Deep Learning Network Model Research Articles

Maturity discrimination of tobacco leaves for tobacco harvesting robots based on a Multi-Scale branch attention neural network

Traditional machine vision methods are susceptible to environmental influences, low recognition accuracy, and poor real-time performance in identifying tobacco leaf maturity. Currently, the maturity of tobacco leaves heavily relies on manual observation, and existing research on tobacco maturity discrimination focuses on standard light sources and standard background, which difficult to be applied to tobacco harvesting robots in natural environment. To tackle this problem, a maturity discrimination method based on the branch attention mechanism is proposed. Firstly, the region of interest in the extracted tobacco leaf images is obtained as a foundation for subsequent image processing. Then, a deep learning network model called the Multi-Scale Branch Attention Network (MSBANet) is designed for identification of tobacco leaf maturity. The effectiveness of MSBANet stems from its newly designed modules for multi-scale information extraction and branch attention. These modules help overcome the impact of lighting conditions and enable the extraction of more comprehensive and informative feature representations, capturing the relationship between phenotype and maturity. On the self-constructed dataset of tobacco leaf maturity (color images are 576 × 768 pixels), MSBANet-L achieves 89.6 % accuracy with 7.1GFLOPs and MSBANet-S achieves 88.2 % accuracy with 1.7 GFLOPs. When compared with widely used BP and MobileNet, as well as state-of-the-art structures such as YOLOv7-tiny-SiLU and FasterNet-T0, MSBANet demonstrates excellent performance. This method preliminarily explores the maturity discrimination of tobacco leaves in actual environments and shows a potential for development of tobacco harvesting robots.

A New Dual-Branch Embedded Multivariate Attention Network for Hyperspectral Remote Sensing Classification

With the continuous maturity of hyperspectral remote sensing imaging technology, it has been widely adopted by scholars to improve the performance of feature classification. However, due to the challenges in acquiring hyperspectral images and producing training samples, the limited training sample is a common problem that researchers often face. Furthermore, efficient algorithms are necessary to excavate the spatial and spectral information from these images, and then, make full use of this information with limited training samples. To solve this problem, a novel two-branch deep learning network model is proposed for extracting hyperspectral remote sensing features in this paper. In this model, one branch focuses on extracting spectral features using multi-scale convolution and a normalization-based attention module, while the other branch captures spatial features through small-scale dilation convolution and Euclidean Similarity Attention. Subsequently, pooling and layering techniques are employed to further extract abstract features after feature fusion. In the experiments conducted on two public datasets, namely, IP and UP, as well as our own labeled dataset, namely, YRE, the proposed DMAN achieves the best classification results, with overall accuracies of 96.74%, 97.4%, and 98.08%, respectively. Compared to the sub-optimal state-of-the-art methods, the overall accuracies are improved by 1.05, 0.42, and 0.51 percentage points, respectively. The advantage of this network structure is particularly evident in unbalanced sample environments. Additionally, we introduce a new strategy based on the RPNet, which utilizes a small number of principal components for feature classification after dimensionality reduction. The results demonstrate its effectiveness in uncovering compressed feature information, with an overall accuracy improvement of 0.68 percentage points. Consequently, our model helps mitigate the impact of data scarcity on model performance, thereby contributing positively to the advancement of hyperspectral remote sensing technology in practical applications.

