Classification Model Research Articles

A dual-path model merging CNN and RNN with attention mechanism for crop classification

Rapid and accurate crop classification is essential for estimating crop information and improving cropland management. The application of deep learning models for crop classification using time-series data has become the most promising method. However, most approaches rely on single models for data processing result in lower classification accuracy and poor stability. Therefore, this study proposes a dual-path approach with attention mechanisms (DPACR) to promote the performance of this model architecture in crop classification using time series data. Specifically, the model comprises two branches, the Recurrent neural network (RNN) branch with bidirectional gated recurrent units (GRU) with a self-attention mechanism, and the convolutional neural network (CNN) branch based on SE-ResNet. Crop classification is accomplished by a main classifier, supported by auxiliary classifiers from the two branches. Using the Google Earth Engine and the Sentinel-2 satellite data, DPACR was tested in the Hetao irrigation district in Inner Mongolia, China. The comparison experiment demonstrated that the DPACR achieved the highest overall accuracy (OA = 0.959) and Kappa coefficient (Kappa = 0.941) compared to other five models (MLP, SE-ResNet, Bi-At-GRU, SVM, and RF). DPACR excelled in classifying six crops, maintaining high accuracy across multiple classes. Compared to attention mechanisms, auxiliary classifiers can significantly improve classification performance. This study highlights the effective combination of cloud computing and deep learning for large-scale crop classification, providing a practical method for agricultural monitoring and management.

Artificial intelligence in ovarian cancer drug resistance advanced 3PM approach: subtype classification and prognostic modeling

Artificial intelligence in ovarian cancer drug resistance advanced 3PM approach: subtype classification and prognostic modeling

A Lightweight Multi-Label Classification Method for Urban Green Space in High-Resolution Remote Sensing Imagery

Urban green spaces are an indispensable part of the ecology of cities, serving as the city’s “purifier” and playing a crucial role in promoting sustainable urban development. Therefore, the refined classification of urban green spaces is an important task in urban planning and management. Traditional methods for the refined classification of urban green spaces heavily rely on expert knowledge, often requiring substantial time and cost. Hence, our study presents a multi-label image classification model based on MobileViT. This model integrates the Triplet Attention module, along with the LSTM module, to enhance its label prediction capabilities while maintaining its lightweight characteristic for standalone operation on mobile devices. Trial outcomes in our UGS dataset in this study demonstrate that the approach we used outperforms the baseline by 1.64%, 3.25%, 3.67%, and 2.71% in mAP,F1,precision, and recall, respectively. This indicates that the model can uncover the latent dependencies among labels to enhance the multi-label image classification device’s performance. This study provides a practical solution for the intelligent and detailed classification of urban green spaces, which holds significant importance for the management and planning of urban green spaces.

Generative Adversarial Networks for SAR Automatic Target Recognition and Classification Models Enhanced Explainability: Perspectives and Challenges

Generative adversarial networks (or GANs) are a specific deep learning architecture often used for different usages, such as data generation or image-to-image translation. In recent years, this structure has gained increased popularity and has been used in different fields. One area of expertise currently in vogue is the use of GANs to produce synthetic aperture radar (SAR) data, and especially expand training datasets for SAR automatic target recognition (ATR). In effect, the complex SAR image formation makes these kind of data rich in information, leading to the use of deep networks in deep learning-based methods. Yet, deep networks also require sufficient data for training. However, contrary to optical images, we generally do not have a substantial number of available SAR images because of their acquisition and labelling cost; GANs are then an interesting tool. Concurrently, how to improve explainability for SAR ATR deep neural networks and how to make their reasoning more transparent have been increasingly explored as model opacity deteriorates trust of users. This paper aims at reviewing how GANs are used with SAR images, but also giving perspectives on how GANs could be used to improve interpretability and explainability of SAR classifiers.

Preoperative Prediction of Muscle Invasiveness in Bladder Cancer: The Role of 3D Volumetric Radiomics Using Diffusion-Weighted MRI, the VI-RADS Score, or a Combination of Both.

