Search Results (966)

Beat-level synchrony among cardiomyocyte clusters is a critical indicator of cardiac electromechanical function. Traditional invasive approaches have substantial limitations, and conventional computer vision methods are poorly suited for resolving densely packed, adherent clusters. To address these challenges, we developed an analysis framework to characterize the beating characteristics of cardiomyocyte clusters from microscopic imaging data. Specifically, we propose CardioSegNet, a multi-task deep learning model that combines attention mechanisms with three prediction heads (semantic segmentation, contour detection, and distance transform), followed by a watershed algorithm to achieve high-accuracy cluster-level segmentation of cardiomyocyte clusters. The Pixel-Difference method is applied to extract time-series beating signals from each segmented cluster and compute several dynamic parameters, including beating amplitude, period, frequency, and the Beat Rate Irregularity (BRI). We further introduce PeriodAwareNAPTD_ij to quantify the beating synchrony among different clusters. Our experimental results show that CardioSegNet achieves a Dice coefficient of 0.8868 and an HD95 of 93.02 µm on an independent test set, demonstrating strong segmentation performance. The cardiomyocyte populations are not uniformly globally synchronized; rather, they consist of multiple local subgroups with high internal synchrony, and the degree of synchronization between clusters is positively correlated with their physical distance. This label-free analytical pipeline provides an efficient tool for myocardial function evaluation and cardiotoxicity screening in vitro. Full article

Porosity-related quantities from industrial X-ray CT depend on segmentation and inference choices. When inference defaults are omitted from the report, void or phase fractions can shift by amounts comparable to slice-to-slice variability. The contribution is metrological rather than architectural: we document a reproducible nnU-Net 2D workflow on Dataset601 CTVoid from semantic labels to slice-wise void fraction, optional two-dimensional connected-component pore summaries, isotropic three-dimensional stacking at 0.058 mm spacing, and spatial axis diagnostics, with region of interest and voxel spacing stated explicitly. The main results pair a weak export policy, defined as a single forward pass per slice without multi-scale fusion or test-time augmentation, with a strong policy that enables multi-scale fusion and flip-based augmentation on the same slice exports and identical weights, on one hundred consecutive slices from one cementitious industrial stack of 1028 × 1028 pixels. In parallel we report trainer validation on eight named Dataset601 validation cases and mirroring-based test-time augmentation off versus on re-inference on those same cases; case identifiers and the cross-validation split appear in the main text. These quantities answer different questions and must not be substituted for one another or for independent full-stack ground truth. Porosity-related scalars from industrial X-ray CT depend on how segmentation and inference are configured; when defaults are omitted, void fractions can shift by amounts comparable to slice-to-slice variability. For fixed nnU-Net weights on one cementitious industrial slice stack (1028 × 1028 pixels), we benchmark weak inference (single forward pass, no multi-scale fusion or test-time augmentation) against a strong export policy (multi-scale fusion and flip-based augmentation) on 100 paired slices, and report parallel trainer validation and TTA-off versus TTA-on re-inference on eight Dataset601 hold-out cases. For the industrial dataset, mean void-class IoU between modes is 0.716 (SD 0.043), while strong inference is ~2.6× slower and predicts lower mean void area (2.37% vs. 3.04%). The full weak export gives a 3D void ratio of 2.44% and integrated void volume of 5175 mm³. On validation patches, mean void Dice/IoU against the reference are 0.835/0.728, while weak–strong void IoU reaches 0.924 under the nnU-Net-native TTA contrast—quantities that must not be interchanged across domains or definitions. The present benchmark does not include a systematic polymer dosage series, and the study does not equate semantic void with open porosity but provides a reproducible disclosure template relevant to porous and polymer-modified cementitious CT reporting. Full article

Conventional ultrasound-based assessment of aortic stiffness relies on two-point distension metrics using maximum and minimum vessel diameters within a cardiac cycle, which may not fully reflect time-resolved aortic wall dynamics. This retrospective pilot study investigated the feasibility and clinical relevance of a time-series-based stiffness parameter, termed pulsatile stiffness-β, derived from automated segmentation of archived aortic M-mode ultrasound images. Seventy-nine cases with available aortic M-mode images were analyzed. Automated image processing was used to segment the anterior and posterior aortic walls and reconstruct diameter waveforms. Conventional stiffness-β, pulsatile stiffness-β, and wall asymmetry-related parameters were calculated and compared with demographic, tonometry-derived, hemodynamic, coronary burden, cardiovascular risk, and echocardiographic variables. Conventional and pulsatile stiffness-β were strongly correlated and showed directionally consistent associations with established vascular functional parameters, including systolic blood pressure, pulse pressure, augmentation pressure, age, and cardiovascular risk burden. Pulsatile stiffness-β demonstrated association patterns broadly comparable to conventional stiffness-β, suggesting its role as a waveform-informed extension rather than a superior alternative. Wall asymmetry-related parameters were associated with the Syntax score. Automated analysis of archived aortic M-mode images may provide feasible time-resolved vascular biomarkers for stiffness and wall motion assessment. Full article