COVID-19 Forecasting Using Recurrent Neural Network and Machine Learning

The COVID-19 variant’s complexity and the dire consequences of the variant spread are the inspirations behind the profound research on modeling and predicting new emerging variant surges. Multiple factors, including the variant characteristics, vaccination rate, the immune response of vaccinated individuals, and disease prevention health policies, impact the COVID-19 infection trend. The advancements in machine learning and neural network models, combined with the growth in computing, have demonstrated outstanding potential in modeling and predicting epidemic diseases. This research presents the modeling and prediction of the combined COVID-19 variant infection trends using the Holt-Winters exponential smoothing and seasonal auto-regressive integrated moving average with exogenous factors (SARIMAX) time-series machine learning models and recurrent neural network (RNN) long short-term memory (LSTM) model. Real-world United States COVID variant data from the Centers for Disease Control and Prevention (CDC) is used for prediction. The SARIMAX model factoring the seasonality of COVID-19 infections showed higher prediction accuracy than the Holt-Winters model, which is heavily weighted towards the most recent trends. The LSTM model had the best prediction accuracy of 91% with the lowest root mean square error (RMSE) values due to its property of selectively remembering patterns for long duration and the forget gates that correct the vanishing gradient problem, minimizing the error losses. This research demonstrates the promising application of the neural network deep learning models for epidemic disease modeling and prediction, enabling timely assessment of different policy decisions to mitigate the impact of an epidemic.

A Random Forest-Convolutional Neural Network Deep Learning Model for Predicting the Wholesale Price Index of Potato in India

A Random Forest-Convolutional Neural Network Deep Learning Model for Predicting the Wholesale Price Index of Potato in India

Copula deep learning control chart for multivariate zero inflated count response variables

With the increasing popularity of big data analysis, research on zero-inflated count data with the copula method has garnered significant attention because zero-inflated count data do not follow a normal distribution and have high correlation among variables. Within the domain of quality control, there has been limited emphasis on multivariate statistical process control (SPC) techniques that specifically address the challenge of multicollinearity within regression models for multivariate zero-inflated count responses. In this paper, we explain a computational challenge in handling big data with parametric generalized linear models, such as the zero-inflated Poisson model. This challenge motivates us to introduce a copula-based deep learning and neural network model involving multivariate zero-inflated count response variables with highly correlated explanatory variables. This approach builds upon existing deep learning and neural network models designed for univariate count responses, expanding them to encompass multivariate count response models through the integration of a copula regression framework. To evaluate the performance of our proposed methodology, we conduct a comparative analysis of accuracies using the zero-inflated Poisson model, univariate deep learning and neural network models, multivariate count response deep learning and neural network models, and our proposed copula deep learning and neural network models. This assessment involves employing metrics such as root mean square error (RMSE), weighted mean absolute percentage error (WMAPE), and mean absolute deviation (MAD). We include both copula-based asymmetrical zero-inflated simulated data and real-world data. We also propose a temporal dependence control chart for assessing the temporal dependence between bivariate zero-inflated count response variables. Our proposed copula deep learning and neural network temporal control charts can check time-varying dependence and outliers by leveraging the Shewhart statistical process control chart and t-copula ARMA-GARCH dynamic control correlation.

Multi-scale collaborative prediction of optimal configuration for carbon fiber woven composites based on deep learning neural networks

Since carbon fiber woven composite panels contain complex fiber bundle structures, the aim of this paper is to investigate the factors influencing the macro-equivalent properties and topological flexibility of composite panels resulting from variations in fiber bundle configuration. Firstly, the structural topology optimization design of carbon fiber composite plates was achieved using a multi-scale modeling approach with the variable density method. Nonlinear relationships between carbon fiber bundle configuration parameters and macroscopic equivalent properties, as well as macroscale topological flexibility, were determined using deep learning methods, respectively. The maximum error in the predicted values of the deep learning network model does not exceed 0.5% when compared with the results of the finite element calculations. In addition, utilizing deep learning networks to predict macro-equivalent properties of composites can significantly decrease computational time. The impact of varying the number of filaments in fiber bundles and the spacing of the fiber bundles on the flexibility of the topology was determined using a deep learning network. The optimal topological flexibility, along with its corresponding fiber bundle spacing and number of fiber bundle filaments, are globally searched within the design domain using an adaptive genetic algorithm. The study in this paper can provide a reference for designing fiber bundle configurations of composites for practical applications.