Bladder cancer treatment decisions hinge on detecting muscle invasion. The 2018 "Vesical Imaging Reporting and Data System" (VI-RADS) standardizes multiparametric MRI (mp-MRI) use. Radiomics, an analysis framework, provides more insightful information than conventional methods. To determine how well MIBC (Muscle Invasive Bladder Cancer) and NMIBC (Non-Muscle Invasive Bladder Cancer) can be distinguished using mp-MRI radiomics features. We conducted a study with 73 bladder cancer patients diagnosed pathologically, who underwent preoperative mp-MRI from January 2020 to July 2022. Utilizing 3D Slicer (version 4.8.1) and Pyradiomics, we manually extracted radiomic features from apparent diffusion coefficient (ADC) maps created from diffusion-weighted imaging. The LASSO approach identified optimal features, and we addressed sample imbalance using SMOTE. We developed a classification model using textural features alone or combined with VI-RADS, employing a random forest classifier with 10-fold cross-validation. Diagnostic performance was assessed using the area under the ROC curve analysis. Among 73 patients (63 men, 10 women; median age: 63 years), 41 had muscle-invasive and 32 had superficial bladder cancer. Muscle invasion was observed in 25 of 41 patients with VI-RADS 4 and 5 scores and 12 of 32 patients with VI-RADS 1, 2, and 3 scores (accuracy: 77.5%, sensitivity: 67.7%, specificity: 88.8%). The combined VI-RADS score and radiomics model (AUC = 0.92±0.12) outperformed the single radiomics model using ADC MRI (AUC = 0.83±0.22 with 10-fold cross-validation) in this dataset. Before undergoing surgery, bladder cancer invasion in muscle might potentially be predicted using a radiomics signature based on mp-MRI.

Development of Predictive Classification Models and Extraction of Signature Wavelengths for the Identification of Spoilage in Chicken Breast Fillets During Storage Using Near Infrared Spectroscopy

Development of Predictive Classification Models and Extraction of Signature Wavelengths for the Identification of Spoilage in Chicken Breast Fillets During Storage Using Near Infrared Spectroscopy

Serum lipidomic profiling of dairy calves fed milk replacers containing animal or vegetable fats

Vegetable fat blends are commonly used as fat sources in milk replacers (MR) for calves, but their composition differs considerably from that of bovine milk fat. The aim of this study was to investigate the serum lipid profile of pre-weaned calves fed twice-daily MR containing 30% fat (% DM). Upon arrival, 30 male Holstein-Friesian calves (BW = 45.6 ± 4.0 kg, age = 2.29 ± 0.8 d) were randomly assigned to 2 experimental diets (n = 15 per treatment): one MR was derived from either vegetable fats (VG; 80% rapeseed and 20% coconut fats) or animal fats (AN; 65% Packer's lard and 35% dairy cream). The 2 MR formulas contained 30% fat, 24% CP, and 36% lactose. Calves were housed indoors in individual pens with ad libitum access to chopped straw and water. Daily milk allowances were 6.0 L from d 1 to 5, 7.0 L from d 6 to 9, and 8.0 L from d 10 to 35, divided into 2 equal meals and prepared at 13.5% solids. An untargeted liquid chromatography-quadrupole time-of-flight mass spectrometry (LC-QTOF-MS) method was employed to analyze the lipid profiles in the serum of calves sampled from the jugular vein at 35 d of age. In total, 594 lipids were characterized, comprising 25 different lipid classes. Principal component analysis (PCA) showed significant separation between VG and AN, indicating different lipid profiles in the serum. An orthogonal partial least squares discriminant analysis (OPLS-DA) classification model was used to further validate the distinction between the 2 treatment groups. The model exhibited a robust class separation and high predictive accuracy. Using a Volcano plot (fold change threshold ≥1.5 and false discovery rate ≤0.05), it was observed that calves fed AN had higher levels of 39 lipid species in serum than calves fed VG, whereas 171 lipid species were lower in the AN group. Lipid classes, such as phosphatidylcholine (PC), phosphatidylethanolamine (PE), sphingomyelin (SM), triglycerides (TG), lysophosphatidylcholine (LPC), and lysophosphatidylethanolamine (LPE), were different. In particular, PC and PE were observed at lower levels in calves fed AN, possibly indicating shifts in cell membrane characteristics, intracellular signaling, and liver functions. In addition, a decrease in certain triglyceride (TG) species was observed in calves fed AN, including a decrease in TG species such as TG 36:0 and TG 38:0, possibly related to variations in the content of certain fatty acids (FA) within the AN MR, such as C10:0, C12:0, C14:0, and C18:0 compared with the VG MR. Calves fed AN had lower levels of LPC and LPE, and lyso-phosphatidylinositol (LPI), SM, and phosphatidylinositol (PI) species than calves fed VG, suggesting shifts in lipoprotein and lipid metabolic pathways. In conclusion, these results deepen the understanding of how lipid sources in MR can modulate the serum lipidome profiles of dairy calves.

