Dense Neural Network Research Articles

When Sparse Neural Network Meets Label Noise Learning: A Multistage Learning Framework.

Recent methods in network pruning have indicated that a dense neural network involves a sparse subnetwork (called a winning ticket), which can achieve similar test accuracy to its dense counterpart with much fewer network parameters. Generally, these methods search for the winning tickets on well-labeled data. Unfortunately, in many real-world applications, the training data are unavoidably contaminated with noisy labels, thereby leading to performance deterioration of these methods. To address the above-mentioned problem, we propose a novel two-stream sample selection network (TS3-Net), which consists of a sparse subnetwork and a dense subnetwork, to effectively identify the winning ticket with noisy labels. The training of TS3-Net contains an iterative procedure that switches between training both subnetworks and pruning the smallest magnitude weights of the sparse subnetwork. In particular, we develop a multistage learning framework including a warm-up stage, a semisupervised alternate learning stage, and a label refinement stage, to progressively train the two subnetworks. In this way, the classification capability of the sparse subnetwork can be gradually improved at a high sparsity level. Extensive experimental results on both synthetic and real-world noisy datasets (including MNIST, CIFAR-10, CIFAR-100, ANIMAL-10N, Clothing1M, and WebVision) demonstrate that our proposed method achieves state-of-the-art performance with very small memory consumption for label noise learning. Code is available at https://github.com/Runqing-forMost/TS3-Net/tree/master.

Generating Stellar Spectra Using Neural Networks

A new generative technique is presented in this paper that uses Deep Learning to reconstruct stellar spectra based on a set of stellar parameters. Two different Neural Networks were trained allowing the generation of new spectra. First, an autoencoder is trained on a set of BAFGK synthetic data calculated using ATLAS9 model atmospheres and SYNSPEC radiative transfer code. These spectra are calculated in the wavelength range of Gaia RVS between 8400 and 8800 Å. Second, we trained a Fully Dense Neural Network to relate the stellar parameters to the Latent Space of the autoencoder. Finally, we linked the Fully Dense Neural Network to the decoder part of the autoencoder and we built a model that uses as input any combination of Teff, logg, vesini, [M/H], and ξt and output a normalized spectrum. The generated spectra are shown to represent all the line profiles and flux values as the ones calculated using the classical radiative transfer code. The accuracy of our technique is tested using a stellar parameter determination procedure and the results show that the generated spectra have the same characteristics as the synthetic ones.

Designing a use-error robust machine learning model for quantitative analysis of diffuse reflectance spectra.

Machine learning (ML)-enabled diffuse reflectance spectroscopy (DRS) is increasingly used as an alternative to the computation-intensive inverse Monte Carlo (MCI) simulation to predict tissue's optical properties, including the absorption coefficient, and reduced scattering coefficient, . We aim to develop a use-error-robust ML algorithm for optical property prediction from DRS spectra. We developed a wavelength-independent regressor (WIR) to predict optical properties from DRS data. For validation, we generated 1520 simulated DRS spectra with the forward Monte Carlo model, where to , and to . We introduced common use-errors, such as wavelength miscalibrations and intensity fluctuations. Finally, we collected 882 experimental DRS images from 170 tissue-mimicking phantoms and compared performances of the WIR model, a dense neural network, and the MCI model. When compounding all use-errors on simulated data, the WIR model best balanced accuracy and speed, yielding errors of 1.75% for and 1.53% for , compared to the MCI's 50.9% for and 24.6% for . Regarding experimental data, WIR model had mean errors of 13.2% and 6.1% for and , respectively. The errors for MCI were about eight times higher. The WIR model presents reliable use-error-robust optical property predictions from DRS data.

