OceanTensorCube and ST-MMT: A Multimodal Data and Transformer System for Seven-Day Forecasting of Pyropia Bleaching

Gwangjin Kim; Cheolyoung Jang; Yosep Jo

cs.LGOct 19, 2026

OceanTensorCube and ST-MMT: A Multimodal Data and Transformer System for Seven-Day Forecasting of Pyropia Bleaching

Authors: Gwangjin Kim, Cheolyoung Jang, Yosep Jo

Organizations: NCS Instructor, Smart Human Resources Development Institute, Republic of Korea · NH Networks LLC, Republic of Korea · DreamTech Co., Ltd., Republic of Korea

Abstract

Bleaching of cultured Pyropia (laver) causes losses in Korean seaweed farming. We present a system that forecasts bleaching seven days ahead for farm clusters. OceanTensorCube aligns 32 channels from public marine, meteorological, ocean-model and satellite sources on a 30 m grid, and ST-MMT, a 9.2-million-parameter spatiotemporal transformer with hub-style multi-head cross-attention between variable groups, maps a 14-day input window to a seven-day-ahead bleaching probability. On the validation set, a six-seed ensemble detects every reported positive at a false-alarm rate of at most 3.5%, an upper bound because the report-based labels are positive-unlabeled. A same-date control on separately trained models indicates that the architecture discriminates between clusters when the calendar is held fixed. We also report the system's weaknesses: skill falls steeply as the quarterly nutrient survey ages (ROC-AUC 0.94 at 30 days, 0.53 at 129 days), and a single LSTM on cluster-average series outperforms a single ST-MMT, with only the ensemble leading at the operating point. We describe the labeling procedure, the evaluation protocol and the operational gate that withholds forecasts when inputs are stale. The forecasting service is operated at https://oceantensor.ai.kr.

Figures & tables

Panel of Figure 1 — Figure 1: Laver-farming grounds photographed in connection with the 2022 Haenam bleaching case. Photographs courtesy of the National Institute of Fisheries Science (NIFS).

Source	Variables (channels)	Type	Cadence
KOEM monitoring network	SST, DIN, DIP, SiO 2 , salinity, chlorophyll, DO, pH, NO 3 , NH 4 , suspended solids (11)	point stations	quarterly surveys
KMA ASOS	precipitation, solar radiation, wind speed, wind direction (sin, cos), air temperature (6)	point stations	hourly $\rightarrow$ daily
HYCOM / KHOA; ocean model	surface current $u,v$ ; mixed-layer depth (3)	model grid / stations	sub-daily $\rightarrow$ daily
Sentinel-2 L2A	near-infrared index (1)	30 m imagery	8-day composite
Derived	N:P ratio, SST anomaly, 7- and 30-day means, SST gradient, current speed, days since rain, NIR daily change (9)	computed	as parents
Calendar	month sine and cosine (2)	static	monthly

Table 1: Data sources in the OceanTensorCube. Cadence is the native update interval of the variables as used.

Panel of Figure 2 — Figure 2: The OceanTensorCube pipeline and serving path. Raw records (Bronze) are aligned to a 30 m tile grid (Silver), forward-filled and cropped to $64\times 64$ training patches (Gold). The six-seed ensemble writes a forecast pack that the web service reads through a three-stage output gate (Section 8.6 ); the service never runs the network itself.

ch	Name	Source	Unit	ch	Name	Source	Unit
0	SST	KOEM	∘ C	16	Wind speed	KMA	m/s
1	DIN	KOEM	$\mu$ mol/L	17	Wind direction (sin)	KMA	–
2	DIP	KOEM	$\mu$ mol/L	18	Wind direction (cos)	KMA	–
3	SiO 2	KOEM	$\mu$ mol/L	19	Air temperature	KMA	∘ C
4	DIN/DIP ratio	derived	ratio	20	pH	KOEM	–
5	Salinity	KOEM	psu	21	NO 3	KOEM	$\mu$ mol/L

Table 2: Channel definitions. Source is the source actually used by the extraction code; where it differs from the original channel master (turbidity, solar-radiation units, currents, mixed-layer depth), the code is authoritative.