Learning Effective Good Variables from Physical Data

We assume that a sufficiently large database is available, where a physical property of interest and a number of associated ruling primitive variables or observables are stored. We introduce and test two machine learning approaches to discover possible groups or combinations of primitive variables, regardless of data origin, being it numerical or experimental: the first approach is based on regression models, whereas the second on classification models. The variable group (here referred to as the new effective good variable) can be considered as successfully found when the physical property of interest is characterized by the following effective invariant behavior: in the first method, invariance of the group implies invariance of the property up to a given accuracy; in the other method, upon partition of the physical property values into two or more classes, invariance of the group implies invariance of the class. For the sake of illustration, the two methods are successfully applied to two popular empirical correlations describing the convective heat transfer phenomenon and to the Newton’s law of universal gravitation.

DATACORTECH: artificial intelligence platform for the virtual screen of aluminum corrosion inhibitors

The machine learning framework reported herein can greatly accelerate the development of more effective and sustainable corrosion inhibitors for aluminum alloys, which still rely mostly on the experience of corrosion scientists, and trial and error laboratory testing. It can be used to design inhibitors for specific applications, which can be immobilized into nanocontainers or included directly into coatings in the search for less hazardous corrosion protective technologies. Therefore, a machine learning (ML) classification model that allows to identify promising compounds ( > 70% inhibitor efficiency) among less promising ones, and an online application (https://datacor.shinyapps.io/datacortech/) were developed for the virtual screen (simulation) of potential inhibitors for aluminum alloys, capable of considering the molecular structure and the influence of pH as an input.

RNDDNet: A residual nested dilated DenseNet based deep-learning model for chilli plant disease classification

The most significant peril to food safety arises from plant diseases, capable of substantially diminishing both the quantity and quality of agricultural yields. Identifying these plant diseases stands out as the foremost challenge within the agricultural sector. Convolutional and deep neural networks prove effective in resolving image classification challenges within the realm of computer vision. Numerous Deep Neural Network(DNN)-based structures have been employed to diagnose plant diseases. Many DNN models in the field make use of various iterations of Dense and DenseNet layers in order to enhance the receptive field and capture intricate features within the data. However, it is important to note that such models often come with a significant computational burden and can introduce aliasing artifacts due to their complexity and resource-intensive nature. To overcome those limitations, we proposed a novel Residual Nested Dilated DenseNet based deep-learning (RNDDNet) model in this paper. Residual Nested Dilated DenseNet model residual connections are achieving the required receptive field, and their dilation factors are effective in extracting more features. The RNDDNet model exhibits the highest level of accuracy in identifying plant diseases. This research introduces a less computational cost and compact model to detect diseases in plant leaves. The proposed model functions to identify diseases, utilizing a dataset comprising 3,800 photographs of chilli leaves, categorized into six distinct classes: five disorder classes and one healthy chilli class. Through experimentation, the outcomes indicate that the suggested model achieves an accuracy of 98.09 %, along with a precision of 97 %, a recall of 97.25 %, and an F1 score of 97.25%. The presented approach demonstrates its superiority over existing methodologies.

A Comparative Study on Data Balancing Methods for Alzheimer's Disease Classification

Alzheimer's disease is a prevalent neurological disorder affecting millions of people worldwide, often associated with the aging process, leading to the death of nerve cells in the brain and loss of connections. Recently, promising results have been demonstrated in diagnosing Alzheimer's disease using deep learning models, and various approaches for early diagnosis have been proposed. However, the imbalance in health datasets, particularly those containing rare cases, can lead to performance losses and misleading results during model training. This study focuses on these imbalance issues, evaluating the effectiveness of different balancing methods using the Alzheimer's MRI dataset. In this context, the performance of SMOTE, ADASYN, and Weight Balancing methods is compared using a custom model. Experimental results indicate that, compared to the original imbalanced dataset, Weight balancing outperforms in terms of accuracy, precision, recall, and F1 score. While SMOTE and ADASYN show improvement in various metrics, they are considered inferior to the Weight Balancing method. This study contributes to selecting data-balancing methods to enhance the accuracy of deep learning models in Alzheimer's disease classification and emphasizes the importance of addressing class imbalances in health datasets.