Deep learning based joint fusion approach to exploit anatomical and functional brain information in autism spectrum disorders

Background:The integration of the information encoded in multiparametric MRI images can enhance the performance of machine-learning classifiers. In this study, we investigate whether the combination of structural and functional MRI might improve the performances of a deep learning (DL) model trained to discriminate subjects with Autism Spectrum Disorders (ASD) with respect to typically developing controls (TD).Material and methodsWe analyzed both structural and functional MRI brain scans publicly available within the ABIDE I and II data collections. We considered 1383 male subjects with age between 5 and 40 years, including 680 subjects with ASD and 703 TD from 35 different acquisition sites. We extracted morphometric and functional brain features from MRI scans with the Freesurfer and the CPAC analysis packages, respectively. Then, due to the multisite nature of the dataset, we implemented a data harmonization protocol. The ASD vs. TD classification was carried out with a multiple-input DL model, consisting in a neural network which generates a fixed-length feature representation of the data of each modality (FR-NN), and a Dense Neural Network for classification (C-NN). Specifically, we implemented a joint fusion approach to multiple source data integration. The main advantage of the latter is that the loss is propagated back to the FR-NN during the training, thus creating informative feature representations for each data modality. Then, a C-NN, with a number of layers and neurons per layer to be optimized during the model training, performs the ASD-TD discrimination. The performance was evaluated by computing the Area under the Receiver Operating Characteristic curve within a nested 10-fold cross-validation. The brain features that drive the DL classification were identified by the SHAP explainability framework.ResultsThe AUC values of 0.66±0.05 and of 0.76±0.04 were obtained in the ASD vs. TD discrimination when only structural or functional features are considered, respectively. The joint fusion approach led to an AUC of 0.78±0.04. The set of structural and functional connectivity features identified as the most important for the two-class discrimination supports the idea that brain changes tend to occur in individuals with ASD in regions belonging to the Default Mode Network and to the Social Brain.ConclusionsOur results demonstrate that the multimodal joint fusion approach outperforms the classification results obtained with data acquired by a single MRI modality as it efficiently exploits the complementarity of structural and functional brain information.

Transformer models for astrophysical time series and the GRB prompt–afterglow relation

Abstract Transformer models have recently become very successful in the natural language domain. Their value as sequence-to-sequence translators there also makes them a highly interesting technique for learning relationships between astrophysical time series. Our aim is investigating how well such a transformer neural network can establish causal temporal relations between different channels of a single-source signal. We thus apply a transformer model to the two phases of gamma-ray bursts (GRBs), reconstructing one phase from the other. GRBs are unique instances where a single process and event produces two distinct time variable phenomena: the prompt emission and the afterglow. We here investigate if a transformer model can predict the afterglow flux from the prompt emission. If successful, such a predictive scheme might then be distilled to the most important underlying physics drivers in the future. We combine the transformer model with a novel dense neural network set-up to directly estimate the starting value of the prediction. We find that the transformer model can, in some instances, successfully predict different phases of canonical afterglows, including the plateau phase. Hence it is a useful and promising new astrophysical analysis technique. For the GRB test case, the method marginally exceeds the baseline model overall, but still achieves accurate recovery of the prompt–afterglow fluence–fluence correlation in reconstructed light curves. Despite this progress, we conclude that consistent improvement over the baseline model is not yet achieved for the GRB case. We discuss the future improvements in data and modelling that are required to identify new physical-relation parameters or new insights into the single process driving both GRB phases.

Classification and Comparative Evaluation of Text and Emoji‐Based Tweets With Deep Neural Network Models

