Convolutional Neural Network Method Research Articles

Abnormal behavior detection in industrial control systems based on CNN

With the widespread application of Internet of Things technology in industrial control systems, abnormal behavior detection has become a key task to ensure the safety and stable operation of the system. We propose a multi-branch convolutional fusion neural network method to improve the accuracy and efficiency of abnormal behavior detection in Internet of Things industrial control systems. This method achieves efficient detection of abnormal behavior in video data by combining ResNet152 for spatial feature extraction and GRU for temporal feature extraction. Unlike traditional methods, this method adopts a multi-branch structure, which can simultaneously capture multi-scale feature information, significantly enhancing the richness of feature expression and the accuracy of detection. Experimental results show that on the UCF-Crime dataset, the accuracy of this method reaches 85.76%, which is significantly better than that of traditional methods. In addition, on the larger UCF-101 dataset, the accuracy of this method reaches 92.21%, further verifying its excellent generalization performance. Compared with the C3D network, this method improves the accuracy by nearly 6% while maintaining a high processing speed. These results show that the proposed method has great potential in practical applications.

An innovative method for improving rainfall prediction in Gujarat state through a fusion model DWT, 1DCNN and LSTM

In countries like India where agriculture is a major industry, accurate prediction of rainfall is crucial. The ability to forecast rainfall helps in planning crop cultivation and has a positive impact on both local and national economies. Nevertheless, conventional statistical techniques prove inadequate for forecasting precipitation over extended periods of time owing to the erratic characteristics inherent in climate phenomena. To address this issue, various ML models such as Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), and Discrete Wavelet Transform, have been explored for precipitation prediction. Large-scale data analysis and overcoming intricate challenges have become achievable through the utilization of deep learning, which has proven to be a potent approach. It combines the advantages of multiple methodologies and algorithms in order to develop accurate precipitation forecasting models. In this study, researchers treated rainfall time series as signals and used the Discrete Wavelet Transform method along with Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network methods to analyze and reconstruct these signals. The research employed a dataset acquired from the Indian Meteorological Department and NASA, covering the years 1901 to 2022. This extensive collection comprised diverse meteorological parameters such as relative humidity, wet bulb temperature, Surface Pressure, soil dampness levels, relative humidity, temperatures (min and max), wind speed data in addition to rainfall records. The fusion of these three methods - DWT-CNN-LSTM deep model - consistently demonstrated excellent performance in key evaluation metrics such as MAE and RMSE. This approach achieved remarkable results in accurately predicting precipitation patterns for Gujarat State over the specified time period.

Detection of Typical Transient Signals in Water by XGBoost Classifier Based on Shape Statistical Features: Application to the Call of Southern Right Whale

Whale sound is a typical transient signal. The escalating demands of ecological research and marine conservation necessitate advanced technologies for the automatic detection and classification of underwater acoustic signals. Traditional energy detection methods, which focus primarily on amplitude, often perform poorly in the non-Gaussian noise conditions typical of oceanic environments. This study introduces a classified-before-detect approach that overcomes the limitations of amplitude-focused techniques. We also address the challenges posed by deep learning models, such as high data labeling costs and extensive computational requirements. By extracting shape statistical features from audio and using the XGBoost classifier, our method not only outperforms the traditional convolutional neural network (CNN) method in accuracy but also reduces the dependence on labeled data, thus improving the detection efficiency. The integration of these features significantly enhances model performance, promoting the broader application of marine acoustic remote sensing technologies. This research contributes to the advancement of marine bioacoustic monitoring, offering a reliable, rapid, and training-efficient method suitable for practical deployment.

Multimodal fusion: A study on speech-text emotion recognition with the integration of deep learning

Recognition of various human emotions holds significant value in numerous real-world scenarios. This paper focuses on the multimodal fusion of speech and text for emotion recognition. A 39-dimensional Mel-frequency cepstral coefficient (MFCC) was used as a feature for speech emotion. A 300-dimensional word vector obtained through the Glove algorithm was used as the feature for text emotion. The bidirectional gate recurrent unit (BiGRU) method in deep learning was added for extracting deep features. Subsequently, it was combined with the multi-head self-attention (MHA) mechanism and the improved sparrow search algorithm (ISSA) to obtain the ISSA-BiGRU-MHA method for emotion recognition. It was validated on the IEMOCAP and MELD datasets. It was found that MFCC and Glove word vectors exhibited superior recognition effects as features. Comparisons with the support vector machine and convolutional neural network methods revealed that the ISSA-BiGRU-MHA method demonstrated the highest weighted accuracy and unweighted accuracy. Multimodal fusion achieved weighted accuracies of 76.52 %, 71.84 %, 66.72 %, and 62.12 % on the IEMOCAP, MELD, MOSI, and MOSEI datasets, suggesting better performance than unimodal fusion. These results affirm the reliability of the multimodal fusion recognition method, showing its practical applicability.