In this work, we investigated the incorporation of a fluorescent side group, fluorescein octyl ester (FOE), in benzodithiophene-based donor–acceptor terpolymers as a strategy to modulate excited-state behavior. Three FOE-containing terpolymers (P2-iIa-c), obtained at different polymerization times, were systematically evaluated against an analogous material without the fluorescent pendant unit (P1-iI). Thermal analysis revealed good thermal stability and an increase in glass transition temperature upon FOE incorporation, suggesting restricted segmental mobility and increased conformational constraints within the conjugated backbone. Optical characterization showed distinct absorption spectra with reaction time and shorter fluorescence lifetimes for the FOE-containing materials, consistent with the presence of additional excited-state deactivation pathways and intramolecular energy transfer processes within the terpolymer backbone. An approximate estimation of energy transfer efficiencies (≈60–65%) suggested that such processes may be operative within the system. Cyclic voltammetry measurements showed only minor variations in HOMO and LUMO energy levels between P1-iI and P2-iIa-c series, indicating that the conjugated backbone predominantly determined the frontier orbital energies despite side chain modification. Furthermore, photocurrent measurements from the bilayer device configuration exhibited a systematic increase in photocurrent for the FOE-containing material, supporting the role of excitonic modulation, rather than significant changes in interfacial energetic alignment. These results suggest that fluorescent side chain incorporation provides an effective strategy for regulating exciton dynamics while maintaining the electronic structure of the donor–acceptor terpolymer. Full article

Rainstorm-induced disaster chains are characterized by high suddenness, immense destructive power, and complex chain propagation mechanisms. Traditional static assessment methods rely on fixed parameters and struggle to depict the dynamic evolution of such disasters. Existing dynamic models are mostly based on predefined structures and lack the capability to integrate multi-source data and quantify uncertainty, thereby constraining the accurate prediction of rainstorm disaster chains. To address these issues, this study proposes a rainstorm disaster chain prediction model (SW-DBN) that integrates a large language model (LLM)-optimized sliding window mechanism with a dynamic Bayesian network (DBN). The model first performs dynamic segmentation and feature extraction on multi-source time-series data through the sliding window mechanism and constructs an LLM-driven module for semantic understanding of multi-source information and latent parameter mining. By leveraging the LLM’s in-depth analysis of data pattern variations within the window, the model excavates latent parameters, adaptively adjusts the DBN network topology, and feeds back to optimize the window width and sliding step, thereby maintaining adaptive alignment between the sliding window’s feature extraction and the dynamic evolution of the disaster chain. Ultimately, the cascade propagation process of the rainstorm disaster chain is modeled, reasoned, and validated through the DBN, forming an integrated prediction framework of “perception–reasoning, dynamic regulation, and cascade verification.” A case study in the Xi’an area demonstrates that the proposed model can effectively simulate the temporal evolution of rainstorm disaster chains. The average prediction accuracy for four key types of disaster nodes reaches 84.8%, representing an improvement of 7.5 percentage points over the standard DBN model, with clear advantages in early warning timeliness for critical nodes. The proposed model provides technical support for the probabilistic prediction of rainstorm disaster chains and disaster prevention decision-making, featuring both dynamic adaptability and interpretability. Full article