Emojis have become increasingly prevalent in today’s digital world, allowing individuals to convey a wide range of emotions, from uncomplicated to intricate, to a greater extent than previously. Consequently, emojis are being utilized in sentiment analysis and tailored marketing strategies. The ongoing research on conducting emotion detection on both tweets and a symbolic expression dataset sourced from Kaggle. Given that tweets are largely commentaries, we utilized two end‐to‐end sentence embedding models, the DistilBERT, USE‐Large3, and RoBERTa, which generate embeddings. These embeddings are further utilized for training with dense neural networks (NNs) and LSTM techniques. Remarkably, it was perceived that the text classification accuracy for both models was consistently high, hovering around 98%. However, when the validation set is constructed with that of symbolic expression or emojis not included in the training dataset, a significant drop in accuracy for both models, plummeting to 75%, had been observed. Additionally, a distributed training methodology is utilized as a substitute for the conventional single‐threaded model to enhance scalability. This approach resulted in a roughly 17% reduction in the runtime while maintaining accuracy. Lastly, in pursuit of explainable AI, the SHAP and LIME algorithms are employed to elucidate the model’s behavior and assess any potential biases in the dataset. The creative use of advanced deep NN techniques customized for the delicate complexities of hybrid‐data sentiment analysis indicates a significant leap forward. Our proposed work provides the critical gap in existing sentiment analysis methods, which primarily aimed at either text or emojis in isolation, thereby exploring more holistic understanding of sentiment in digital communications. Moreover, the application of explainable AI techniques, SHAP and LIME, to demystify model decisions emphasizes commitment to advancing transparent and trustworthy deep learning technologies in sentiment analysis.

An Efficient Sarcasm Detection using Linguistic Features and Ensemble Machine Learning

Sarcasm detection in written text has emerged as a significant research area within natural language processing (NLP). Sarcasm, characterized by conveying the opposite of the intended meaning often for humor, irony, or ridicule, poses a challenge due to its contextual and tonal nuances. This study investigates the application of machine learning methods to detect sarcasm in text due to its potential to reverse the overall sentiment expressed in a sentence. A total of 13 linguistic manually crafted features related to text meaning, word usage, lexical diversity, and readability are extracted. These features are then employed to train a variety of machine learning models including Gradient Boosting, Decision Tree, Random Forest, Support Vector Machine, Gaussian Nave Bayes, K-Nearest Neighbor, and Logistic Regression classifiers. Additionally, an Ensemble Model and a Dense Neural Network is developed, both trained on the extracted handcrafted features to showcase performance. The results reveal that the suggested ensemble model achieves a peak accuracy of 93% in sarcasm detection. The amalgamation of these 13 linguistic features enhances model performance when compared to other contemporary models, exhibiting an improvement of up to 5% in terms of F1-score using the publicly available gold standard News Headline Dataset.

Dense Neural Network for Classification of Seafloor Sediment using Backscatter Mosaic Feature

Water transportation plays a vital role in global economic activities, facilitating more than 85% of international trade and serving as a cost-effective and essential means to fulfill the demand for goods and services. Similarly, the Benoa Port, situated in the southern part of Denpasar City, operates in the same manner. By utilizing Multibeam Echo Sounder (MBES) backscatter data, backscatter mosaics can be generated to identify various seafloor sediment types, which consist of rock fragments, minerals, and organic materials. The characteristics of these sediments, such as grain size, density, composition, and others, can be observed. To improve the classification of sediments, the integration of backscatter data and backscatter features, such as ASM (Angular Second Moment), Energy, Contrast, and Correlation, can be employed. Supervised classification models like Dense Neural Network (DNN) can be utilized to accurately determine the types of seafloor sediments. The application of DNN modeling resulted in a training accuracy rate of 88% and a testing accuracy rate of 100%. The accuracy results delineated six distinct sediment types. Notably, sandy silt exhibited the highest distribution, accounting for 49.30%, whereas soft clayey silt registered the lowest distribution at 0.53%, as determined by their respective spatial prevalence.