Automatic lip-reading classification using deep learning approaches and optimized quaternion meixner moments by GWO algorithm

Lip-reading classification has received a lot of interest in recent decades because it is widely used in a variety of fields. It plays an important role in interpreting spoken words in noisy situations and reconstructing communication processes for those with hearing impairments. Despite significant advancements in this field, there are still several drawbacks in existing work such as feature extraction and Model capability for visual speech recognition. For these reasons, the current paper suggests an Optimized Quaternion Meixner Moments Convolutional Neural Network (OQMMs-CNN) method that intends to develop a Visual Speech Recognition (VSR) system based only on video images. This unique method combines OQMMs optimized for the GWO algorithm and convolutional neural networks taken from deep learning techniques with the aim of recognizing digits, words, or letters displayed as input videos.The OQMMs are used here as descriptors with the purpose of identifying, holding, and extracting essential information from video images (lips images) and generating moments for CNN input. The latter uses Meixner polynomials, which are defined by local parameters α and β. Then, the Grey Wolf optimization method (GWO) is applied to enssure excellent classification accuracy by optimizing those local parameters. After being tested on three public datasets such as AVLetters, Grid, AVDigits, and LRW, and comparing to several ways using complicated models and deep architecture, the method emerges as an excellent solution for reducing the high dimensionality of video pictures and training time.

Penerapan Algoritma Convolutional Neural Network Pada Identifikasi Foto Hero Game Mobile Legends

This study aims to develop a hero photo identification system in the Mobile Legends game using the Convolutional Neural Network (CNN) algorithm. The background of this study is based on the need to speed up and simplify the hero identification process which often requires time and special knowledge. The CNN algorithm was chosen because of its superior ability in pattern recognition and image classification. The research method includes several stages, namely collecting Mobile Legends hero image data, data preprocessing including resizing and normalization, and dividing the data into training, validation, and testing data. The CNN model used consists of several convolution and pooling layers for feature extraction, and a fully connected layer for the final classification. Model training is carried out using a dataset that has been processed with augmentation techniques to increase data variation. The results of this study are that the data used amount to 600 image data divided into 30 classes. By implementing the CNN method, researchers have succeeded in creating a system that can recognize images of mobile legends heroes. Based on the scenario created by the researcher, the highest accuracy is a combination of ReLu activation, Dropout 0.2 and using epoch 20, resulting in an accuracy of 62.17%.

Remora‐CNN: A Novel and Effective Method for Rice Leaf Disease Detection and Classification

ABSTRACTFor millions of people worldwide, rice is one of the main food crops. Nevertheless, while being grown, rice is susceptible to many diseases. Most rice plant diseases are influenced by biotic and abiotic factors, including nematodes, viroids, fungus, viruses, bacteria, and other microorganisms, as well as temperature and other environmental factors. Thus, an automatic early classification of leaf disease is necessary to improve the rice yield. In this paper, for identifying and categorizing the rice leaf disease, a convolutional neural network (CNN) model is used, and the CNN is trained using the Remora Optimization Algorithm (ROA). A better classification outcome is attained by performing the segmentation process using K‐means with the Fractional Tangential‐Spherical Kernel (FTSK) algorithm. Furthermore, the developed Remora Optimization‐ Convolutional Neural Network (Remora‐CNN) method achieved the optimal performance based on the testing accuracy, sensitivity and specificity of 0.925, 0.931, and 0.941 using the Rice Leaf Disease Image Samples Dataset.