The use of image classification to estimate flamingo abundance from aerial, drone and satellite imagery

The Lesser Flamingo Phoeniconaias minor is a Near Threatened species known to be highly gregarious and that can concentrate in large numbers at core foraging and breeding sites, yet is also known to disperse widely in search of suitable foraging habitat. The density of flamingos at occupied sites, together with frequent dispersal between foraging sites, poses significant challenges to obtaining reliable counts and associated population estimates. Our study tested the use of unmanned aerial vehicles (drones) imagery and high-resolution satellite imagery for producing density estimates of Lesser Flamingos at the only known breeding site in South Africa (Kamfers Dam, Northern Cape Province). We implemented a supervised classification approach using multiple supervised classification algorithms, including xgbTree, randomForest, nnet and GBM. The four supervised classification models performed very well at classifying flamingos from both drone and satellite imagery, with classification accuracies ranging between 0.98 and 0.99. The inclusion of an additional predictor (other than RGB bands) yielded the highest variable importance score and greatly increased model accuracy and predictive performance. Estimates of adult flamingos from drone imagery acquired across nesting mudflats on 5 February (n = 1 320) and 14 February (n = 330) 2019 suggested a 75% decline in incubating adults. Furthermore, a total of 16 134 (15 973–16 295) flamingos were estimated from drone imagery on 1 March 2019, and 17 291 (17 118–17 290) flamingos from satellite imagery on 19 March 2019. The modelled size of the creche across dates suggested a loss of 234 birds (4.5%) during the period of assessment, with the highest count estimating 5 218 juveniles. Our study successfully implemented machine learning classifiers that were able to detect and derive highly accurate counts of Lesser Flamingos from both drone and satellite imagery. This use of machine learning could greatly improve the efficiency of monitoring and associated conservation efforts underway for the Lesser Flamingo.

The use of image classification to estimate flamingo abundance from aerial, drone and satellite imagery

The Lesser Flamingo Phoeniconaias minor is a Near Threatened species known to be highly gregarious and that can concentrate in large numbers at core foraging and breeding sites, yet is also known to disperse widely in search of suitable foraging habitat. The density of flamingos at occupied sites, together with frequent dispersal between foraging sites, poses significant challenges to obtaining reliable counts and associated population estimates. Our study tested the use of unmanned aerial vehicles (drones) imagery and high-resolution satellite imagery for producing density estimates of Lesser Flamingos at the only known breeding site in South Africa (Kamfers Dam, Northern Cape Province). We implemented a supervised classification approach using multiple supervised classification algorithms, including xgbTree, randomForest, nnet and GBM. The four supervised classification models performed very well at classifying flamingos from both drone and satellite imagery, with classification accuracies ranging between 0.98 and 0.99. The inclusion of an additional predictor (other than RGB bands) yielded the highest variable importance score and greatly increased model accuracy and predictive performance. Estimates of adult flamingos from drone imagery acquired across nesting mudflats on 5 February (n = 1 320) and 14 February (n = 330) 2019 suggested a 75% decline in incubating adults. Furthermore, a total of 16 134 (15 973–16 295) flamingos were estimated from drone imagery on 1 March 2019, and 17 291 (17 118–17 290) flamingos from satellite imagery on 19 March 2019. The modelled size of the creche across dates suggested a loss of 234 birds (4.5%) during the period of assessment, with the highest count estimating 5 218 juveniles. Our study successfully implemented machine learning classifiers that were able to detect and derive highly accurate counts of Lesser Flamingos from both drone and satellite imagery. This use of machine learning could greatly improve the efficiency of monitoring and associated conservation efforts underway for the Lesser Flamingo.

A method of identifying domain-specific academic user information needs based on academic Q&A communities

Purpose To address the shortcomings of existing academic user information needs identification methods, such as low efficiency and high subjectivity, this study aims to propose an automated method of identifying online academic user information needs. Design/methodology/approach This study’s method consists of two main parts: the first is the automatic classification of academic user information needs based on the bidirectional encoder representations from transformers (BERT) model. The second is the key content extraction of academic user information needs based on the improved MDERank key phrase extraction (KPE) algorithm. Finally, the applicability and effectiveness of the method are verified by an example of identifying the information needs of academic users in the field of materials science. Findings Experimental results show that the BERT-based information needs classification model achieved the highest weighted average F1 score of 91.61%. The improved MDERank KPE algorithm achieves the highest F1 score of 61%. The empirical analysis results reveal that the information needs of the categories “methods,” “experimental phenomena” and “experimental materials” are relatively high in the materials science field. Originality/value This study provides a solution for automated identification of academic user information needs. It helps online academic resource platforms to better understand their users’ information needs, which in turn facilitates the platform’s academic resource organization and services.