Forecasting Gas Well Classification Based on a Two-Dimensional Convolutional Neural Network Deep Learning Model

In response to the limitations of existing evaluation methods for gas well types in tight sandstone gas reservoirs, characterized by low indicator dimensions and a reliance on traditional methods with low prediction accuracy, therefore, a novel approach based on a two-dimensional convolutional neural network (2D-CNN) is proposed for predicting gas well types. First, gas well features are hierarchically selected using variance filtering, correlation coefficients, and the XGBoost algorithm. Then, gas well types are determined via spectral clustering, with each gas well labeled accordingly. Finally, the selected features are inputted, and classification labels are outputted into the 2D-CNN, where convolutional layers extract features of gas well indicators, and the pooling layer, which, trained by the backpropagation of CNN, performs secondary dimensionality reduction. A 2D-CNN gas well classification prediction model is constructed, and the softmax function is employed to determine well classifications. This methodology is applied to a specific tight gas reservoir. The study findings indicate the following: (1) Via two rounds of feature selection using the new algorithm, the number of gas well indicator dimensions is reduced from 29 to 15, thereby reducing the computational complexity of the model. (2) Gas wells are categorized into high, medium, and low types, addressing a deep learning multi-class prediction problem. (3) The new method achieves an accuracy of 0.99 and a loss value of 0.03, outperforming BP neural networks, XGBoost, LightGBM, long short-term memory networks (LSTMs), and one-dimensional convolutional neural networks (1D-CNNs). Overall, this innovative approach demonstrates superior efficacy in predicting gas well types, which is particularly valuable for tight sandstone gas reservoirs.

Research on inter-shaft bearing fault diagnosis method based on STFT-CNN modeling

Inter-shaft bearings are an essential component of aircraft engines, and their operational status determines the safety of aircraft engine operation. Therefore, to improve the accuracy of fault type prediction and enrich the feature information in vibration signals of aircraft engine inter-shaft bearings, this paper proposes an STFT-CNN model based on the AlexNet architecture, extending its application to the research of aircraft engine inter-shaft bearing fault diagnosis. This approach addresses the common reliance on personnel experience for fault type diagnosis in traditional aircraft engine inter-shaft bearing fault diagnosis. Firstly, real vibration fault signals from inter-shaft bearings are collected through experiments to enrich feature information in non-stationary signals using STFT time-frequency methods. Secondly, utilizing the high interpretability of the STFT-CNN model, fault feature data from inter-shaft bearings under various operating conditions are extracted to refine our understanding of fault feature information. Finally, leveraging the robustness of the STFT-CNN model, fault types are classified and predicted. The training process involves comparative analysis using different pooling algorithms, time-frequency analysis methods, and various deep learning network models. The results demonstrate that the STFT-CNN model, employing the maximum pooling algorithm, outperforms other models in predicting inter-shaft bearing faults, achieving an average fault prediction accuracy of 98.8% .

GSB: GNGS and SAG-BiGRU network for malware dynamic detection.

With the rapid development of the Internet, the continuous increase of malware and its variants have brought greatly challenges for cyber security. Due to the imbalance of the data distribution, the research on malware detection focuses on the accuracy of the whole data sample, while ignoring the detection rate of the minority categories' malware. In the dataset sample, the normal data samples account for the majority, while the attacks' malware accounts for the minority. However, the minority categories' attacks will bring great losses to countries, enterprises, or individuals. For solving the problem, this study proposed the GNGS algorithm to construct a new balance dataset for the model algorithm to pay more attention to the feature learning of the minority attacks' malware to improve the detection rate of attacks' malware. The traditional malware detection method is highly dependent on professional knowledge and static analysis, so we used the Self-Attention with Gate mechanism (SAG) based on the Transformer to carry out feature extraction between the local and global features and filter irrelevant noise information, then extracted the long-distance dependency temporal sequence features by the BiGRU network, and obtained the classification results through the SoftMax classifier. In the study, we used the Alibaba Cloud dataset for malware multi-classification. Compared the GSB deep learning network model with other current studies, the experimental results showed that the Gaussian noise generation strategy (GNGS) could solve the unbalanced distribution of minority categories' malware and the SAG-BiGRU algorithm obtained the accuracy rate of 88.7% on the eight-classification, which has better performance than other existing algorithms, and the GSB model also has a good effect on the NSL-KDD dataset, which showed the GSB model is effective for other network intrusion detection.