Diagnosing Cognitive Proficiency of Students Using Dense Neural Networks for Adaptive Assistance

Diagnosing Cognitive Proficiency of Students Using Dense Neural Networks for Adaptive Assistance

Parameter Estimation Model for Stellar Spectral Data based on RDN

The temperature, gravity and other key parameters of stars change during their evolution, and finding the relationship between these parameters and the life cycle of stars has always been the research focus in astrophysics. How to estimate the parameters of massive spectral data more accurately is of great significance for studying the properties of stars. However, the current conventional parameter estimation methods have problems such as exploding gradient, vanishing gradient and mismatch of newly added stellar parameters in the face of explosive growth of astronomical data, resulting in low parameter estimation accuracy. The Residual Dense neural network (RDN) model proposed in this paper is mainly improved based on two more advanced neural networks, ResNet and DenseNet. The core of RDN is a new Residual Dense Block (RDB), which includes two modules: residual and dense. The purpose of the residual module is to learn the residual between input and output and add the identity map to it, which aims to solve the problem of vanishing gradient and exploding gradient in deep network training. The dense module is where each layer is directly connected to all the layers before it, allowing better utilization of gradients and feature reuse. Its main purpose is feature extraction. The proposed model was trained on the preprocessed LAMOST DR7 dataset, making uncertainty predictions for 17 stellar parameters in LAMOST DR7 spectra with a signal-to-noise ratio (SNR) equal to or greater than 10. The results show that the proposed model has high estimation accuracy and solves the problems existing in previous methods. Compared with ResNet, DenseNet and StarNet, the key indicators such as the mean absolute error of RDN are optimized.

Development Edge Device Monitoring System Stunting and Malnutrition in Golden age 0–5 years Integrated with AI

Golden age is the best period for a child's growth, monitoring of growth and development must be carried out regularly. Growth and development disorders in children are stunting and malnutrition. This incident also prompted the government through the Ministry of Health to create a special program towards a golden Indonesia 2045 to monitor stunting and disease, especially in children. The prevalence of stunting decreased to 21.6% from 24.4% in 2022. Early detection of stunting and malnutrition, where the research object is children aged 0-5 years. This prototype was built using a load cell sensor, a study of the use of optical sensors and ultrasonic sensors to measure body height, and a MAX sensor to detect children's anemia. The integration of this tool combines IoT and AI. The results obtained to validate the use of load cells have a reading error of 0.01% with an accuracy of 99%. Comparison using optical sensors and ultrasonic sensors. Optical sensors have result average error of 0.01, accuracy 98.99%, ultrasonic sensors error was 0. 15 with 85% accuracy. To measure malnutrition, the anemia parameters were processed using the Dense Neural Network (DNN) model with 256 neurons showing an accuracy of 98.03%.

From nuclear track characterization to machine learning based image classification in neutron autoradiography for boron neutron capture therapy.

Knowledge of the 10B microdistribution is of great relevance in BNCT studies. Since 10B concentration assesment through neutron autoradiography depends on the correct quantification of tracks in a nuclear track detector, image acquisition and processing conditions should be controlled and verified, in order to obtain accurate results to be applied in the frame of BNCT. With this aim, an image verification process was proposed, based on parameters extracted from the quantified nuclear tracks. Track characterization was performed by selecting a set of morphological and pixel-intensity uniformity parameters from the quantified objects (area, diameter, roundness, aspect ratio, heterogeneity and clumpiness). Their distributions were studied, leading to the observation of varying behaviours in images generated by different samples and acquisition conditions. The distributions corresponding to samples coming from the BNC reaction showed similar attributes in each analyzed parameter, proving to be robust to the experimental process, but sensitive to light and focus conditions. Considering those observations, a manual feature extraction was performed as a pre-processing step. A Support Vector Machine (SVM) and a fully dense Neural Network (NN) were optimized, trained, and tested. The final performance metrics were similar for both models: 93%-93% for the SVM, vs 94%-95% for the NN in accuracy and precision respectively. Based on the distribution of the predicted class probabilities, the latter had a better capacity to reject inadequate images, so the NN was selected to perform the image verification step prior to quantification. The trained NN was able to correctly classify the images regardless of their track density. The exhaustive characterization of the nuclear tracks provided new knowledge related to the autoradiographic images generation. The inclusion of machine learning in the analysis workflow proves to optimize the boron determination process and paves the way for further applications in the field of boron imaging.