Figure 3: Single-channel AUC of the 14-day window mean on the validation set (2,792 samples, 176 positives), coloured by information source. The calendar and season-correlated channels rank near the top, which motivates the same-date control in Section 7.3 .

Factor	Channel	Condition $\phi_{k}$	$w_{k}$
DIN deficit	ch1	below the regional lower quartile, exponentially weighted over 10 days	30
DIP deficit	ch2	below the regional lower quartile, exponentially weighted over 10 days	15
N:P imbalance	ch4	$<10$ (nitrogen-limited) or $>30$ (phosphorus-limited)	15
SST anomaly	ch12	outside $\pm 2\,^{\circ}$ C of normal during the growing season	15
Short-term chlorophyll drop	ch28	7-day mean at least 30% below the 14-day mean	15
Rain deficit	ch14	more than 14 days since rain	10

Table 3: BLC six-factor weak-label score. Thresholds are relative to the regional normal; the weights are initial values intended to be tuned by ablation.

Figure 4: BLC time rules around a reported event: the time-decay factor

\delta(t)

that multiplies the normalized six-factor score inside the

\pm 14

-day window, the confidence schedule

c(t)

, and the cut-offs that map the soft label

y

to severity grades 0–3.

All samples			Validation (8 clusters $\times$ 2 years)
Period	Positive	Negative	Year	Total	Positive	Negative
2024, Jan–Dec	0	18,144	2024	1,568	0	1,568
2025, Jan–May	0	12,052	2025	1,224	176	1,048
2025, 1–20 Oct (2 clusters)	40	0
2025, 1–17 Nov (20 clusters)	340	0
Total	380	30,196	Total	2,792	176	2,616

Table 4: Composition of the labeled samples (left) and of the validation set (right).

Panel of Figure 5 — Figure 5: ST-MMT architecture ( STMMTCrossAttnV3 ). A joint spatiotemporal encoder over all channels is refined by modality-specific encoders fused through two rounds of multi-head cross-attention with the in-situ ocean group as the query hub. The refinement enters through a learnable gate $\alpha$ . The head outputs a $16\times 16$ map of class probabilities; the forecast pack stores its spatial mean and maximum.

Input	$14\times 32\times 64\times 64$	Hidden size $d$	256
Horizon	7 days	Attention heads	8
Patch size	$4\times 4$ (256 tokens)	Joint / group ST blocks	4 / 2
Groups (A/B/C)	20 / 7 / 3 channels (+2 calendar)	Cross-attention	A $\leftarrow$ B, then $\cdot\leftarrow$ C
Loss	focal, $\gamma=2$ , class weights 0.5/1.5	Fusion	$\mathrm{LN}(\mathbf{F}_{\text{joint}}+\alpha\mathbf{R})$
Ensemble	6 seeds, mean probability	Parameters	9,218,179

Table 5: ST-MMT configuration as deployed ( xattn_v3_focal_7day ).

Metric	Value		Predicted +	Predicted $-$
ROC-AUC	0.9940	Reported positive	176	0
PR-AUC (average precision)	0.9097	Negative (2024)	92	1,476
Recall at $\theta^{\star}$	100.0%	Negative (2025)	0	1,048
Precision at $\theta^{\star}$	65.7%	Negative, total	92	2,524
False-alarm rate at $\theta^{\star}$ a	3.5%
F1 at $\theta^{\star}$	0.7928

Table 6: Validation accuracy of the six-seed ST-MMT ensemble (left) and the confusion matrix at the operating threshold

\theta^{\star}=0.0940

(right).

Figure 6: Validation samples by ensemble probability band. Boundaries are the recall-100% threshold (0.0940) and the median of positive probabilities (0.1938); the early-warning and warning grades share the middle band (split at 0.1120).