Defect Recognition and Classification Using Ultrasonic Scattered Signals Based on Convolutional Neural Network

Abstract Carbon Fiber Reinforced Polymer (CFRP) is widely used in the aerospace, aircraft, and railway industries because of its remarkable advantages, such as its lightweight and high strength. Detecting and identifying defects in CFRP is crucial for ensuring its secure and efficient use. Ultrasonic non-destructive testing is one of the most practical and effective methods for defect detection and identification. The ultrasonic scattered signals of defects, which contain abundant defect feature information, are advantageous for defect recognition and classification. However, the anisotropy and inhomogeneity of CFRP, which lead to a low signal-to-noise ratio and poor image quality, pose challenges in flaw recognition. To address the issue, this paper proposes using Convolutional Neural Networks (CNN) methods for defect classification, which can adequately extract information from ultrasonic scattered signals directly. The scattering signals from various types of defects in CFRP, including the different sizes of holes and folds, were selected as inputs for CNN models. Different network models combining Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) with 1D-CNN were employed to classify and identify defect signals. To generate sufficient input sample signals, the data sets were further enlarged by incorporating noise into the original signals. The experimental results showed that using the 1D-LSTM-CNN and 1D-GRU-CNN network models can reduce computational resources while ensuring the accuracy of defect recognition.

Sunspot Group Detection and Classification by Dual Stream Convolutional Neural Network Method

The automatic detection and analysis of sunspots play a crucial role in understanding solar dynamics and predicting space weather events. This paper proposes a novel method for sunspot group detection and classification called the dual stream Convolutional Neural Network with Attention Mechanism (DSCNN-AM). The network consists of two parallel streams each processing different input data allowing for joint processing of spatial and temporal information while classifying sunspots. It takes in the white light images as well as the corresponding magnetic images that reveal both the optical and magnetic features of sunspots. The extracted features are then fused and processed by fully connected layers to perform detection and classification. The attention mechanism is further integrated to address the “edge dimming” problem which improves the model’s ability to handle sunspots near the edge of the solar disk. The network is trained and tested on the SOLAR-STORM1 data set. The results demonstrate that the DSCNN-AM achieves superior performance compared to existing methods, with a total accuracy exceeding 90%.

Integration of LIME Explainable AI to Enhance Interpretability of Deep Learning Models in Box Palette Classification

In the food production system, manual box palette arrangement often encounters errors, such as incorrect stacking patterns, mismatched box quantities, and improper mixing of product variants. This issue occurs in a soy sauce production company with limited infrastructure, leading to the mixing of two product variants with the same box size in a single production line. This results in disruptions to the production flow and significant potential losses. This study proposes a solution using deep learning to detect box palette arrangement patterns. The Convolutional Neural Network (CNN) method is chosen because it has proven effective in image classification. Additionally, this study implements Explainable AI (XAI) to provide explanations related to classification results, increasing user confidence in the system. The Local Interpretable Model-agnostic Explanations (LIME) technique will be used to provide interpretations. This research produces the development of a deep learning model to classify box palette arrangements. Furthermore, the implementation of LIME in this study successfully provides interpretations of the model's predictions. The result is evident in the result where it shows that MobileNetV2 give an F1-Score of 100%, and LIME fidelity score of 0.2 and stability score of 0.2.

Customized Convolutional Neural Network for Glaucoma Detection in Retinal Fundus Images

Glaucoma is one of the leading causes of permanent blindness and remains a current challenge in the field of ophthalmology. This research aims to present a comprehensive investigation into the development and evaluation of new technology for glaucoma detection in retinal fundus images. The development and evaluation are presented on a customized architecture, using the Convolutional Neural Network (CNN) method. The proposed CNN architecture is designed to address the complex characteristics of glaucoma changes in the identification process. The research dataset consists of 506 retinal images categorized into 117 glaucoma images, 19 suspected glaucoma images, and 370 healthy images. Through our in-depth exploration, we conducted a careful analysis to uncover patterns and fundamental trends related to glaucoma-related features. During the training phase, the proposed CNN achieved outstanding average accuracy, sensitivity, and specificity values of 92.88%, 94.66%, and 89.31%, respectively. In the unseen test dataset, the model demonstrated competitive performance with an accuracy of 80.87%, sensitivity of 85.65%, and specificity of 71.26%. These findings emphasize the potential of the model as a reliable tool for glaucoma detection. The results indicate that the proposed method utilizing a customized CNN architecture is designed for glaucoma detection in retinal fundus images. The presented output results also hold promise for clinical relevance and can be considered an improvement in the care of retinal fundus patients.

