distributions reflect distinct turbidity regimes.Fig. 3 summarizes monthly variability, while Fig. 4 reports turbidity distributions aggregated by region.Source: D. Dataset Analysis and Statistical… · p. 7, Figure 3 · p. 7, Figure 4 · p. 7
  • The adopted representation preserves mean spectral information over each reference observation’s local support in a compact onboard-compatible input.Source: D. Dataset Analysis and Statistical… · p. 7

IV. METHODOLOGY

The methodology pairs simulated Φsat-2 multispectral imagery with CMEMS turbidity references and trains an onboard-oriented regressor under a leakage-aware geospatial split. Channel-wise patch averages provide compact inputs, while parameter sharing supports dense spatial outputs.

IV. METHODOLOGY

  • The workflow pairs simulated Φsat-2 patches with CMEMS HR-OC turbidity references and uses a geospatially robust split before producing spatial maps and threshold-based anomaly masks.Patch-wise maps can be mosaicked into larger-area products and prioritized for downlink.Source: §IV · p. 7

A. Data Representation

  • Each CMEMS observation is paired with a colocated simulated patch, and turbidity is estimated by regression from the patch's multispectral content.This provides consistent spatial support between inputs and labels.Source: A. Data Representation · p. 7
  • Channel-wise spatial averaging reduces each patch to an eight-band descriptor representing its mean spectral signature over the CMEMS-aligned area.The compact representation is intended to be robust to small geolocation uncertainties.Source: A. Data Representation · p. 7
  • Targets are min–max normalized using training-subset parameters, which are reused for evaluation and conversion of predictions back to FNU.The design targets consistent spatial support, realistic generalization assessment, and low computational complexity.Source: A. Data Representation · p. 7, §IV · p. 7

B. Training Strategy

Training uses spatial blocks to mitigate leakage between correlated remote-sensing samples while preserving turbidity balance across subsets. A compact MLP is optimized with Adam and Huber loss, with validation-based early stopping and stability controls.

B. Training Strategy

  • Train, validation, and test subsets have similar turbidity distributions, preserving overall turbidity balance while maintaining spatial separation.Fig. 6 presents the distributions obtained with the geospatial split.Source: Figure 6 · p. 8
  • Spatial block splitting assigns all samples in each 15 km UTM grid cell to one subset, mitigating short-range spatial dependence while targeting a 60%/20%/20% partition.The protocol does not impose a temporal constraint, so potential temporal leakage is not fully eliminated.Source: B. Training Strategy · pp. 7–8

C. Regression Model and Optimization

  • The default regressor is an MLP with three hidden layers of widths (512, 512, 256) and ReLU activations.Configurations are varied to balance predictive capability and computational cost.Source: C. Regression Model and Optimization · p. 8
  • Adam optimization uses validation-driven early stopping and Huber loss with δ = 1.0 in normalized space to reduce sensitivity to occasional large errors.Gradient clipping and deterministic seeding support stability and repeatability.Source: C. Regression Model and Optimization · p. 8

D. Model Selection

Model selection balances validation accuracy with parameter count, and the selected regressor is reformulated for dense spatial prediction through shared 1 × 1 convolutional layers.

D. Model Selection

  • Candidate MLPs are compared using validation RMSE/MAE and parameter count, favoring compact architectures that preserve accuracy.The selected configuration is reported in the results section.Source: D. Model Selection · pp. 8–9

E. From Regression to Spatial Prediction

  • The trained MLP is reformulated as shared 1 × 1 convolutional layers, applying the learned spectral mapping at each pixel to produce a dense turbidity map without extra trainable parameters or retraining.The CNN-style reformulation exploits VPU acceleration; thresholding predicted turbidity yields anomaly masks.Source: E. From Regression to Spatial Prediction · p. 9

V. RESULTS

On the robust geospatial split, AquaCubeAI achieves strong agreement across the full turbidity range, while errors rise toward the less represented high-turbidity tail. The selected model also reduces parameter count relative to the previous baseline.

A. Architecture Comparison and Selected Model

B. Performance Across Turbidity Ranges and Baseline Comparison

  • R2 reaches 0.995 across the complete 0–140 FNU interval, although errors increase as rarer, more heterogeneous high-turbidity conditions are included.Performance is strongest in the most represented 0–40 FNU range, and extreme conditions remain more challenging.Source: B. Performance Across Turbidity Ranges… · pp. 9–10, Table VI · p. 10
  • RMSE = 1.39 FNU and MAE = 0.86 FNU on the current validation subset in the 0–40 FNU comparison regime.The operating range matches the previous study's reference range, although experimental conditions are not fully matched.Source: B. Performance Across Turbidity Ranges… · p. 10, Table VII · p. 10
  • 401,417 parameters in AquaCubeAI versus 818,349 in the baseline, reducing model size for onboard memory and compute.The comparison is against the previous work, whose evaluation used a local case study; the present pilot covers selected localized European coastal sites under a stricter spatial split.Source: B. Performance Across Turbidity Ranges… · p. 10

C. External Hold-Out Evaluation

An independent North-West Shelf acquisition excluded from all development subsets was used for external hold-out evaluation. Performance was stable across training seeds, with residual dispersion concentrated at higher turbidity levels.