Feature Selection and Performance Evaluation of Buzzer Classification Model

In the rapidly evolving digital age, social media platforms have transformed into battleground for shaping public opinion. Among these platforms, X has been particularly susceptible to the phenomenon of 'buzzers', paid or coordinated actors who manipulate online discussions and influence public sentiment. This manipulation poses significant challenges for users, researchers, and policymakers alike, necessitating robust detection measures and strategic feature selection for accurate classification models. This research explores the utilization of various feature selection techniques to identify the most influential features among the 24 features employed in the classification modeling using Support Vector Machine. This study found that selecting 11 key features yields a remarkably effective classification model, achieving an impressive F1-score of 87.54 in distinguishing between buzzer and non-buzzer accounts. These results suggest that focusing on the relevant features can improve the accuracy and efficiency of buzzer detection models. By providing a more robust and adaptable solution to buzzer detection, our research has the potential to advance social media research and policy. This enabling researchers and policymakers to devise strategies aimed at mitigating misinformation dissemination and cultivating an environment of trust and integrity within social media platforms, thus fostering healthier online interactions and discourse.

A Novel Hybrid Attention-Based Dilated Network for Depression Classification Model from Multimodal Data Using Improved Heuristic Approach

Automatic depression classification from multimodal input data is a challenging task. Modern methods use paralinguistic information such as audio and video signals. Using linguistic information such as speech signals and text data for depression classification is a complicated task in deep learning models. Best audio and video features are built to produce a dependable depression classification system. Textual signals related to depression classification are analyzed using text-based content data. Moreover, to increase the achievements of the depression classification system, audio, visual, and text descriptors are used. So, a deep learning-based depression classification model is developed to detect the person with depression from multimodal data. The EEG signal, Speech signal, video, and text are gathered from standard databases. Four stages of feature extraction take place. In the first stage, the features from the decomposed EEG signals are attained by the empirical mode decomposition (EMD) method, and features are extracted by means of linear and nonlinear feature extraction. In the second stage, the spectral features of the speech signals from the Mel-frequency cepstral coefficients (MFCC) are extracted. In the third stage, the facial texture features from the input video are extracted. In the fourth stage of feature extraction, the input text data are pre-processed, and from the pre-processed data, the textual features are extracted by using the Transformer Net. All four sets of features are optimally selected and combined with the optimal weights to get the weighted fused features using the enhanced mountaineering team-based optimization algorithm (EMTOA). The optimal weighted fused features are finally given to the hybrid attention-based dilated network (HADN). The HDAN is developed by combining temporal convolutional network (TCN) with bidirectional long short-term memory (Bi-LSTM). The parameters in the HDAN are optimized with the assistance of the developed EMTOA algorithm. At last, the classified output of depression is obtained from the HDAN. The efficiency of the developed deep learning HDAN is validated by comparing it with various traditional classification models.

MutaPT: A Multi-Task Pre-Trained Transformer for Classifying State of Disorders of Consciousness Using EEG Signal

Deep learning (DL) has been demonstrated to be a valuable tool for classifying state of disorders of consciousness (DOC) using EEG signals. However, the performance of the DL-based DOC state classification is often challenged by the limited size of EEG datasets. To overcome this issue, we introduce multiple open-source EEG datasets to increase data volume and train a novel multi-task pre-training Transformer model named MutaPT. Furthermore, we propose a cross-distribution self-supervised (CDS) pre-training strategy to enhance the model’s generalization ability, addressing data distribution shifts across multiple datasets. An EEG dataset of DOC patients is used to validate the effectiveness of our methods for the task of classifying DOC states. Experimental results show the superiority of our MutaPT over several DL models for EEG classification.

Leveraging Hybrid Adaptive Sine Cosine Algorithm with Deep Learning for Arabic Poem Meter Detection

