Learning Convolutional Neural Network Research Articles

Prediction of Potential Evapotranspiration via Machine Learning and Deep Learning for Sustainable Water Management in the Murat River Basin

Potential evapotranspiration (PET) is a significant factor contributing to water loss in hydrological systems, making it a critical area of research. However, accurately calculating and measuring PET remains challenging due to the limited availability of comprehensive data. This study presents a detailed sustainable model for predicting PET using the Thornthwaite equation, which requires only mean monthly temperature (Tmean) and latitude, with calculations performed using R-Studio. A geographic information system (GIS) was employed to interpolate meteorological data, ensuring coverage of all sub-basins within the Murat River basin, the study area. Additionally, Python libraries were utilized to implement artificial intelligence-driven models, incorporating both machine learning and deep learning techniques. The study harnesses the power of artificial intelligence (AI), applying deep learning through a convolutional neural network (CNN) and machine learning techniques, including support vector machine (SVM) and random forest (RF). The results demonstrate promising performance across the models. For CNN, the coefficient of determination (R2) varied from 96.2 to 98.7%, the mean squared error (MSE) ranged from 0.287 to 0.408, and the root mean squared error (RMSE) was between 0.541 and 0.649. For SVM, the R2 varied from 94.5 to 95.6%, MSE ranged between 0.981 and 1.013, and RMSE ranged from 0.990 to 1.014. RF showed the best performance, achieving an R2 of 100%, MSE values of 0.326 and 0.640, and corresponding RMSE values of 0.571 and 0.800. The climate and topography data used for all algorithms were consistent, and the results indicate that the RF model outperforms the others. Consequently, The RF model’s superior accuracy highlights its potential as a reliable tool for sustainable PET prediction, supporting informed decision-making in water resource planning. By leveraging GIS, AI, and machine learning, this study enhances PET modeling methodologies, addressing critical water management challenges and promoting sustainable hydrological practices in the face of climate change and resource limitations.

Image analysis-based identification of high risk ER-positive, HER2-negative breast cancers

BackgroundBreast cancer subtypes Luminal A and Luminal B are classified by the expression of PAM50 genes and may benefit from different treatment strategies. Machine learning models based on H&E images may contain features associated with subtype, allowing early identification of tumors with higher risk of recurrence.MethodsH&E images (n = 630 ER+/HER2-breast cancers) were pixel-level segmented into epithelium and stroma. Convolutional neural network and multiple instance learning were used to extract image features from original and segmented images. Patient-level classification models were trained to discriminate Luminal A versus B image features in tenfold cross-validation, with or without grade adjustment. The best-performing visual classifier was incorporated into envisioned diagnostic protocols as an alternative to genomic testing (PAM50). The protocols were then compared in time-to-recurrence models.ResultsAmong ER+/HER2-tumors, the image-based protocol differentiated recurrence times with a hazard ratio (HR) of 2.81 (95% CI: 1.73–4.56), which was similar to the HR for PAM50 (2.66, 95% CI: 1.65–4.28). Grade adjustment did not improve subtype prediction accuracy, but did help balance sensitivity and specificity. Among high grade participants, sensitivity and specificity (0.734 and 0.474, respectively) became more similar (0.732 and 0.624, respectively) in grade-adjusted models. The original and epithelium-specific images had similar performance and highest accuracy, followed by stroma or binarized images showing only the epithelial-stromal interface.ConclusionsGiven low rates of genomic testing uptake nationally, image-based methods may help identify ER+/HER2-patients who could benefit from testing.

Cable fault diagnosis with generalization capability using incremental learning and deep convolutional neural network

Power cable faults may exhibit dynamic and differential changes attributable to variations in operating times and environmental conditions. However, the existing artificial intelligence (AI)-based fault diagnosis methods are lack of capability to handle this complexity of fault development. They can only diagnose given faulty types and will malfunction when dealing with newly-emerged faults, which accounts for the lack of generalization capability for AI-based methods. Given that, this paper proposes a cable fault diagnosis method with generalization capability utilizing incremental learning and deep convolutional neural network (DCNN). The idea of knowledge distillation is introduced to reserve the diagnosis information of the DCNN for original faults, and meanwhile, the DCNN is also modified by the classification loss of newly-emerged faulty samples, which equips the DCNN with generalization capability. The category features are obtained utilizing the refined DCNN, and sent to the constructed classifier for accurate incremental fault diagnosis. Experiments are conducted with field and simulated data. The obtained results validate the feasibility and effectiveness of the proposed method.