Generic Framework for Vehicle Identification System with Deep Learning Models

In contemporary vehicle scenario, different kinds of vehicles are playing vital role for the customers. Day by day, vehicles are increasing with different properties and identifications. Monitoring and maintain the records of vehicles in digital platform is the challenging task as well as a crucial task for the countries. Here, we have to take the scenario of digitalization of vehicles with deep learning network models for identification and maneuvering process. In our country, different states are following multiple vehicle identification like numbering, coloring and imaging. Many researchers were contributing towards this identification mechanism like supervised, unsupervised models for getting optimum accuracy. Text and Image classification are not sufficient to solve this vehicle identification problem. We are incorporating text, image and video related unstructured data set to solve this problem. We have framed a generic architecture for all kind of vehicles analyze the data. For instance, a vehicle is moving towards CCTV camera and it has been recorded as video content. In this paper, we come across video to still pictures and image segmentation and all the identification parameters like White, Yellow and green. We performed both text classification and image classification algorithms to acquire effective result compared to existing algorithms. Bi Directional RNN (LSTM/GRU) and LeNet, Alex Net, Inception, VggNet and ResNet algorithms are analyzed for this entire process. This is one of the globally challenging scenarios for identification with feature extraction, data cleaning and processing. At the end of the result, we achieved 98.94 % accuracy compare than other existing systems.

Deep learning radiomics-based prediction model of metachronous distant metastasis following curative resection for retroperitoneal leiomyosarcoma: a bicentric study

BackgroundCombining conventional radiomics models with deep learning features can result in superior performance in predicting the prognosis of patients with tumors; however, this approach has never been evaluated for the prediction of metachronous distant metastasis (MDM) among patients with retroperitoneal leiomyosarcoma (RLS). Thus, the purpose of this study was to develop and validate a preoperative contrast-enhanced computed tomography (CECT)-based deep learning radiomics model for predicting the occurrence of MDM in patients with RLS undergoing complete surgical resection.MethodsA total of 179 patients who had undergone surgery for the treatment of histologically confirmed RLS were retrospectively recruited from two tertiary sarcoma centers. Semantic segmentation features derived from a convolutional neural network deep learning model as well as conventional hand-crafted radiomics features were extracted from preoperative three-phase CECT images to quantify the sarcoma phenotypes. A conventional radiomics signature (RS) and a deep learning radiomics signature (DLRS) that incorporated hand-crafted radiomics and deep learning features were developed to predict the risk of MDM. Additionally, a deep learning radiomics nomogram (DLRN) was established to evaluate the incremental prognostic significance of the DLRS in combination with clinico-radiological predictors.ResultsThe comparison of the area under the curve (AUC) values in the external validation set, as determined by the DeLong test, demonstrated that the integrated DLRN, DLRS, and RS models all exhibited superior predictive performance compared with that of the clinical model (AUC 0.786 [95% confidence interval 0.649–0.923] vs. 0.822 [0.692–0.952] vs. 0.733 [0.573–0.892] vs. 0.511 [0.359–0.662]; both P < 0.05). The decision curve analyses graphically indicated that utilizing the DLRN for risk stratification provided greater net benefits than those achieved using the DLRS, RS and clinical models. Good alignment with the calibration curve indicated that the DLRN also exhibited good performance.ConclusionsThe novel CECT-based DLRN developed in this study demonstrated promising performance in the preoperative prediction of the risk of MDM following curative resection in patients with RLS. The DLRN, which outperformed the other three models, could provide valuable information for predicting surgical efficacy and tailoring individualized treatment plans in this patient population.Trial registration: Not applicable.