In many university and healthcare projects, models are built for very different data types such as tables, institutional time series, and medical images, but they are deployed as separate applications. In this work, that separation made testing and maintenance difficult because each module had its own pipeline and runtime requirements. This paper presents an integrated AI lakehouse-style implementation that runs three model pipelines inside one containerized backend. For medical imaging, we used MRI datasets from IEEE DataPort: a four-class classification set with 7012 images (5708 train/1304 test) and a segmentation set with 3063 image–mask pairs. The classification model (ResNet50 transfer learning) is evaluated using a proper train–validation–test protocol across multiple splits (80/10/10, 70/10/20, 60/10/30, and 10/30/60), achieving a test accuracy of 99.00% under the standard 80/10/10 split. Additionally, a patient-level evaluation is conducted using an external glioma dataset to provide a more realistic assessment without data leakage. The segmentation model (DeepLabV3-ResNet50) achieved 83.09% validation mIoU and 88.79% Dice score. For university KPI forecasting, we used annual IPEDS and NSF HERD data from 2010 to 2023 for three universities (BSU, EOU, and UAB). To examine the effect of preprocessing on forecasting performance, two case studies are conducted. In the first case, linear interpolation is applied to generate semester-level data. In the second case, the original annual data is used directly without interpolation. Random Forest regression and ARIMA models are evaluated using MAE, RMSE, MAPE, and $R^2$. The results showed that interpolation improved apparent forecasting performance due to smoothing, while evaluation on the original annual data provided a more realistic assessment of model behavior. To further validate the framework on a larger dataset, an additional case study is conducted using a student dropout dataset. For water potability, we trained and compared multiple tabular classifiers on a large dataset (1,048,575 samples). A Random Forest model (100 trees, max depth 10) achieved 85.86% test accuracy and high recall for unsafe samples (0.8447). All modules are served via FastAPI and deployed together using Docker, with workflow automation routing requests to the correct endpoint. System-level benchmarking indicates that the backend maintains stable throughput and latency under concurrent requests. Full article

Low back pain (LBP) is a major global health problem and can result in a variety of movement impairments. Advances in smart technology have enabled the collection of novel streams of movement data, and machine learning (ML) methods have been increasingly used for data analysis. However, many existing technologies remain expensive and unsuitable for widespread clinical use, and ML approaches have largely focused on distinguishing people with LBP from healthy controls rather than identifying meaningful subgroups within the LBP population. Motion Tape (MT) is a recently developed wearable strain sensor that translates skin deformation from underlying movement and muscle engagement into electrical signals. In this exploratory study involving 10 participants with LBP, we demonstrate that MT data from six sensors applied on the lower back capture rich movement information capable of characterizing movement patterns among participants with LBP. We propose a feature engineering approach based on biomechanical features as well as time-series causal discovery applied to multivariate sensor time-series data to extract directed inter-segment coordination patterns. We further develop an exploratory subgroup discovery pipeline by aggregating clustering coassociation information across diverse movement tasks. Our causal coordination features show promising discriminative information across several movement types, capturing aspects of motor control not reflected in amplitude-based or embedding-based features alone, such as asymmetries and movement restrictions. Preliminary ensemble clustering analysis indicates three potential LBP subgroups distinguished by biomechanical and inter-segment coordination patterns, which may reflect varied strategies under different movement demands. We investigate the differences in clinical characteristics among these LBP subgroups. We show that time-series foundation models are not well suited for LBP subgrouping due to their uninterpretability, which is improved in our feature engineering pipeline. This framework could reveal additional subgroups with larger cohorts and may generalize to other sensor modalities. Full article

Background/Objectives: Robotic gastrectomy (RG) for gastric cancer requires structured implementation to ensure oncological safety, particularly in Western centers with lower case volumes. Indocyanine green (ICG)-guided near-infrared lymphography may facilitate adequate lymphadenectomy and reliable tumor localization. We report our stepwise institutional introduction of RG and evaluate the perioperative outcomes and diagnostic accuracy of ICG-guided lymphography. Methods: All consecutive patients who underwent curative-intent RG at the University Medical Center Ljubljana between June 2022 and September 2025 were retrospectively analyzed. The implementation followed a structured stepwise approach, beginning with subtotal gastrectomy and progressing to total gastrectomy after formal training at Severance Hospital, Yonsei University Health System, under the mentorship of Prof. Woo Jin Hyung. ICG was administered endoscopically the day before surgery for tumor localization and intraoperative lymphatic mapping. The operative learning curve was assessed by CUSUM analysis, segmented regression, and bootstrapped plateau estimation. Results: Thirty-eight patients underwent RG (17 subtotal and 21 total). R0 resection was achieved in 100% of cases. The conversion rate was 2.6%. Major complications (Clavien–Dindo ≥ IIIb) occurred in six patients (15.8%). The 30-day mortality rate was 0%, and the 90-day mortality rate was 2.6%. Bootstrapped plateau operative times were 321.2 min (95% Bias-corrected and accelerated confidence interval (BCa CI): 278.4–344.1) for subtotal and 413.5 min (95% BCa CI: 378.1–476.1) for total gastrectomy, with the steepest learning phase confined to the first 2–4 cases. ICG was used in 23 patients. In a validation subset of five patients (259 lymph node stations), the sensitivity and negative predictive value were both 100%, with zero false negatives in 57 ICG-negative stations. Conclusions: RG can be safely introduced using a structured, stepwise strategy supported by training at a high-volume expert center. ICG-guided lymphography demonstrated 100% sensitivity for tumor-draining nodal basins in a small validation cohort (n = 5), supporting the feasibility of the technique during program introduction and warranting prospective evaluation in larger series. Full article