Replicate a	Within-date AUC	Bootstrap lower bound
1	0.9778	0.9466
2	0.9763	0.9429
3	0.9857	0.9621

Table 7: Same-date control: within-date AUC among 95 clusters on November 2025 dates (out-of-fold predictions of the cross-validated control training), where calendar features are identical for all clusters.

Model	Parameters	ROC-AUC	PR-AUC	Precision	False alarm	MCC
XGBoost (3 seeds)	– a	0.9150	0.6288	14.62%	39.67%	0.2970
VGG16 (3 seeds) b	39,913,602	0.9810	0.7704	41.49%	36.44%	0.7505
LSTM (3 seeds)	823,810	0.9953	0.9487	60.15%	4.74%	0.7550
ST-MMT (6 single seeds)	9,218,179	0.9907	0.8583	48.71%	7.67%	0.6681
ST-MMT (6-seed ensemble) c	$6\times$ 9,218,179	0.9940	0.9096	65.67%	3.52%	0.7960

Table 8: Benchmark at each model’s own recall-100% threshold. Single-model rows are means over seeds; the ensemble row is computed from averaged probabilities.

Figure 7: Matthews correlation coefficient at each model’s recall-100% threshold (benchmark of 7 October 2026). Single-model bars are means over seeds; the VGG16 bar excludes one seed whose threshold collapsed to zero.

Condition	ROC-AUC	Detection	False alarm	False alarm 2024	Verdict
Original	0.9940	100.0%	3.5%	5.9%	reference
Nine channels flattened in time	0.9943	100.0%	3.9%	6.4%	usable
Nine channels set to zero	0.9778	88.6%	3.1%	5.2%	not usable

Table 9: Experiment A: effect of degrading nine channels (ch10–13, 24, 27–30). Pre-registered criterion: detection 100% and AUC drop

<0.02

Check	Previous pack	Regenerated pack
Share severe, September 2024	100%	0.0%
Share severe, October 2024	92.5%	0.0%
Share severe, November 2025	10.9%	43.8%
Share of “early” grade	0.0% (all 17 months)	9.9%
Range across clusters, 1 Nov 2025	0.013	0.509

Table 10: Forecast pack before and after regeneration with the ensemble.

Source	Latest nutrient value	Age (days)	Note
KOEM monitoring network	28 May 2026	132	August survey not yet published
NIFS serial oceanographic stations	16 Dec 2025	295	no 2026 nutrient records
NIFS fishing-ground stations	5 Nov 2025	336	sparse responses
NIFS real-time stations	–	–	temperature only
CMEMS biogeochemistry (model)	5 Oct 2026	2	daily; not used in training

Table 11: Most recent nutrient observation per source on 7 October 2026.

Figure 8: Experiment B: ROC-AUC (left) and false-alarm rate (right) when the eleven KOEM channels are frozen

N

days before the input window. At 129 days, close to the 132-day age of the freshest nutrient value on 7 October 2026, discrimination is close to chance.

$N$ (days)	Samples	ROC-AUC	Detection	False alarm	False alarm 2024
0	2,792	0.9940	100.0%	3.5%	5.9%
30	2,537	0.9405	97.7%	12.1%	19.7%
60	2,087	0.8333	67.6%	24.4%	37.6%
90	1,637	0.6778	87.5%	42.7%	60.6%
129	1,052	0.5267	–	67.4%	78.0%
180	820	0.3964	–	84.9%	84.9%

Table 12: Experiment B in detail. Detection rates for 129 and 180 days are omitted because the evaluable sample changes composition.