DesTrans: A medical image fusion method based on Transformer and improved DenseNet

Medical image fusion can provide doctors with more detailed data and thus improve the accuracy of disease diagnosis. In recent years, deep learning has been widely used in the field of medical image fusion. The traditional method of medical image fusion is to operate by superimposing and other methods of pixels. The introduction of deep learning methods has improved the effectiveness of medical image fusion. However, these methods still have problems such as edge blurring and information redundancy. In this paper, we propose a deep learning network model based on Transformer and an improved DenseNet network module integration that can be applied to medical images and solve the above problems. At the same time, the method can be moved to natural images. The use of Transformer and dense concatenation enhances the feature extraction capability of the method by limiting the feature loss which reduces the risk of edge blurring. We compared several representative traditional methods and more advanced deep learning methods with this method. The experimental results show that the Transformer and the improved DenseNet network module have a strong capability of feature extraction. The method yields good results both in terms of visual quality and objective image evaluation metrics.

LEET: stock market forecast with long-term emotional change enhanced temporal model.

The stock market serves as a macroeconomic indicator, and stock price forecasting aids investors in analysing market trends and industry dynamics. Several deep learning network models have been proposed and extensively applied for stock price prediction and trading scenarios in recent times. Although numerous studies have indicated a significant correlation between market sentiment and stock prices, the majority of stock price predictions rely solely on historical indicator data, with minimal effort to incorporate sentiment analysis into stock price forecasting. Additionally, many deep learning models struggle with handling the long-distance dependencies of large datasets. This can cause them to overlook unexpected stock price fluctuations that may arise from long-term market sentiment, making it challenging to effectively utilise long-term market sentiment information. To address the aforementioned issues, this investigation suggests implementing a new technique called Long-term Sentiment Change Enhanced Temporal Analysis (LEET) which effectively incorporates long-term market sentiment and enhances the precision of stock price forecasts. The LEET method proposes two market sentiment index estimation methods: Exponential Weighted Sentiment Analysis (EWSA) and Weighted Average Sentiment Analysis (WASA). These methods are utilized to extract the market sentiment index. Additionally, the study proposes a Transformer architecture based on ProbAttention with rotational position encoding for enhanced positional information capture of long-term emotions. The LEET methodology underwent validation using the Standard & Poor's 500 (SP500) and FTSE 100 indices. These indices accurately reflect the state of the US and UK equity markets, respectively. The experimental results obtained from a genuine dataset demonstrate that this method is superior to the majority of deep learning network architectures when it comes to predicting stock prices.

Intelligent Identification and Quantitative Characterization of Pores in Shale SEM Images Based on Pore-Net Deep-Learning Network Model

Among the various shale reservoir evaluation methods, the scanning electron microscope (SEM) image method is widely used. Its image can intuitively reflect the development stage of a shale reservoir and is often used for the qualitative characterization of shale pores. However, manual image processing is inefficient and cannot quantitatively characterize pores. The semantic segmentation method of deep learning greatly improves the efficiency of image analysis and can calculate the face rate of shale SEM images to achieve quantitative characterization. In this paper, the high-maturity shale of the Longmaxi Formation in the Changning area of Yibin City, Sichuan Province, and the low-maturity shale of Beibu Gulf Basin in China are studied. Based on the Pore-net network model, the intelligent identification and quantitative characterization of pores in shale SEM images are realized. The pore-net model is improved from the U-net deep-learning network model, which improves the ability of the network model to identify pores. The results show that the pore-net model performs better than the U-net model, FCN model, DeepLab V3 + model, and traditional binarization method. The problem of low accuracy of the traditional pore identification method is solved. The porosity of SEM images of high-maturity shale calculated by the pore-net network model is between 12 and 19% deviation from the experimental data. The calculated porosity of the SEM image of the low-maturity shale has a large deviation from the experimental data, which is between 14 and 47%. Compared with the porosity results calculated by other methods, the results calculated by pore-net are closer to the true value, which proves that the porosity calculated by the pore-net network model is reliable. The deep-learning semantic image segmentation method is suitable for pore recognition of shale SEM images. The fully convolutional neural network model is used to train the manually labeled shale SEM images, which can realize the intelligent recognition and quantitative characterization of the pores in the shale SEM images. It provides a certain reference value for the evaluation of shale oil and gas reservoirs and the study of other porous materials.