APPLICATION OF DEEP LEARNING CROP CLASSIFICATION MODEL BASED ON MULTISPECTRAL AND SAR SATELLITE IMAGERY

Abstract. Classifying crops using satellite data is a challenge, especially since most crops have similar growth cycles. Due to their different characteristics and chlorophyll content, different crops exhibit subtle differences in their reflectance spectra. This study uses a data-driven approach to build a series of deep learning models to classify 36 different land covers in Steele County and Traill Country, North Dakota, US. A Google Earth Engine workflow was implemented to generate a composite layer containing Sentinel 1 and Sentinel 2 satellite data and surface crop data over the study area. 200,000 sample points were generated on this layer, 140,000 for training dataset, 30,000 for validation dataset and 30,000 for testing dataset. Each sample point contains the values of 12 months of SAR and spectral data. In this way, a two-dimensional feature matrix of the time dimension and spectral band dimension (bands refer to specific wavelengths of data in remote sensing imagery and other type of data like NDVI) is generated for each sample point. The training dataset of the model is composed of the feature matrix of these sample points, and the surface crops as labels correspond to the feature matrix. Since this is a dataset with two-dimensional features, this research uses four deep learning models: Dense Neural Network (DNN), Long short-term memory (LSTM), Convolutional neural network (CNN) and Transformer. Among them, the Transformer model based on the self-attention mechanism performed the best, with a comprehensive accuracy rate of 85%, and the classification accuracy rate of crops with more than 2,000 sample points in the training data set reached more than 90%.

Methodological Approach for Identifying Websites with Infringing Content via Text Transformers and Dense Neural Networks

The rapid evolution of the Internet of Everything (IoE) has significantly enhanced global connectivity and multimedia content sharing, simultaneously escalating the unauthorized distribution of multimedia content, posing risks to intellectual property rights. In 2022 alone, about 130 billion accesses to potentially non-compliant websites were recorded, underscoring the challenges for industries reliant on copyright-protected assets. Amidst prevailing uncertainties and the need for technical and AI-integrated solutions, this study introduces two pivotal contributions. First, it establishes a novel taxonomy aimed at safeguarding and identifying IoE-based content infringements. Second, it proposes an innovative architecture combining IoE components with automated sensors to compile a dataset reflective of potential copyright breaches. This dataset is analyzed using a Bidirectional Encoder Representations from Transformers-based advanced Natural Language Processing (NLP) algorithm, further fine-tuned by a dense neural network (DNN), achieving a remarkable 98.71% accuracy in pinpointing websites that violate copyright.

GADNN: A graph attention-based method for drug-drug association prediction considering the contribution rate of different types of drug-related features

The simultaneous use of multiple drugs, known as drug combinations, is increasingly common in the treatment of complex diseases such as cancer. However, drug-drug interactions (DDIs) caused by the simultaneous use of multiple drugs can lead to unexpected side effects and even death, resulting in high medical costs. Therefore, understanding potential DDI is an essential step in reducing the risk of adverse drug reactions before administering clinical drug compounds. Recently, many machine learning-based methods for DDI prediction have been introduced, but in these methods, no attention is paid to determining the contribution rate of each type of drug feature. Here, a Graph Attention-based Deep Neural Network method called GADNN is proposed to predict DDIs based on considering the influence or contribution rate of different drug-related features. The introduced method consists of two stages. In the first stage, four basic drug datasets, including target, enzyme, infrastructure, and pathway, are used as types of drug features. Using a graph neural network-based method, the embedding vector is generated for each drug based on each dataset separately. In the second stage, the contribution coefficient of each dataset is calculated dynamically through a new graph attention mechanism. Then, the weighted combination of the separate vectors is used to predict the probability of drug-drug interactions using a dense neural network. The experimental results confirmed that the presented innovation has been able to improve the accuracy compared to state-of-the-art methods.