Background and Objectives: The SARS-CoV-2 pandemic disrupted oral and maxillofacial surgery (OMS) services worldwide because of the high aerosol-generating nature of head-and-neck procedures, restricted access to elective dental care, and systemic reallocation of hospital resources. Continuous longitudinal multi-year data covering both the pandemic and the post-pandemic phases from regional Romanian (and more broadly central and southeastern European) emergency centers remain scarce. We aimed to quantify the impact of the pandemic on OMS activity in a large Romanian regional referral center and to evaluate post-pandemic resilience. Materials and Methods: We conducted a retrospective single-center study of all inpatient admissions to the OMS Clinic of a tertiary emergency hospital in western Romania between 1 January 2018 and 31 December 2024. Three periods were pre-specified: pre-pandemic (2018–2019), pandemic (2020–2022) and post-pandemic (2023–2024). A Newey–West segmented interrupted-time-series (ITS) regression and a negative-binomial monthly count model with Fourier seasonality were fitted; length of hospital stay was further analyzed with a multivariable gamma-log generalized linear model adjusted for age, sex, county, primary ICD-10 chapter and total ICD-10 codes. Variables analyzed included case volume, demographics, primary and secondary ICD-10 diagnoses, length of hospital stay (LOS), case complexity (total ICD-10 codes per admission) and in-hospital mortality. Results: A total of 11,628 inpatient admissions corresponding to 8084 unique patients (56.5% male; mean age 52.2 ± 19.2 years) were analyzed. Compared with the pre-pandemic baseline (mean 2037 admissions/year), annual volume dropped by 45.1% in 2020, 44.0% in 2021 and 32.3% in 2022, with a nadir of −76% during the first state of emergency (April 2020; n = 34 admissions). Recovery was rapid; 2024 exceeded the pre-pandemic baseline by +10.1% on raw counts and by +16.2% on admissions per 100,000 catchment population using year-specific INS denominators. The segmented ITS regression confirmed an immediate level drop of −114.2 admissions/month in March 2020 (95% CI −133.1 to −95.3; p < 0.001) and a positive post-intervention slope of +2.06 admissions/month (95% CI 1.23–2.88; p < 0.001), with observed monthly volume returning to the counterfactual projection by October 2023. The case mix shifted significantly (χ² = 406.9, p < 0.0001); elective benign neoplasm admissions were reduced from 7.2% to 2.0%, while neoplasms of uncertain behavior nearly doubled from 15.7% to 27.5%. Case complexity increased during the pandemic (mean ICD codes 4.08 ± 2.42 vs. 3.44 ± 2.30; p < 0.001); after exclusion of administrative codes (whole Z chapter and U07.x), the difference attenuated to 3.34 vs. 3.17 codes (still p < 0.001 by Kruskal–Wallis), indicating that the largest portion of the unadjusted increase was driven by the new mandatory pre-admission SARS-CoV-2 screening code Z11.5 rather than true clinical complexity. Notably, the clinically interpretable proxy R63.3 (feeding difficulty) independently rose from 41.5% to 53.1%. The crude median LOS did not differ between the pre-pandemic and pandemic periods (3.07 vs. 3.06 d; p = 0.19) and dropped significantly post-pandemic (2.22 d; p < 0.001); however, after multivariable adjustment for case mix, age, sex, county and code count, the LOS was 15.7% shorter during the pandemic (adjusted ratio 0.84, 95% CI 0.82–0.87; p < 0.001) and 22.8% shorter post-pandemic (adjusted ratio 0.77, 95% CI 0.75–0.80; p < 0.001) relative to baseline. Conclusions: The pandemic caused a severe but transient contraction of OMS activity accompanied by increased case complexity and a marked shift away from elective surgery. Inpatient volume returned to and exceeded the pre-pandemic baseline by 2024. These results support the value of standing pandemic-preparedness protocols, sustained access to preventive dental care, and integrated tele-triage pathways for future public-health crises. Full article