Spatiotemporal prediction of surface roughness evolution of C/C composites based on recurrent neural network

Carbon/carbon (C/C) composites are widely used in aerospace applications due to their excellent properties. Their surface roughness is an important factor affecting the material properties and ablation resistance, which is essential for the accurate prediction of the material. Images are often used to summarize and characterize microstructural features, which include information such as fiber microstructure and fiber orientations. An ablation shape evolution model based on optimization of attention mechanism, which combines multi-scale convolutional neural network (MSCNN) and long short-term memory (LSTM) methods for the C/C composites, is proposed. MSCNN encodes the microstructural information of the ablation image, which is fed into the LSTM model to extract the spatiotemporal features and long-term dependencies latent in the temporal law. The attention mechanism is used to assign different weights to the extracted features which makes the model focus more on the important time steps in the sequence. The results show that the proposed model achieves excellent effects in predicting the evolution of ablative morphology, with high agreement with the simulated values. The method saves time and cost, while providing an efficient and reliable method for the design and optimization of thermal protection materials.

Resnet50 and logistic Gaussian map-based zero-watermarking algorithm for medical color images

Medical image copyright protection is becoming increasingly relevant as medical images are used more frequently in medical networks and institutions. The traditional embedded watermarking system is inappropriate for medical images since it degrades the original images’ quality. Furthermore, medical-colored image watermarking options are constrained since most medical watermarking systems are built for gray-scale images. This paper proposes a zero-watermarking scheme for medical color image copyright protection based on a chaotic system and Resnet50, which is a convolutional neural network method. The network Resnet50 is used to extract features from the color medical image, and then a logistic Gaussian map is used to scramble these features and scramble the binary image. Finally, an exclusive OR operation is performed (scrambled binary image, scrambled features for the medical color image) to form a zero watermarking. The experimental result proves that our scheme is effective and robust to geometric and common image processing attacks. The BER values of the extracted watermarks are below 0.0039, and the NCC values are above 0.9942, while the average PSNR values of the attacked images are 29.0056 dB. Also, it is superior to other zero-watermark schemes for medical images in terms of robustness to conventional image processing and geometric attacks. Furthermore, the experimental results show that the Resnet50 model outperforms other models in terms of reducing the mean squared errors of the features between the attacked and original image.

Unsupervised deep learning bias correction of CMIP6 global ensemble precipitation predictions with cycle generative adversarial network

Abstract Climate change significantly impacts agricultural production, ecosystem stability, and socioeconomic development. Global climate models (GCMs) serve as the primary tool for simulating historical and future precipitation patterns. However, due to issues such as coarse resolution, boundary condition, and parameterization, model outputs require bias correction (BC). With the evolution of deep learning techniques, supervised convolutional neural network (CNN) frameworks have gained popularity in the area of climate model BC but face limitations in spatial correlation assumptions and data sparsity, particularly for extreme precipitation This study proposed an unsupervised learning approach using cycle generative adversarial network (CycleGAN) to correct the ensemble mean bias of models and compare its performance with CNN and Quantile Mapping methods. The results demonstrate that the proposed CycleGAN approach outperforms both CNN and Quantile Mapping in ensemble mean BC. It effectively learns the overall distribution of precipitation through an adversarial process and yields better extreme precipitation predictions.

An attribution graph-based interpretable method for CNNs

Convolutional Neural Networks (CNNs) have demonstrated outstanding performance in various domains, such as face recognition, object detection, and image segmentation. However, the lack of transparency and limited interpretability inherent in CNNs pose challenges in fields such as medical diagnosis, autonomous driving, finance, and military applications. Several studies have explored the interpretability of CNNs and proposed various post-hoc interpretable methods. The majority of these methods are feature-based, focusing on the influence of input variables on outputs. Few methods undertake the analysis of parameters in CNNs and their overall structure. To explore the structure of CNNs and intuitively comprehend the role of their internal parameters, we propose an Attribution Graph-based Interpretable method for CNNs (AGIC) which models the overall structure of CNNs as graphs and provides interpretability from global and local perspectives. The runtime parameters of CNNs and feature maps of each image sample are applied to construct attribution graphs (At-GCs), where the convolutional kernels are represented as nodes and the SHAP values between kernel outputs are assigned as edges. These At-GCs are then employed to pretrain a newly designed heterogeneous graph encoder based on Deep Graph Infomax (DGI). To comprehensively delve into the overall structure of CNNs, the pretrained encoder is used for two types of interpretable tasks: (1) a classifier is attached to the pretrained encoder for the classification of At-GCs, revealing the dependency of At-GC’s topological characteristics on the image sample categories, and (2) a scoring aggregation (SA) network is constructed to assess the importance of each node in At-GCs, thus reflecting the relative importance of kernels in CNNs. The experimental results indicate that the topological characteristics of At-GC exhibit a dependency on the sample category used in its construction, which reveals that kernels in CNNs show distinct combined activation patterns for processing different image categories, meanwhile, the kernels that receive high scores from SA network are crucial for feature extraction, whereas low-scoring kernels can be pruned without affecting model performance, thereby enhancing the interpretability of CNNs.