CoGNN: An Algorithm-Hardware Co-Design Approach to Accelerate GNN Inference With Minibatch Sampling

As a new algorithm of graph embedding, Graph Neural Networks (GNNs) have been widely used in many fields. However, GNN computing has the characteristics of both sparse graph processing and dense neural network, which make it difficult to be deployed efficiently on the existing graph processing accelerators or neural network accelerators. Recently, some GNN accelerators have been proposed, but the following challenges have not been fully solved: 1) The mini-batch GNN inference scenario has the potential of software and hardware co-design, which can bring 30% computation amount reduction, and this is not well utilized. Besides, the cost of Message Flow Graph construction is large and may account for more than 50% of the total delay. 2) The feature aggregation has a large amount of data access and relatively small amount of computation, which leads to low on-chip data reuse, only 10% of dense computing. 3) Without the optimization of sparse computing units, simple memory bank and cross bar architecture can easily lead to bank access conflict and load imbalance, reducing the utilization of computing units to less than 60%. In order to solve the above problems, we propose a algorithm-hardware co-design scheme to accelerate GNN inference, which includes three technologies: 1) A reuse-aware sampling method is proposed for mini-batch inference scenarios, which reduces 30% of the calculation and improves the on-chip reusability of local data. 2) Through the node-wise parallelism-aware quantization, the features and weights are quantized to integers with eight or four bits, which reduces the amount of memory access by at least four times. 3) An accelerator supporting the above technologies is designed and evaluated, and different operations are supported by the sampling-inference integration architecture. The multi-bank on-chip memory pool is designed to support data reuse, and edge stream reordering is used to reduce data access conflicts, improving the utilization of computing units by 1.5x. Combined with the above technologies, the experiments show that our design achieves 9.2× speedup and 29× energy efficiency improvement compared with Deep Graph Library framework running on servers equiped with CPU and GPU.

Advantages of Hybrid Neural Network Architectures to Enhance Prediction of Tensile Properties in Laser Powder Bed Fusion

The properties of AlSi10Mg produced by Laser Powder Bed Fusion (PBF-LB) are defined by a multitude of different machine and laser parameters. This multi-parameter space presents the challenge of optimizing the material properties for a given application by the sheer amount of possible parameter combinations. Characterizing this multi-parameter space empirically is limited by time and resources and thus yields an incomplete picture of the process capabilities and local optima, respectively. To improve on this situation, machine learning to map the process parameters on the tensile properties of AlSi10Mg was used. The Hybrid Neural Network (HNN) used in this study consisted of a Convolutional Neural Network (CNN) to process the micrographs and a Dense Neural Network (DNN) to process the LPBF process parameters as well as the output of the CNN. The micrographs given to the CNN part of the network were printed with the same parameters given to the DNN part to include the information of the bulk microstructure as it strongly influences the tensile properties of the material. With the HNN, we observed good accuracy of the predicted tensile properties, given the small amount of training data. Furthermore, we explore which features of the micrographs were extracted by the CNN.

Dense neural network outperforms other machine learning models for scaling-up lichen cover maps in Eastern Canada.

Lichen mapping is vital for caribou management plans and sustainable land conservation. Previous studies have used random forest, dense neural network, and convolutional neural network models for mapping lichen coverage. However, to date, it is not clear how these models rank in this task. In this study, these machine learning models were evaluated on their ability to predict lichen percent coverage in Sentinel-2 imagery in Québec and Labrador, Canada. The models were trained on 10-m resolution lichen coverage (%) maps created from 20 drone surveys collected in July 2019 and 2022. The dense neural network achieved a higher accuracy than the other two, with a reported mean absolute error of 5.2% and an R2 of 0.76. By comparison, the random forest model returned a mean absolute error of 5.5% (R2: 0.74) and the convolutional neural network had a mean absolute error of 5.3% (R2: 0.74). A regional lichen map was created using the trained dense neural network and a Sentinel-2 imagery mosaic. There was greater uncertainty on land covers that the model was not exposed to in training, such as mines and deep lakes. While the dense neural network requires more computational effort to train than a random forest model, the 5.9% performance gain in the test pixel comparison renders it the most suitable for lichen mapping. This study represents progress toward determining the appropriate methodology for generating accurate lichen maps from satellite imagery for caribou conservation and sustainable land management.