Convolutional neural network deep learning model accurately detects rectal cancer in endoanal ultrasounds

BackgroundImaging is vital for assessing rectal cancer, with endoanal ultrasound (EAUS) being highly accurate in large tertiary medical centers. However, EAUS accuracy drops outside such settings, possibly due to varied examiner experience and fewer examinations. This underscores the need for an AI-based system to enhance accuracy in non-specialized centers. This study aimed to develop and validate deep learning (DL) models to differentiate rectal cancer in standard EAUS images.MethodsA transfer learning approach with fine-tuned DL architectures was employed, utilizing a dataset of 294 images. The performance of DL models was assessed through a tenfold cross-validation.ResultsThe DL diagnostics model exhibited a sensitivity and accuracy of 0.78 each. In the identification phase, the automatic diagnostic platform achieved an area under the curve performance of 0.85 for diagnosing rectal cancer.ConclusionsThis research demonstrates the potential of DL models in enhancing rectal cancer detection during EAUS, especially in settings with lower examiner experience. The achieved sensitivity and accuracy suggest the viability of incorporating AI support for improved diagnostic outcomes in non-specialized medical centers.

Detection of coal gangue by YOLO deep learning method based on channel pruning

ABSTRACT Rapid and accurate identification of coal and gangue is of great significance to improve sorting efficiency and reduce environmental pollution. Image recognition based on deep learning is the current research hotspot, but most of the existing deep learning network models for coal gangue recognition are complex in structure and rely on high-performance computing resources, which is difficult to deploy in edge devices. To resolve this problem, based on the YOLOv5s model and on the condition of ensuring the accuracy of coal gangue identification, this paper uses channel pruning and transfer learning technology to realize the lightweight of the coal gangue identification and detection model. Results show that the model size is significantly reduced by this method. The parameter and the amount of computation were reduced by 78.7% and 62%, and the reasoning speed was 6.8 ms per image. The superiority of this model is further verified by comparing it with 7 mainstream target detection networks. The proposed lightweight treatment method of coal gangue identification model effectively reduces the model volume and improves the real-time identification.

Computed tomography-based 3D convolutional neural network deep learning model for predicting micropapillary or solid growth pattern of invasive lung adenocarcinoma

PurposeTo investigate the value of a computed tomography (CT)-based deep learning (DL) model to predict the presence of micropapillary or solid (M/S) growth pattern in invasive lung adenocarcinoma (ILADC).Materials and MethodsFrom June 2019 to October 2022, 617 patients with ILADC who underwent preoperative chest CT scans in our institution were randomly placed into training and internal validation sets in a 4:1 ratio, and 353 patients with ILADC from another institution were included as an external validation set. Then, a self-paced learning (SPL) 3D Net was used to establish two DL models: model 1 was used to predict the M/S growth pattern in ILADC, and model 2 was used to predict that pattern in ≤ 2-cm-diameter ILADC.ResultsFor model 1, the training cohort’s area under the curve (AUC), accuracy, recall, precision, and F1-score were 0.924, 0.845, 0.851, 0.842, and 0.843; the internal validation cohort’s were 0.807, 0.744, 0.756, 0.750, and 0.743; and the external validation cohort’s were 0.857, 0.805, 0.804, 0.806, and 0.804, respectively. For model 2, the training cohort’s AUC, accuracy, recall, precision, and F1-score were 0.946, 0.858, 0.881,0.844, and 0.851; the internal validation cohort’s were 0.869, 0.809, 0.786, 0.794, and 0.790; and the external validation cohort’s were 0.831, 0.792, 0.789, 0.790, and 0.790, respectively. The SPL 3D Net model performed better than the ResNet34, ResNet50, ResNeXt50, and DenseNet121 models.ConclusionThe CT-based DL model performed well as a noninvasive screening tool capable of reliably detecting and distinguishing the subtypes of ILADC, even in small-sized tumors.