C. External Hold-Out Evaluation

  • The independent North-West Shelf acquisition was excluded from training, validation, and test, and its predictions clustered around the 1:1 line.Residual dispersion was mainly observed at higher turbidity levels, where fewer samples were available and local variability was stronger.Source: C. External Hold-Out Evaluation · p. 10, Figure 9 · p. 11
  • The external 2024 hold-out set yielded an RMSE of 1.59±0.15 FNU, an MAE of 0.95±0.14 FNU, and a bias of 0.19±0.23 FNU.Across five training seeds, the internal test set yielded an RMSE of 2.70 ± 0.17 FNU, an MAE of 1.25 ± 0.13 FNU, and a bias of −0.55 ± 0.23 FNU.Source: C. External Hold-Out Evaluation · p. 11
  • Across five independent random seeds, the limited standard deviations indicate that reported performance was stable with respect to training seed.The split, preprocessing, architecture, and hyperparameters were held fixed across runs.Source: C. External Hold-Out Evaluation · p. 11

D. Global Mapping and Spatial Products

The patch-averaged MLP outperformed a lightweight CNN across validation, test, and external hold-out splits, and was retained for dense spatial prediction. The spatial model produced turbidity maps and threshold-based anomaly masks, with embedded inference demonstrated on Myriad 2.

D. Global Mapping and Spatial Products

  • The patch-averaged MLP achieved lower RMSE than the CNN on validation (2.78 versus 3.28 FNU), test (2.47 versus 3.51 FNU), and external hold-out (1.34 versus 1.43 FNU) splits.The MLP also achieved lower MAE and higher R2 across all three splits; the CNN used two 3×3 convolutional layers and global average pooling on full 20×20×8 patches.Source: D. Global Mapping and Spatial Products · p. 11
  • The MLP was retained because it provided higher predictive accuracy and a computationally simpler representation for onboard inference.Both models used the same geospatial split, target normalization, and evaluation protocol.Source: D. Global Mapping and Spatial Products · p. 11
  • Applying the spatial predictor patch by patch and mosaicking outputs produced a dense turbidity map, with a TUR > 100 FNU threshold mask shown for the Iberian AoI.The example compares the model map with CMEMS reference turbidity for 2024–12–23.Source: D. Global Mapping and Spatial Products · p. 11

E. Inference on Myriad 2

  • Myriad 2 inference averaged 8.51 ms per patch (117.49 inferences/s), while processing retained valid patches from a 4096 × 4096 scene took 215.6 s.The current pipeline selects fixed-size patches using valid HR-OC turbidity locations; automatic screening to discard non-water areas is not implemented.Source: E. Inference on Myriad 2 · p. 11

VI. DISCUSSION AND CONCLUSIONS

AquaCubeAI combines simulated Φsat-2 imagery and CMEMS HR-OC references with a lightweight regressor reformulated for dense turbidity maps and threshold-based anomaly masks. The discussion reports promising geospatial and embedded results while noting limits tied to simulated inputs, reference uncertainty, and validation scope.

VI. DISCUSSION AND CONCLUSIONS

  • AquaCubeAI combines a scalable dataset pipeline, geographically diverse coastal domains, and a lightweight embedded model; its spatial reformulation supports turbidity maps and threshold-based anomaly masks for onboard decision logic.The Iberian AoI example presents CMEMS reference turbidity, the AquaCubeAI map, and an anomaly mask.Source: §VI · pp. 11–12, Figure 10 · p. 12
  • The patch-averaging strategy trades away sub-patch patterns, while the tested lightweight CNN did not improve accuracy; the 1 × 1 convolutional reformulation increases output resolution but does not recover spatial context.Source: §VI · p. 12
  • Level-1C inputs may be affected by atmospheric conditions underrepresented in training, and CMEMS-derived rather than in situ targets may contribute residual reference uncertainty to the error budget.Source: §VI · pp. 12–13
  • Training and evaluation used simulated mission-like Φsat-2 data exclusively, leaving a possible residual gap with real in-orbit observations from calibration, noise, misregistration, atmospheric variability, and mission effects.Source: §VI · p. 13
  • The geospatial-split results indicate strongest performance in low-to-moderate turbidity and reduced accuracy in the high-turbidity tail; temporal dependence across internal subsets cannot be fully excluded because the split was not date-constrained.The external hold-out provides a first evaluation on an unseen acquisition date; the authors call for additional validation under distribution shifts.Source: §VI · p. 13
  • Embedded tests on an Intel Myriad VPU support execution on low-power hardware with latencies compatible with near-real-time inference.Source: §VI · p. 13
  • Future work targets validation across more regions and seasons, evaluation on real in-orbit Φsat-2 observations, and improved robustness and sampling for rare high-turbidity extremes.Source: §VI · p. 13
Loading 2609.12744v1…