Dense neural network based arrhythmia classification on low-cost and low-compute micro-controller

The electrocardiogram (ECG) monitoring device is an expensive albeit essential device for the treatment and diagnosis of cardiovascular diseases (CVD). The cost of this device typically ranges from $2000 to $10000. Several studies have implemented ECG monitoring systems in micro-controller units (MCU) to reduce industrial development costs by up to 20 times. However, to match industry-grade systems and display heartbeats effectively, it is essential to develop an efficient algorithm for detecting arrhythmia (irregular heartbeat). Hence in this study, a dense neural network to detect arrhythmia on the Arduino Nano is developed. The Nano consists of the ATMega328 microcontroller with a 16 MHz clock, 2 KB of SRAM, and 32 KB of program memory. Additionally, the AD8232 SparkFun Single-Lead Heart Rate Monitor is used as the ECG sensor. The implemented neural network model consists of two layers (excluding the input) with 10 and four neurons respectively with sigmoid activation function. However, four approaches are explored to choose the appropriate activation functions. The model has a size of 1.267 KB, achieves an F1 score (macro-average) of 78.3% for classifying four types of arrhythmia, an accuracy rate of 96.38%, and requires 0.001314 MOps of floating-point operations (FLOPs). It significantly outperformed recent studies in terms of neural network metrics and compute resource usage efficiency.

Forecasting freshwater cyanobacterial harmful algal blooms for Sentinel-3 satellite resolved U.S. lakes and reservoirs

This forecasting approach may be useful for water managers and associated public health managers to predict near-term future high-risk cyanobacterial harmful algal blooms (cyanoHAB) occurrence. Freshwater cyanoHABs may grow to excessive concentrations and cause human, animal, and environmental health concerns in lakes and reservoirs. Knowledge of the timing and location of cyanoHAB events is important for water quality management of recreational and drinking water systems. No quantitative tool exists to forecast cyanoHABs across broad geographic scales and at regular intervals. Publicly available satellite monitoring has proven effective in detecting cyanobacteria biomass near-real time within the United States. Weekly cyanobacteria abundance was quantified from the Ocean and Land Colour Instrument (OLCI) onboard the Sentinel-3 satellite as the response variable. An Integrated Nested Laplace Approximation (INLA) hierarchical Bayesian spatiotemporal model was applied to forecast World Health Organization (WHO) recreation Alert Level 1 exceedance >12 μg L−1 chlorophyll-a with cyanobacteria dominance for 2192 satellite resolved lakes in the United States across nine climate zones. The INLA model was compared against support vector classifier and random forest machine learning models; and Dense Neural Network, Long Short-Term Memory (LSTM), Recurrent Neural Network (RNN), and Gneural Network (GNU) neural network models. Predictors were limited to data sources relevant to cyanobacterial growth, readily available on a weekly basis, and at the national scale for operational forecasting. Relevant predictors included water surface temperature, precipitation, and lake geomorphology. Overall, the INLA model outperformed the machine learning and neural network models with prediction accuracy of 90% with 88% sensitivity, 91% specificity, and 49% precision as demonstrated by training the model with data from 2017 through 2020 and independently assessing predictions with data from the 2021 calendar year. The probability of true positive responses was greater than false positive responses and the probability of true negative responses was less than false negative responses. This indicated the model correctly assigned lower probabilities of events when they didn't exceed the WHO Alert Level 1 threshold and assigned higher probabilities when events did exceed the threshold. The INLA model was robust to missing data and unbalanced sampling between waterbodies.