Statistic	Value
NO 2 share, median / mean / 95th percentile	5.26% / 8.07% / 23.84%
Median relative error, sum of species vs. reported DIN	0.0000
Samples with relative error $>10\%$	1,278 (5.2%)
Correction factor $1/(1-f)$ for NO 3 +NH 4 (median / 95th percentile)	1.055 / 1.313

Table 13: Share of nitrite in DIN (NIFS fishing-ground samples,

n=24{,}513

Figure 9: DIN prediction error of persistence (forward fill) and leave-one-year-out climatology for consecutive KOEM observations, grouped by the gap between observations. Labels give the relative reduction in MAE achieved by climatology.

Group	Channel	2024	2025
Same survey (cannot fill the gap)	NO 3 (ch21)	0.945	0.894
	DIP (ch2)	0.800	0.809
	SiO 2 (ch3)	0.710	0.842
Daily (candidates)	Air temperature (ch19)	$-0.501$	$-0.089$
	Month cosine (ch26)	0.461	0.168
	Month sine (ch25)	0.345	0.046

Table 14: Change correlation with DIN for the leading candidates. The daily candidates that ranked highest in 2024 are seasonal variables whose correlation collapses in 2025.

Stage	Rule	Default
1. Freshness	block if the newest nutrient observation is older than the limit	60 days
2. Data quality	require declared valid channels; otherwise mark the forecast “unknown”	–
3. Persistence	raise a grade after consecutive exceedances, lower it after consecutive clear days	2 / 5 / 7 days

Table 15: Harness Gate applied to every forecast before display.

Configuration	ROC-AUC	F1 (recall 100%)	Nov./summer ratio	Alarm days 2024
Original patches (22 valid channels)	0.9825	0.7756	–	–
Moving averages and currents restored (29)	$0.9922\pm 0.0053$	$0.8656\pm 0.0806$	1.41	132
Empty clusters excluded, SST anomaly and MLD	$0.9820\pm 0.0067$	$0.7647\pm 0.0560$	9.42	0
KOEM nutrients replaced by CMEMS	$0.9355\pm 0.0092$	$0.6929\pm 0.0136$	–	–

Table 16: Retraining against the seasonal confound (8 external clusters, three seeds, mean

\pm

s.d.). The ratio is the mean probability in November 2025 divided by that in June–August; alarm days count 2024 days above an alert threshold, which was 0.30 for the second row and 0.50 for the third (at 0.30 the third configuration has 30 alarm days).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	Change	Params	ROC-AUC	PR-AUC	F1	Precision
v7 (Wando only, 30 ch)	recall-100% threshold introduced	–	0.903–0.927	–	0.907–0.926	83.3–86.2%
v13	two tiles	–	0.9071	0.8522	0.9193	87.72%
v14	+ Jindo (three tiles)	–	0.8378	0.8091	0.7408	70.80%
v16	+ Goheung, 2016 event (four tiles)	–	0.7790	0.7212	0.7580 †	–
v19	N:P ratio bug fixed; self-attention	3,819,650	0.8716	0.8516	0.8953	84.75%
xattn_v1	3-group hub cross-attention	5,922,690	0.9191	0.9110	0.8013	75.36%

Appendix

Table 17: Key development configurations (two-tile Wando+Mokpo task unless noted; F1 and precision at the recall-100% threshold, where some source tables label the F1 as macro-F1; pixel-level counting).

Measurement	Procedure	Date
Overall metrics and confusion matrix (Table 6 )	validation set, 6-seed ensemble, $\theta^{\star}=0.0940$	2026-10-04/05
Same-date control (Table 7 )	separate 5-fold CV training, within-date AUC, cluster bootstrap	2026-10-04/05
Inference port check	element-wise comparison with training path	2026-10-04
Experiments A and B (Tables 9 , 12 )	ablation script with pre-registered criteria	2026-10-06
2026 patch channel audit	zero-fraction check across machines	2026-10-07
Nutrient freshness (Table 11 )	per-source latest record	2026-10-07

Appendix

Table 18: Source and date of each reported measurement.

Time Series Forecasting

Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.