Global climate change and soil salinization pose challenges to licorice cultivation. Evaluating seed vigor based on the dynamic changes in radicle morphology is crucial for screening and cultivating licorice varieties that are tolerant to low temperatures and salts. Traditional manual measurement of licorice radicle characteristics suffers from issues such as high cost, long time consumption, and large errors. The YOLOv11 instance segmentation model in the field of deep learning offers advantages including a simple architecture, strong lightweight properties, and a unified detection-segmentation framework. Therefore, this study selected the YOLOv11 model to build a deep learning framework and used the continuous time-series crop growth vitality monitoring system to collect full-time-series images of 18 groups of licorice seeds germinating under different temperature and salt stress conditions. The YOLOv11-seg model was improved by adding a Spatial Strip Attention mechanism (SSA) to enhance the spatial correlation of radicle features, replacing ordinary convolutions with a Multi-scale Edge Detail Enhancement Module (MEEM) to optimize multi-scale feature extraction capabilities, and embedding a Normalized Weighted Distance (NWD) loss function to strengthen the segmentation ability for tiny targets. The YOLOv11-LicoSeg model was constructed for segmenting and extracting licorice radicle features and calculating root length. The experimental results showed that the mAP50 of the model’s detection reached 97.4%, mAP50–95 reached 81.7%, the mAP50 of the segmentation mask reached 97.0%, and mAP50–95 reached 78.2%. Compared with the unimproved YOLOv11-seg, the mAP50 of detection increased by 0.7%, mAP50–95 increased by 1.3%, the mAP50 of segmentation increased by 0.7%, and mAP50–95 increased by 0.8%. The linear regression coefficient between manual measurement and machine-vision measurement was 0.94218, and the goodness of fit R² was 0.94408. Using this model and the monitoring system, the morphological evolution of the licorice radicle contour characteristics over the germination time was obtained. The study indicated that the growth of licorice radicles was optimal under salt stress of 1200 µs/cm and 1800 µs/cm. YOLOv11-LicoSeg accurately segmented licorice radicles and calculated radicle length, with the performance to segment 100 licorice radicle images within 7 s. After deployment, it significantly reduced the labor cost and time consumption for acquiring licorice radicle phenotypes. In conclusion, YOLOv11-LicoSeg provides a rapid and accurate method for variety screening in licorice breeding and cultivation. Full article

In hybrid-series-converter-based LCC-HVDC systems, controllable capacitor modules can provide additional voltage–time area during commutation, thereby improving inverter-side fault tolerance under AC faults. However, their switching behavior makes the commutation path impedance state-dependent, while most existing commutation-failure prediction methods still rely on fixed-reactance assumptions. To address this problem, this paper proposes a reactance-corrected predictive control and coordinated switching method. First, a capacitor switching coefficient is introduced to describe the insertion state of the controllable capacitor modules, and an equivalent commutation reactance of the HSC valve arm is derived. Then, the corrected reactance is incorporated into an extinction-angle margin index and an energy-margin index to quantify the influence of reactance variation on commutation capability. A segmented firing-angle controller with smooth compensation is further designed, and energy-margin feedback is coordinated with capacitor insertion control. PSCAD/EMTDC simulations verify that the proposed method reduces prediction error, provides a prediction lead time of 0.7–4.5 ms, and improves fault ride-through capability. Full article

This study compares deep learning models for point-level anomaly detection in multichannel satellite telemetry data. Raw event-based telemetry was converted into segment-based multivariate time series without windowing or feature extraction, allowing models to learn system behavior at each time step. Preprocessing included channel alignment, training set-based normalization, missing value imputation, and temporal label smoothing, while Focal Loss and segment-level oversampling addressed class imbalance. Five architectures, BiLSTM, BiGRU, Transformer, Hybrid BiLSTM–Transformer, and Hybrid BiGRU–Transformer, were evaluated, with thresholds optimized on a validation set. The results show that hybrid models combining recurrent networks and attention mechanisms effectively capture both short- and long-term dependencies. The standalone BiGRU model achieves the highest overall classification performance in terms of F1 score and accuracy. In contrast, the Hybrid BiGRU–Transformer architecture does not outperform BiGRU in classification metrics but provides superior temporal stability, improved boundary sensitivity, and better interpretability in anomaly detection tasks. Full article