Using three-dimensional fluorescence spectroscopy and machine learning for rapid detection of adulteration in camellia oil

Camellia oil had been widely utilized in the realms of cooking, healthcare, and beauty. Nevertheless, merchants frequently adulterated pure camellia oil with low-priced oils to cut costs. This study was aimed at identifying the authenticity of camellia oil. Through the employment of three-dimensional fluorescence spectroscopy combined with the parallel factor analysis (PARAFAC) method, the characteristics of different vegetable oils were analyzed to establish a foundation for classification modeling. In the identification of pure vegetable oil types, methods such as partial least squares discriminant analysis (PLS-DA), k-nearest neighbors (KNN), support vector machine (SVM), and random forest (RF) were adopted. The classification accuracy reached 100 %, demonstrating the effectiveness of feature extraction by PARAFAC. For the identification of camellia oil and its adulterants, traditional machine learning methods and convolutional neural network (CNN) models were introduced. The results indicated that traditional methods had limitations in the classification of single and binary adulterated oils. However, the optimized CaoCNN model achieved an accuracy of 97.78 % in identifying adulterated oil types, showcasing the potential of deep learning in adulterated oil detection. Further, feature visualization analysis verified the ability of CaoCNN to effectively capture and distinguish the characteristics of adulterated oils, providing an effective approach for the identification of camellia oil and its adulterated oils.

A hybrid classification approach for automatically recognizing COVID-19 using deep transfer learning using chest radiographs

Coronavirus 2019 causes COVID-19, a worldwide epidemic. It endangers millions globally. Early illness detection improves recovery and control. X-ray image processing is used to categorise and identify COVID-19 in the present study. Preprocessing, feature extraction using local binary pattern (LBP) and edge orient histogram (EOH), and classification utilising K-nearest neighbour (KNN), Navie Bayes, support vector machine (SVM), and transfer learning convolution neural networks (CNNs) are some of the stages that are implemented in the process. Other phases in the process include preprocessing, feature extraction, and preprocessing. LBP+KNN, EOH+KNN, LBP+SVM, EOH +SVM, CNN+LBP, and CNN+EOH are the outputs derived from the combinations of feature extraction operators and classifiers. Other possible outcomes are CNN+EOH and CNN+LBP. A total of 4,000 pictures were used as the basis for conducting an analysis of the performance of six different models. In order to train the models, 10-fold cross-validation was used, and their accuracy was measured accordingly. The evaluation results indicate a high level of accuracy in diagnosis, ranging from 90.2% to 97.56%. The CNN+LBP and CNN+EOH models have demonstrated superior performance compared to other models, achieving average accuracies ranging from 96.66% and 98.54%.

Plant disease classification using novel integration of deep learning CNN and graph convolutional networks

Plant diseases present substantial challenges to global agriculture, significantly affecting crop yields and jeopardizing food security. Accurate and timely detection of these diseases is paramount for mitigating their adverse effects. This paper proposes a novel approach for plant disease classification by integrating convolutional neural networks (CNNs) and graph convolutional networks (GCNs). The model aims to enhance classification accuracy by leveraging both visual features extracted by CNNs and relational information captured by GCNs. Using a Kaggle dataset containing images of diseased and healthy plant leaves from 31 classes, including apple, corn, grape, peach, pepper bell, potato, strawberry, and tomato. Standalone CNN models were trained on image data from each plant type, while standalone GCN models utilized graph-structured data representing plant relationships within each subset. The proposed integrated CNN-GCN model capitalizes on the complementary strengths of CNNs and GCNs to achieve improved classification performance. Through rigorous experimentation and comparative analysis, the effectiveness of the integrated CNN-GCN approach was evaluated alongside standalone CNN and GCN models across all plant types. Results demonstrated the superiority of the integrated model, highlighting its potential for enhancing plant disease classification accuracy.

Artificial intelligence facial recognition of obstructive sleep apnea: a Bayesian meta-analysis.