See all 839 papers

Predicting Pseudo-nitzschia harmful algal blooms along the Portuguese Coast using satellite-derived predictors
Jul 8, 2026Ayman Bnoussaad, El Khalil Cherif, Ligia Pinto +3Spatiotemporal Forecasting
ISR – Institute for Systems and Robotics, Lisbon, Portugal. · MARETEC/LARSyS, Associação do Instituto Superior Técnico para a Investigação e Desenvolvimento, Lisboa, Portugal. · IPMA – Portuguese Institute for Sea and Atmosphere, Algés, Portugal. +1
BG4Sea: Biogeochemical Seasonal Forecastability via Progressive Information Scaling
Jul 18, 2026Gabriela Martinez Balbontin, Anastase Charantonis, Dominique Bereziat +1Multivariate Time Series Forecasting Time Series Forecasting
Mercator Océan International LIP6, Sorbonne Université, Paris, France · INRIA (ARCHES) Paris, France
Neptune: An AI model for Global Ocean Subseasonal Prediction
Sep 8, 2026Davide Donno, Italo Epicoco, Massimo Cafaro +8Neural Surrogate Modeling Climate Modeling
Department of Engineering for Innovation, University of Salento, Via per Monteroni, Lecce, Italy · CMCC Foundation - Euro-Mediterranean Center on Climate Change, Italy · Department of Earth and Environmental Engineering, Columbia University, New York, NY, USA +2

Jul 8, 2026J/K move · Enter open · S save

Predicting Pseudo-nitzschia harmful algal blooms along the Portuguese Coast using satellite-derived predictors

Ayman Bnoussaad, El Khalil Cherif, Ligia Pinto, Ramiro Neves+2

ISR – Institute for Systems and Robotics, Lisbon, Portugal. · MARETEC/LARSyS, Associação do Instituto Superior Técnico para a Investigação e Desenvolvimento, Lisboa, Portugal. · IPMA – Portuguese Institute for Sea and Atmosphere, Algés, Portugal.+1

Spatiotemporal Forecasting

Open paper AskPDF arXiv

Pseudo-nitzschia diatoms pose recurrent risks to coastal ecosystems and shellfish harvesting along the Portuguese Atlantic coast. Here we develop and evaluate a spatio-temporal machine-learning framework to predict harmful algal bloom (HAB) occurrence using exclusively satellite-derived predictors under realistic forecasting constraints. We characterised environmental and biological variability across shellfish production zones (L1-L9) using 5,882 observations, providing system-wide context. Predictive models were developed for zones L1-L2, a hotspot for Pseudo-nitzschia and domoic acid events, using a decade-long dataset (2013-2023; 1,440 observations; more than 1,000 satellite-based predictors including sea surface temperature, an upwelling index, chlorophyll-a, and plankton functional types). Sampling locations were partitioned into ecologically meaningful sub-regions using a river-aware spatial clustering scheme. A stringent spatio-temporal cross-validation strategy that simultaneously withholds entire years and spatial clusters prevents leakage and closely mimics real-world forecasting conditions. HAB occurrence proved moderately predictable across model classes and feature configurations. Ensemble tree-based methods achieved the strongest discrimination: Random Forest reached 0.74 +/- 0.05 with environmental predictors; Extra Trees reached 0.77 +/- 0.06 with biological variables added. Feature-importance analyses revealed that seasonal structure, spatial context, and lagged environmental conditions dominate model decisions, while biological indicators refine bloom likelihood within physically favourable periods. The framework demonstrates operationally relevant skill for satellite-supported HAB early-warning systems along eastern boundary upwelling coasts.

Abstract

Time Series Forecasting

Explore similar work

Predicting Pseudo-nitzschia harmful algal blooms along the Portuguese Coast using satellite-derived predictors

BG4Sea: Biogeochemical Seasonal Forecastability via Progressive Information Scaling

Neptune: An AI model for Global Ocean Subseasonal Prediction