Understanding sea surface temperature (SST) variability is essential for assessing climate-driven changes in semi-enclosed basins such as the Black Sea. This study investigates SST variability in the western Black Sea over 2003–2024 using MODIS Aqua nighttime SST observations. Annual mean SST time series were constructed for coastal, shelf, and open-sea subregions and analysed using linear regression, ARIMA modelling, segmented regression, and spectral methods. SST exhibits a persistent warming signal across all subregions, with an overall increase of ~2.0–2.5 °C and a mean trend of 0.64 ± 0.08 °C decade⁻¹. Warming is spatially heterogeneous, with stronger trends in coastal and shelf regions, indicating a pronounced cross-shelf gradient. Temporal evolution reveals a multi-phase structure, with breakpoints around 2006–2008 and ~2022 marking shifts in warming intensity. Extreme anomalies include basin-wide cooling in 2017 and a sustained warming episode during 2019–2020. Statistical analyses indicate that SST variability is dominated by short-term persistence, while the influence of the North Atlantic Oscillation (NAO) is weak at the annual scale. In addition to the warming trend, SST variability undergoes a systematic reorganization, with variability remaining pronounced and spatially differentiated, particularly in coastal and shelf regions. Near-term projections further suggest that SST evolution may be moderated by internal variability, resulting in limited net change relative to recent peak conditions. Overall, SST variability reflects the combined effects of basin-scale warming, stratification, and regional circulation, indicating a transition toward a more stratified and dynamically variable system with implications for regional climate and marine ecosystems. Full article

Battery formation is important for ensuring the quality and service life of cells in lithium-ion battery (LIB) production. During the formation process, fault cells, such as low open-circuit voltage cells, are screened offline after the charging stage since, in most formation protocols, the online screening process is absent. This can lead to energy waste and extend the rework cycle of the fault cells in the LIB formation process. To address this problem, this paper considers the online fault cell screening problem, the formation pre-screening, in the LIB formation process as a classification task and proposes a data-driven model based on incomplete time-series data for formation pre-screening. First, the proposed model transforms segments of the incomplete charging voltage curve (ICVC) of the LIB as tokens, which is a more compact and less redundant data representation of the ICVC. Then, the attention-based feature encoder, Transformer encoder (TE), captures the dependency between tokens to extract features for the formation pre-screening. Finally, a task-specified decoder, feature enhance decoder (FED), is used to screen out fault cells online. The effectiveness of the proposed model is verified using real-world production data collected from a specific type of 18,650 lithium-ion cell under one formation protocol. The results on the investigated industrial dataset show that the proposed model achieves an accuracy of 98.73% and a miss rate of 1.92% during formation pre-screening, which is a 2.49% improvement in accuracy and an 8.98% decrease in miss rate compared with the deep residual network baseline. These results demonstrate the feasibility of using incomplete formation-stage voltage curves for online fault-cell pre-screening, which has the potential to reduce unnecessary charging and rework time in LIB production. Full article

River flooding and urban inundation require forecasting systems that can anticipate future risk, rather than systems that only estimate the current water state. However, real-world closed-circuit television (CCTV)-based flood datasets often contain imbalanced or temporally inconsistent risk labels. In addition, most image-based approaches remain limited to static scene understanding. This study proposes a dataset reformulation and temporal multi-task forecasting framework for CCTV-based flood-risk prediction. First, we introduce a site-relative relabeling strategy that converts noisy frame-level danger annotations into four risk levels using visual flood indicators and lightweight environmental cues. Second, we transform the original frame-based dataset into site-hour sequences for multi-horizon forecasting at 1 h, 3 h, and 6 h. Third, we evaluate image-only, weather-only, and naive multimodal configurations to examine the role and limitations of heterogeneous sensor fusion. On the reformulated dataset, the image-only temporal model achieved the best overall performance, with a mean Intersection over Union (mIoU) of 0.892, Dice score of 0.940, macro-averaged F1 score (Macro-F1) of 0.532, and high-risk recall of 0.642. In contrast, naive multimodal fusion reduced Macro-F1 to 0.267 and high-risk recall to 0.070. This result indicates that additional weather inputs do not automatically improve prediction when cross-modal signals are noisy, weakly correlated, or temporally misaligned. The ablation results further showed that removing temporal modeling decreased Macro-F1 to 0.227 and high-risk recall to 0.000. These findings demonstrate that dataset reformulation and temporal modeling are essential for extending CCTV-based flood analysis from static estimation to future risk forecasting. They also suggest that robust cross-modal alignment is required before multimodal sensing can provide reliable performance gains. Full article