Conventional obstructive sleep apnea (OSA) diagnosis via polysomnography can be costly and inaccessible. Recent advances in artificial intelligence (AI) have enabled the use of craniofacial photographs to diagnose OSA. This meta-analysis aims to clarify the diagnostic accuracy of this innovative approach. Two blinded reviewers searched PubMed, Embase, Scopus, Web of Science, and IEEE Xplore databases, then selected and graded the risk of bias of observational studies of adults (≥ 18 years) comparing the diagnostic performance of AI algorithms using craniofacial photographs, versus conventional OSA diagnostic criteria (i.e. apnea-hypopnea index [AHI]). Studies were excluded if they detected apneic events without diagnosing OSA. AI models evaluated with a random split test set or k-fold cross-validation were included in a Bayesian bivariate meta-analysis. From 5,147 records, 6 studies were included, containing 10 AI models trained/tested on 1,417/983 participants. The risk of bias was low. AI trained on craniofacial photographs achieved a pooled 84.9% sensitivity (95% credible interval [95% CrI]: 77.1-90.7%) and 71.2% specificity (95% CrI: 60.7-81.4%). Bayesian meta-regression identified deep learning (convolutional neural networks) as the most accurate AI algorithm (91.1% sensitivity, 79.2% specificity) comparable to home sleep apnea tests. AHI cutoffs, OSA prevalence, feature engineering, input data, camera type and informativeness of Bayesian prior did not alter diagnostic accuracy. There was no substantial publication bias. AI trained on craniofacial photographs have high diagnostic accuracy and should be considered as a low-cost OSA screening tool. Future work focused on deep learning using smartphone images could improve the feasibility of this approach in primary care.

Computer-aided diagnosis of early-stage Retinopathy of Prematurity in neonatal fundus images using artificial intelligence

Retinopathy of Prematurity (ROP) is a retinal disorder affecting preterm babies, which can lead to permanent blindness without treatment. Early-stage ROP diagnosis is vital in providing optimal therapy for the neonates. The proposed study predicts early-stage ROP from neonatal fundus images using Machine Learning (ML) classifiers and Convolutional Neural Networks (CNN) based pre-trained networks. The characteristic demarcation lines and ridges in early stage ROP are segmented utilising a novel Swin U-Net. 2000 Scale Invariant Feature Transform (SIFT) descriptors were extracted from the segmented ridges and are dimensionally reduced to 50 features using Principal Component Analysis (PCA). Seven ROP-specific features, including six Gray Level Co-occurrence Matrix (GLCM) and ridge length features, are extracted from the segmented image and are fused with the PCA reduced 50 SIFT features. Finally, three ML classifiers, such as Support Vector Machine (SVM), Random Forest (RF), andk- Nearest Neighbor (k-NN), are used to classify the 50 features to predict the early-stage ROP from Normal images. On the other hand, the raw retinal images are classified directly into normal and early-stage ROP using six pre-trained classifiers, namely ResNet50, ShuffleNet V2, EfficientNet, MobileNet, VGG16, and DarkNet19. It is seen that the ResNet50 network outperformed all other networks in predicting early-stage ROP with 89.5% accuracy, 87.5% sensitivity, 91.5% specificity, 91.1% precision, 88% NPV and an Area Under the Curve (AUC) of 0.92. Swin U-Net Convolutional Neural Networks (CNN) segmented the ridges and demarcation lines with an accuracy of 89.7% with 80.5% precision, 92.6% recall, 75.76% IoU, and 0.86 as the Dice coefficient. The SVM classifier using the 57 features from the segmented images achieved a classification accuracy of 88.75%, sensitivity of 90%, specificity of 87.5%, and an AUC of 0.91. The system can be utilised as a point-of-care diagnostic tool for ROP diagnosis of neonates in remote areas.

Performance Analysis of Various Deep Transfer Learning Models for Bacterial Blight Disease Detection and Classification in Pomegranate Fruits

Background: Detecting and classifying diseases in pomegranate fruit using computer vision remains a challenging task due to the presence of numerous diseases. Recent research findings suggest that models based on Convolutional Neural Networks (CNNs) have shown significant enhancements in accuracy when it comes to classifying images of fruits and leaves. The fungal infection responsible for the disease swiftly propagates through the soil, infiltrating the roots of pomegranate plants. Presently, the sole method to halt its spread entails farmers conducting thorough inspections and promptly eliminating infected plants, an uphill task. Methods: The present study introduces a classification PomeNetV1 model for identifying pomegranate fruit diseases employing CNN and transfer learning techniques. For executing this system, the dataset is created by taking the images directly from the farms in Ballari, Bengaluru, Bagalakote, etc. The proposed pomegranate fruit disease dataset contains 5099 images of five categories: Alternaria, Anthracnose, Bacterial Blight, Cercospora and Healthy. The system under consideration utilizes TensorFlow as its framework for developing deep learning models. Result: This paper evaluates ten different pre-trained models using a transfer learning approach and a proposed novel PomeNetV1 model for classification. All pre-trained models achieved over 99% classification accuracy. However, the PomeNetV1 model stands out with an impressive 99.80% accuracy, making it the most efficient for detecting healthy and bacterial blight diseases compared to others. Surpassing existing models like Vasumathi et al., 2021 deep CNN model with at least 1.63% accuracy, the PomeNetV1 model offers a feasible solution for detecting bacterial blight in pomegranate crops.