MetalTrans: A Biological Language Model-Based Approach for Predicting Disease-Associated Mutations in Protein Metal-Binding Sites.

The critical importance of accurately predicting mutations in protein metal-binding sites for advancing drug discovery and enhancing disease diagnostic processes cannot be overstated. In response to this imperative, MetalTrans emerges as an accurate predictor for disease-associated mutations in protein metal-binding sites. The core innovation of MetalTrans lies in its seamless integration of multifeature splicing with the Transformer framework, a strategy that ensures exhaustive feature extraction. Central to MetalTrans's effectiveness is its deep feature combination strategy, which merges evolutionary-scale modeling amino acid embeddings with ProtTrans embeddings, thus shedding light on the biochemical properties of proteins. Employing the Transformer component, MetalTrans leverages the self-attention mechanism to delve into higher-level representations. Utilizing mutation site information for feature fusion not only enriches the feature set but also sidesteps the common pitfall of overestimation linked to protein sequence-based predictions. This nuanced approach to feature fusion is a key differentiator, enabling MetalTrans to outperform existing methods significantly, as evidenced by comparative analyses. Our evaluations across varied metal binding site data sets (specifically Zn, Ca, Mg, and Mix) underscore MetalTrans's superior performance, which achieved the average AUC values of 0.971, 0.965, 0.980, and 0.945 on multiple 5-fold cross-validation, respectively. Remarkably, against the multichannel convolutional neural network method on a benchmark independent test set, MetalTrans demonstrated unparalleled robustness and superiority, boasting the AUC score of 0.998 on multiple 5-fold cross-validation. Our comprehensive examination of the predicted outcomes further confirms the effectiveness of the model. The source codes, data sets, and prediction results for MetalTrans can be accessed for academic usage at https://github.com/EduardWang/MetalTrans.

Collaborative representation based unsupervised CNN for hyperspectral anomaly detection

The convolutional neural networks (CNNs) have shown high success in supervised analysis of hyperspectral images. But, the use of a supervised CNN is not possible for an unsupervised task such as hyperspectral anomaly detection. So, an unsupervised CNN with pre-determined convolutional kernels without requirement to labeled samples or training process is proposed in this work. The proposed method uses the collaborative representation (CR) for background estimate and introduces the global preserving projection (GPP) for dimensionality reduction of it. Then, the convolutional kernels are randomly selected from the reduced CR data. Moreover, two distances in inner and guard windows are defined, which difference of them results in the anomaly score. The CR based unsupervised CNN (CUCNN) method achieves high detection accuracy compared to its counterparts and is more than 9 times faster than other presented unsupervised CNN detectors.

Revolutionizing malaria diagnosis: deep learning-powered detection of parasite-infected red blood cells

Malaria is a significant global health issue, responsible for the highest rates of morbidity and mortality globally. This paper introduces a very effective and precise convolutional neural network (CNN) method that employs advanced deep learning techniques to automate the detection of malaria in images of red blood cells (RBC). Furthermore, we present an emerging and efficient deep learning method for differentiating between cells infected with malaria and those that are not infected. To thoroughly evaluate the efficiency of our approach, we do a meticulous assessment that involves comparing different deep learning models, such as ResNet-50, MobileNet-v2, and Inception-v3, within the domain of malaria detection. Additionally, we conduct a thorough comparison of our proposed approach with current automated methods for malaria identification. An examination of the most current techniques reveals differences in performance metrics, such as accuracy, specificity, sensitivity, and F1 score, for diagnosing malaria. Moreover, compared to existing models for malaria detection, our method is the most successful, achieving an accurate score of 1.00 in all statistical matrices, confirming its promise as a highly efficient tool for automating malaria detection.