Short term power load forecasting system based on improved neural network deep learning model

AbstractThe electricity load prediction is closely related to production and daily life. The electricity load prediction is also a very important task. With the widespread application of smart grids, load data shows an exponential growth trend. The huge amount of data in the load makes power prediction even more difficult. On the basis of traditional prediction algorithms, a power load prediction model based on machine learning and neural networks is designed. Because the single model prediction has the unstable results, a combined model is obtained based on the ensemble learning idea and two single model prediction method. The prediction results are detected by the load data. From the experimental results, the mean absolute percentage error (MAPE) of the AdaBoost‐GRU data fusion model is 0.066%. Compared to the AdaBoost‐GRU data fusion model, the MAPE decreases by 1.59% and 1.12%, respectively. The relative mass scores of the two groups decrease by 132.57% and 89.14%, respectively. The prediction accuracy is improved, which has advantages compared to traditional combination models. It can effectively enhance the accuracy of short‐term power grid load forecasting. It is an important scientific and practical reference for power grid decision‐making.

A Survey of Deep Learning Road Extraction Algorithms Using High-Resolution Remote Sensing Images.

Roads are the fundamental elements of transportation, connecting cities and rural areas, as well as people's lives and work. They play a significant role in various areas such as map updates, economic development, tourism, and disaster management. The automatic extraction of road features from high-resolution remote sensing images has always been a hot and challenging topic in the field of remote sensing, and deep learning network models are widely used to extract roads from remote sensing images in recent years. In light of this, this paper systematically reviews and summarizes the deep-learning-based techniques for automatic road extraction from high-resolution remote sensing images. It reviews the application of deep learning network models in road extraction tasks and classifies these models into fully supervised learning, semi-supervised learning, and weakly supervised learning based on their use of labels. Finally, a summary and outlook of the current development of deep learning techniques in road extraction are provided.

Deep learning-based forecasting model for chlorophyll-a response to tropical cyclones in the Western North Pacific

Tropical cyclones cause increases in sea surface chlorophyll-a concentration, which is important for studying variations in the regional marine environment. Precisely forecasting the variations of sea surface chlorophyll-a concentration induced by tropical cyclones remains a challenge. In this research, a bidirectional long short-term memory (BiLSTM) neural network deep learning model was applied to predict the variations of sea surface chlorophyll-a concentration induced by typhoons in the Western North Pacific (WNP). Typhoons occurring between 2011 and 2020 were used as training cases and those from 2021 to 2022 as forecasting and testing cases. The input variables of the deep learning model include the sea surface wind at 10 meters (U10 and V10), sea surface temperature anomaly (SSTA), and sea surface chlorophyll-a concentration. The output variable was the chlorophyll-a concentration one day after the passage of the typhoon. Data from the previous 7 days were used to predict the chlorophyll-a concentration one day after the typhoon's passage, and the rolling forecast method was employed to predict chlorophyll-a concentration in the following 7 days. To assess the impact of input variables on the model's forecasting performance, ablation experiments were conducted. The results showed that when using U10, V10, and chlorophyll-a from the previous seven days as inputs, the model exhibited the best overall forecasting performance. Taking Typhoons Chanthu, In-fa, and Malou as examples, the root mean square error (RMSE) for the forecast results are 0.0143 mg · m−3, 0.0087 mg · m−3, and 0.0030 mg · m−3, the mean absolute errors (MAE) are 0.0072 mg · m−3, 0.0074 mg · m−3, and 0.0025 mg · m−3, and the spatial anomaly correlation coefficients (ACC) are 0.9968, 0.9775, and 0.9721, respectively. The results reveal that the most accurate forecasting performance was observed during the mid-phase of the moderate-intensity Typhoon Muifa, with RMSE, MAE, and ACC values of 0.0040 mg · m−3, 0.0032 mg · m−3, and 0.9894, respectively. The BiLSTM neural network model had the best forecasting performance for typhoons of moderate intensity and during the mid-term phase. This is because moderate-intensity typhoons or the mature phase of any typhoon tend to have relatively stable and more predictable paths, resulting in better predictions of chlorophyll-a concentrations. In future work, we intend to extend our training and forecasting to typhoons of various intensities, aiming to further refine and enhance predictive performance.