Detection of hate: speech tweets based convolutional neural network and machine learning algorithms

There is no doubt that social media sites have provided many benefits to humanity, such as sharing information continuously and communicating with others easily. It also seems that social media sites have many advantages, but in addition to these advantages, there are disadvantages that we always strive to find a solution. One of these disadvantages is sharing hate speech. In our study, we’re discussing a way to solve this phenomenon by using Term Frequency-Inverse Document Frequency (TF-IDF) based approach to feature engineering on eleven classifiers for machine and deep learning that can automatically identify hate speech. Three different databases were used, the first of which “Hate speech offensive tweets by Davidson et al.”, the second called "Twitter hate speech" and finally we merged the second data with (Cyberbullying dataset (toxicity_parsed_dataset)". The classifiers involved are Logistic Regression (LR), Naive Bayes (NB), Multi-layer Perceptron (MLP), and Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbor (KNN), K-Means, Decision Tree (DT), Gradient Boosting classifier (GBC), and the Extra Trees (ET) in addition to the convolutional neural network (CNN). Maximum accuracy was attained, which exceeded 99%.

Machine learning-Assisted spiral fiber Bragg Grating-Based flexible dual-parameter sensing

The ability of flexible sensors to bend or fold freely offers great advantages in sensing ability and adaptability to harsh environments. Moreover, compared with typical electrical effect based flexible sensors, optical based fiber Bragg grating (FBG) flexible sensors offer much greater ease of networking and resistance to electromagnetic interference, making them suitable for distributed multi-point strain measurements in complex environments. In this paper, two FBGs with no overlapping reflective spectra are shallowly embedded in the surface layer of a flexible thin-cylinder substrate to form a dual-parameter flexible strain sensor. However, it is crucial for the changes in direction and curvature parameters of FBG strain sensors under deformation to be accurately understood to characterize the current deformation state of the flexible sensor. Moreover, conventional peak tracking demodulation methods often fail to account for distortion in the reflected spectrum of a spiral FBG under stress. Hence, a multi-output convolutional neural network learning model is constructed to simultaneously identify the bending direction and curvature radius of the flexible sensor using machine learning methods. Experimental results show that the flexible dual-parameter FBG sensor has precisely recognize angles to within 2° across a 360° range, with a curvature recognition accuracy of 99.1%, offering precision sensing performance suitable for highly demanding application scenarios such as bionic robots and flexible medical devices.

Transformer learning-based neural network algorithms for identification and detection of electronic bullying in social media

Abstract The global phenomenon known as cyberbullying is a form of modern harassment that cannot be entirely stopped but can be avoided. Most current solutions to the cyberbullying problem have relied on tools and methods to identify online bullying. However, end users do not have free access to these tools. The goal of this study is to create a model to combat cyberbullying on social media sites based on users’ appearance. In this article, we present a cyberbullying detection system constructed using the Word2Vec word-embedding method and a deep learning convolutional neural network combined with bidirectional long short-term memory (CNN-BiLSTM), as well as the XLM-Roberta transformer, to develop a model for cyberbullying detection. We carried out two experiments based on binary (hate speech or non-hate speech bullying comments) and multiclass (religion, age, gender, ethnicity, and non-bullying tweets) datasets collected from Kaggle online discussions and Twitter. To evaluate the model’s performance, we used standard measurement metrics, such as precision, recall, F1-score, and accuracy. Through a comparison of the results, it is noted that the XLM-Roberta model outperformed the CNN-BiLSTM model, resulting in 84% accuracy using the Kaggle online discussion dataset and 94% accuracy using the Twitter dataset.