Generative adversarial network for Multimodal Contrastive Domain Sharing based on efficient invariant feature-centric growth analysis improved brain tumor classification

ABSTRACT Efficient and accurate classification of brain tumor categories remains a critical challenge in medical imaging. While existing techniques have made strides, their reliance on generic features often leads to suboptimal results. To overcome these issues, Multimodal Contrastive Domain Sharing Generative Adversarial Network for Improved Brain Tumor Classification Based on Efficient Invariant Feature Centric Growth Analysis (MCDS-GNN-IBTC-CGA) is proposed in this manuscript.Here, the input imagesare amassed from brain tumor dataset. Then the input images are preprocesssed using Range – Doppler Matched Filter (RDMF) for improving the quality of the image. Then Ternary Pattern and Discrete Wavelet Transforms (TPDWT) is employed for feature extraction and focusing on white, gray mass, edge correlation, and depth features. The proposed method leverages Multimodal Contrastive Domain Sharing Generative Adversarial Network (MCDS-GNN) to categorize brain tumor images into Glioma, Meningioma, and Pituitary tumors. Finally, Coati Optimization Algorithm (COA) optimizes MCDS-GNN’s weight parameters. The proposed MCDS-GNN-IBTC-CGA is empirically evaluated utilizing accuracy, specificity, sensitivity, Precision, F1-score,Mean Square Error (MSE). Here, MCDS-GNN-IBTC-CGA attains 12.75%, 11.39%, 13.35%, 11.42% and 12.98% greater accuracy comparing to the existingstate-of-the-arts techniques, likeMRI brain tumor categorization utilizing parallel deep convolutional neural networks (PDCNN-BTC), attention-guided convolutional neural network for the categorization of braintumor (AGCNN-BTC), intelligent driven deep residual learning method for the categorization of braintumor (DCRN-BTC),fully convolutional neural networks method for the classification of braintumor (FCNN-BTC), Convolutional Neural Network and Multi-Layer Perceptron based brain tumor classification (CNN-MLP-BTC) respectively.

A convolutional neural networks method for tropospheric ozone vertical distribution retrieval from Multi-AXis Differential Optical Absorption Spectroscopy measurements

The vertical distribution of tropospheric ozone (O3) is crucial for understanding atmospheric physicochemical processes. A Convolutional Neural Networks (CNN) method for the retrieval of tropospheric O3 vertical distribution from ground-based Multi-AXis Differential Optical Absorption Spectroscopy (MAX-DOAS) measurements to tackle the issue of stratospheric O3 absorption interference faced by MAX-DOAS in obtaining tropospheric O3 profiles. Firstly, a hybrid model, named PCA-F_Regression-SVR, is developed to screen features sensitive to O3 inversion based on the MAX-DOAS spectra and EAC4 reanalysis O3 profiles, which incorporates Principal Component Analysis (PCA), F_Regression function, and Support Vector Regression (SVR) algorithm. Thus, these screened features for ancillary inversion include the profiles of temperature, specific humidity, fraction of cloud coverage, eastward and northward wind, the profiles of SO2, NO2, and HCHO, as well as season and time features to serve as sensitive factors. Secondly, the preprocessed MAX-DOAS spectra dataset and the sensitive factor dataset are utilized as input, while the O3 profiles of the EAC4 reanalysis dataset incorporating the surface O3 concentrations are employed as output for constructing the CNN model. The Mean Absolute Percentage Error (MAPE) decreases from 26 % to approximately 19 %. Finally, the CNN model is applied for inversion and comparison of tropospheric O3 profiles using independent input data. The CNN model effectively reproduces the O3 profiles of the EAC4 dataset, showing a Gaussian-like spatial distribution with peaks primarily around 950 hPa (550 m). Since the reanalysis data used for model training has been smoothed, the CNN model is insensitive to extreme values. This behavior can be attributed to the MAPE loss function, which evaluates Absolute Percentage Errors (APEs) of O₃ concentration at all altitudes, resulting in varying retrieval accuracy across different altitudes while maintaining overall MAPE control. Temporally, the CNN model tends to overestimate surface O3 in summer by around 20 μg/m3, primarily due to the influence of the temperature feature in the sensitivity factor dataset. In conclusion, leveraging MAX-DOAS spectra enables the retrieval of tropospheric O3 vertical distribution through the established CNN model.