Poetry is a significant aspect of any language. Many cultures and the history of nations are recognized in poems. Compared to prose, each poem has a rhythmic structure that is quite different. The language has its set of lyrical structures for poems, known as meters. Detecting the meters of Arabic poems is a complicated and lengthy procedure. The text must be encrypted using the Arudi method to classify the poem's meter, which requires complex rule-based transformation before another set of rules classifies the meters. Applying deep learning (DL) to meter classification in Arabic poems includes constructing a neural network to discern rhythmic patterns inherent in various meters. The model can extract essential features, like word lengths or syllable patterns, by tokenizing and preprocessing text datasets. Architectures such as Long Short-Term Memory Networks (LSTM) or Recurrent Neural Networks (RNNs) are fitting solutions to capture temporal relations in poetic verses. This research introduces a Hybrid Meta-heuristics with Deep Learning for the Arabic Poem Meter Detection and Classification (HMDL-APMDC) model. The main intention of the HMDL-APMDC system is to recognize various kinds of meters in Arabic poems. The HMDL-APMDC technique primarily preprocesses the input dataset to make it compatible with the classification process. Besides, the HMDL-APMDC technique applies Convolution and Attention with a Bi-directional Gated Recurrent Unit (CAT-BiGRU) for the automated recognition of meter classes. Furthermore, the adaptive sine-s-cosine particle swarm optimization (ASCA-PSO) algorithm is applied to optimize the hyperparameter tuning of the CAT-BiGRU model, enhancing the meter detection results. A detailed simulation analysis is made to highlight the improved performance of the HMDL-APMDC technique. The empirical outcomes stated that the HMDL-APMDC technique had a superior outcome of 98.53% over recent models under the MetRec dataset.

Meta-Analysis and Machine Learning Models for Anaerobic Biodegradation Rates of Organic Contaminants in Sediments and Sludge.

Anaerobic biodegradation rates (half-lives) of organic chemicals are pivotal for environmental risk assessment and remediation. Traditional experimental evaluation, constrained by prolonged, oxygen-free conditions, struggles to keep pace with emerging contaminants. Data-driven machine learning (ML) models serve as promising complements. However, reported quantitative structure-biodegradation relationships or ML models on anaerobic biodegradation are mostly based on small data sets (<100 records) and neglect experimental conditions, usually achieving compromised predictions. This work aimed to develop ML models for predicting the biodegradation half-lives of organic pollutants in anaerobic environments (i.e., sediment/soil and sludge). Focusing on important features of both chemicals and experimental conditions, we first curated two data sets, one for sediment/soil (SED) and the other for sludge (SLD), covering 978 records for 206 chemicals from the literature, and then conducted a meta-analysis. Next, we built a binary classification (half-life of 30 days as the cutoff) model with an accuracy of 81% and a regression model with R2 of 0.56 for SED based on LightGBM (80% and 0.31 for SLD based on Extra tree, respectively). The model interpretations underscored the significance of experimental conditions (e.g., temperature and inoculum dosage), as evidenced by their high feature importance, and the models were found to correctly capture the effects of chemical substructures, for example, branched structures and aromatic rings prolonged half-lives while methyl group and ortho-substitution on rings shortened half-lives. The applicability domains of the models were also defined, resulting in reasonable prediction for the half-lives of 41% (SED) or 67% (SLD) of over 4000 persistent, bioaccumulative, and toxic chemicals. Overall, this study pioneers ML models for predicting the anaerobic degradation half-lives, offering valuable support for future evaluation and implementation of chemical anaerobic biodegradation.

Optimization of a Novel Engineered Ecosystem Integrating Carbon, Nitrogen, Phosphorus, and Sulfur Biotransformation for Saline Wastewater Treatment Using an Interpretable Machine Learning Approach.

The denitrifying sulfur (S) conversion-associated enhanced biological phosphorus removal (DS-EBPR) process for treating saline wastewater is characterized by its unique microbial ecology that integrates carbon (C), nitrogen (N), phosphorus (P), and S biotransformation. However, operational instability arises due to the numerous parameters and intricates bacterial interactions. This study introduces a two-stage interpretable machine learning approach to predict S conversion-driven P removal efficiency and optimize DS-EBPR process. Stage one utilized the XGBoost regression model, achieving an R2 value of 0.948 for predicting sulfate reduction (SR) intensity from anaerobic parameters with feature engineering. Stage two involved the CatBoost classification and regression model integrating anoxic parameters with the predicted SR values for predicting P removal, reaching an accuracy of 94% and an R2 value of 0.93, respectively. This study identified key environmental factors, including SR intensity (20-45 mg S/L), influent P concentration (<9.0 mg P/L), mixed liquor volatile suspended solids (MLVSS)/mixed liquor suspended solids (MLSS) ratio (0.55-0.72), influent C/S ratio (0.5-1.0), anoxic reaction time (5-6 h), and MLSS concentration (>6.50 g/L). A user-friendly graphic interface was developed to facilitate easier optimization and control. This approach streamlines the determination of optimal conditions for enhancing P removal in the DS-EBPR process.