Enhanced Typhoon Center Localization Using Geostationary Satellite Imagery

AbstractAn accurate center localization in near real‐time is critical for tropical cyclone (TC) monitoring and forecasting. This study presents a robust algorithm for localizing typhoon centers using the Chinese geostationary (GEO) meteorological satellite. The results using the Advanced Geostationary Radiation Imager (AGRI) onboard Fengyun‐4A (FY‐4A) satellite data, achieving a mean absolute error (MAE) of 29.4 km across various typhoon intensities in the Western North Pacific, superior to other baseline methods. By harnessing the multi‐spectral imagery from the FY‐4A and incorporating an attention mechanism, it significantly boosts the deep learning convolutional neural network's ability to identify typhoon cloud features and their centers, even during their initial and weakest stages, which is laudable because these are the most difficult for center fixing even for human analysts. Remarkably, it requires just a single moment satellite imagery to locate the center of typhoon, enabling automated updates of the typhoon centers in near real‐time applications.

A Comparison of Convolutional Neural Network (CNN) and Transfer Learning MobileNetV2 Performance on Spices Images Classification

This research was conducted to analyze the performance of the CNN algorithm without transfer learning in classifying spice images and compare it with the CNN algorithm using transfer learning on the MobileNetV2 architecture. This comparison aims to evaluate both methods' accuracy, efficiency, and overall performance and analyze the impact of transfer learning on classification results in the context of spices. The dataset consists of 1500 spice images divided into 10 classes, with each class of 150 images. In the first experiment, CNN without transfer learning resulted in 93% accuracy performance. For the second experiment using MobileNetV2, there was an increase in accuracy, reaching a value of 99% for all spice classes. The results of this study confirm that MobileNetV2 architecture significantly improves the accuracy and performance of spice classification compared to CNN without transfer learning, which can be recommended for spice image classification.

Artificial intelligence-assisted magnetic resonance imaging technology in the differential diagnosis and prognosis prediction of endometrial cancer

It aimed to analyze the value of deep learning algorithm combined with magnetic resonance imaging (MRI) in the risk diagnosis and prognosis of endometrial cancer (EC). Based on the deep learning convolutional neural network (CNN) architecture residual network with 101 layers (ResNet-101), spatial attention and channel attention modules were introduced to optimize the model. A retrospective collection of MRI image data from 210 EC patients was used for model segmentation and reconstruction, with 140 cases as the test set and 70 cases as the validation set. The performance was compared with traditional ResNet-101 model, ResNet-101 model based on spatial attention mechanism (SA-ResNet-101), and ResNet-101 model based on channel attention mechanism (CA-ResNet-101), using accuracy (AC), precision (PR), recall (RE), and F1 score as evaluation metrics. Among the 70 cases in the validation set, there were 45 cases of low-risk EC and 25 cases of high-risk EC. Using ROC curve analysis, it was found that the area under the curve (AUC) for the diagnosis of high-risk EC of the proposed model in this article (0.918) was visibly larger as against traditional ResNet-101 (0.613), SA-ResNet-101 (0.760), and CA-ResNet-101 models (0.758). The AC, PR, RE, and F1 values of the proposed model for the diagnosis of EC risk were visibly higher (P < 0.05). In the validation set, postoperative recurrence occurred in 13 cases and did not occur in 57 cases. Using ROC curve analysis, it was found that the AUC for postoperative recurrence prediction of the patients by the proposed model (0.926) was visibly larger as against traditional ResNet-101 (0.620), SA-ResNet-101 (0.729), and CA-ResNet-101 models (0.767). The AC, PR, RE, and F1 values of the proposed model for postoperative recurrence prediction were visibly higher (P < 0.05). The proposed model in this article, assisted by MRI, presented superior performance in diagnosing high-risk EC patients, with higher sensitivity (Sen) and specificity (Spe), and also demonstrated excellent predictive AC in postoperative recurrence prediction.

Classifying driver mutations of papillary thyroid carcinoma on whole slide image: an automated workflow applying deep convolutional neural network.

Informative biomarkers play a vital role in guiding clinical decisions regarding management of cancers. We have previously demonstrated the potential of a deep convolutional neural network (CNN) for predicting cancer driver gene mutations from expert-curated histopathologic images in papillary thyroid carcinomas (PTCs). Recognizing the importance of whole slide image (WSI) analysis for clinical application, we aimed to develop an automated image preprocessing workflow that uses WSI inputs to categorize PTCs based on driver mutations. Histopathology slides from The Cancer Genome Atlas (TCGA) repository were utilized for diagnostic purposes. These slides underwent an automated tile extraction and preprocessing pipeline to ensure analysis-ready quality. Next, the extracted image tiles were utilized to train a deep learning CNN model, specifically Google's Inception v3, for the classification of PTCs. The model was trained to distinguish between different groups based on BRAFV600E or RAS mutations. The newly developed pipeline performed equally well as the expert-curated image classifier. The best model achieved Area Under the Curve (AUC) values of 0.86 (ranging from 0.847 to 0.872) for validation and 0.865 (ranging from 0.854 to 0.876) for the final testing subsets. Notably, it accurately predicted 90% of tumors in the validation set and 84.2% in the final testing set. Furthermore, the performance of our new classifier showed a strong correlation with the expert-curated classifier (Spearman rho = 0.726, p = 5.28 e-08), and correlated with the molecular expression-based classifier, BRS (BRAF-RAS scores) (Spearman rho = 0.418, p = 1.92e-13). Utilizing WSIs, we implemented an automated workflow with deep CNN model that accurately classifies driver mutations in PTCs.

Quantifying Shape and Texture Biases for Enhancing Transfer Learning in Convolutional Neural Networks

Neural networks have inductive biases owing to the assumptions associated with the selected learning algorithm, datasets, and network structure. Specifically, convolutional neural networks (CNNs) are known for their tendency to exhibit textural biases. This bias is closely related to image classification accuracy. Aligning the model’s bias with the dataset’s bias can significantly enhance performance in transfer learning, leading to more efficient learning. This study aims to quantitatively demonstrate that increasing shape bias within the network by varying kernel sizes and dilation rates improves accuracy on shape-dominant data and enables efficient learning with less data. Furthermore, we propose a novel method for quantitatively evaluating the balance between texture bias and shape bias. This method enables efficient learning by aligning the biases of the transfer learning dataset with those of the model. Systematically adjusting these biases allows CNNs to better fit data with specific biases. Compared to the original model, an accuracy improvement of up to 9.9% was observed. Our findings underscore the critical role of bias adjustment in CNN design, contributing to developing more efficient and effective image classification models.

Deep Reinforcement Learning of Graph Convolutional Neural Network for Resilient Production Control of Mass Individualized Prototyping Toward Industry 5.0

Deep Reinforcement Learning of Graph Convolutional Neural Network for Resilient Production Control of Mass Individualized Prototyping Toward Industry 5.0

Comparison of transformer, LSTM and coupled algorithms for soil moisture prediction in shallow-groundwater-level areas with interpretability analysis

Accurate quantification of soil moisture is essential for understanding water and energy exchanges between the atmosphere and the Earth’s surface, as well as for agricultural applications. Predicting soil moisture content is vital for efficient water management, irrigation scheduling, and drought monitoring. Traditional forecasting methods, such as numerical regression models, often struggle due to various influencing factors and poor observation data quality. In contrast, deep learning algorithms, particularly recurrent and convolutional neural networks, show promise in predicting nonlinear data like soil moisture. This study focuses on shallow groundwater regions, using groundwater levels and meteorological data as features while coupling the Transformer model with other neural network structures. We investigate the potential of attention-based neural networks for soil moisture time series prediction. Our findings demonstrate that the Transformer model achieves an average R2 of 0.523 across different time lags, outperforming the LSTM model with an R2 of 0.485. The introduction of LSTM enhances the Transformer’s stability in handling temporal changes. Additionally, we verified the importance of groundwater for soil moisture prediction. This study introduces new methods for soil moisture prediction and offers new insights and recommendations for the development of artificial intelligence technology for soil moisture prediction.

A Deep Learning Approach to Identify Potential Sites for Pocket Park Installation in Nanjing, China

ABSTRACT Pocket parks are considered valuable tools for supporting green space provision in high-density environments. As the factors affecting pocket parks vary in fine-grained granularity, analyzing them efficiently on a large scale using traditional methods is challenging. Therefore, considering Nanjing as the study site, a potential identification framework was established, which includes categories of space, vitality, facility, and pleasurability for pocket park installation (PPI) based on visual information from street view images (SVIs). Then, the TrueSkill algorithm was employed to construct four parallel-task training sets according to the four categories and a convolutional neural network (CNN) and active learning (AL) approaches were applied to establish four parallel-task models of deep learning (DL). A two-phase grading score method was designed to integrate the identification results of the four DL models into a comprehensive potential score to determine priority sites for PPI. Lastly, the attributes of the potential sites and corresponding planning strategies in different scenarios were analyzed. The results showed that this modelling approach can integrally assess multiple factors on PPI within an enormous scope, overcoming traditional methods’ spatial and dimensional constraint in high-density environments. Our method and results can inform future pocket park planning and design in Nanjing and similar cities worldwide.