Size Of Input Image Research Articles

DAC: Differentiable Auto-Cropping in Deep Learning

Auto-cropping, the process of automatically adjusting the boundaries of an image to focus on the region of interest, is crucial to improving the diagnostic quality of dental panoramic radiographs. Its importance lies in its ability to standardize the size of different input images with minimal loss of information, thus ensuring consistency and improving the performance of subsequent image-processing tasks. Despite the widespread use of CNNs in many studies, research on auto-cropping for different-sized images remains limited. This study aims to explore the potential of differentiable auto-cropping in dental panoramic radiographs. A unique dataset of 20,973 dental panoramic radiographs, mostly with a resolution of 2836×1536 or close, divided into five classes by 3 dentists, was used, which is the same dataset from the previous study (Top et al. 2023). ResNet-101 model, which was the most successful network for the dataset (Top et al. 2023), was used for the evaluation. To reduce variance, the model was evaluated using 10-fold cross-validation for both non-auto-cropped and auto-cropped trainings. Data augmentation was also used to produce more accurate and robust results. For auto-cropped training, it was adjusted to be much less effective than the non-auto-cropped one. Accuracy was improved by 1.8%, from 92.7% to 94.5%, thanks to the proposed auto-crop optimization developed to reduce dataset-related issues. Its macro-average AUC was also raised from 0.989 to 0.993. The proposed auto-crop optimization can be implemented as a trainable network layer in an end-to-end CNN and can be used for other problems as well. Increasing the accuracy from 92.7% to 94.5% is a very challenging task due to diminishing returns, as there is little room for improvement. The results show the potential of the proposed differentiable auto-crop algorithm and encourages its use in different fields.

Breast tumor segmentation in ultrasound using distance-adapted fuzzy connectedness, convolutional neural network, and active contour

This study addresses computer-aided breast cancer diagnosis through a hybrid framework for breast tumor segmentation in ultrasound images. The core of the three-stage method is based on the autoencoder convolutional neural network. In the first stage, we prepare a hybrid pseudo-color image through multiple instances of fuzzy connectedness analysis with a novel distance-adapted fuzzy affinity. We produce different weight combinations to determine connectivity maps driven by particular image specifics. After the hybrid image is processed by the deep network, we adjust the segmentation outcome with the Chan-Vese active contour model. We find the idea of incorporating fuzzy connectedness into the input data preparation for deep-learning image analysis our main contribution to the study. The method is trained and validated using a combined dataset of 993 breast ultrasound images from three public collections frequently used in recent studies on breast tumor segmentation. The experiments address essential settings and hyperparameters of the method, e.g., the network architecture, input image size, and active contour setup. The tumor segmentation reaches a median Dice index of 0.86 (mean at 0.79) over the combined database. We refer our results to the most recent state-of-the-art from 2022–2023 using the same datasets, finding our model comparable in segmentation performance.

Fusion of visible and fluorescence imaging through deep neural network for color value prediction of pelletized red peppers.

A non-destructive method for determining the color value of pelletized red peppers is crucial for pepper processing factories. This study aimed to investigate the potentiality of visible and fluorescence images for the determination of color value of pelletized red pepper. The imaging problem, caused by the cylindrical shape and irregular cross-sectional features of the pelletized red peppers, was reduced through the extraction of an approximate plane region. To integrate the information in the visible and fluorescence images, a baseline convolutional neural network (CNN) architecture was designed and low level, middle level, and high level fusion models (denoted as LL-CNN, ML-CNN, and HL-CNN, respectively) were developed upon the baseline CNN. The effects of input image size and color space were examined. According to the training result, CNN fusion models were developed using visible image in L*a*b* color space and fluorescence image in RGB color space using 56×56 input image size. Among the three types of CNN fusion models, the HL-CNN obtained the best performance, resulting in Rv 2 of 0.828 and RMSEV of 0.351. This study suggests that the fusion of visible and fluorescence image through CNN is a practical approach to save testing time and replace traditional destructive method. The low cost and compact structure of the imaging systems can maintain the commercial appeal of pepper industry.

Multi-source information fusion for enhanced in-process quality monitoring of laser powder bed fusion additive manufacturing

Defects such as lack of fusion, porosity, and keyhole generated during the laser powder bed fusion (L-PBF) additive manufacturing process pose a challenge, with the absence of effective prediction methods for the process-induced defects and as-printed quality. On-line monitoring becomes imperative to evaluate and enhance the L-PBF in-process quality. Here, a multi-source information fusion strategy using a residual network (ResNet) is introduced for the in-process monitoring during the L-PBF. This approach integrates the melt-pool infrared (IR) images captured layer-by-layer, melt-track top-view photographs, melt-track numerical simulation diagrams, L-PBF process parameters, and characteristic parameters of melt-pool cross-sectional morphology after solidification to enable quality monitoring of the L-PBF processing. To assess the defect severity, a quantitative defect evaluation method based on the defect-specific characteristics is proposed. This method facilitates the quantitative evaluation of defects by extracting pertinent defect indicators related to porosity and deformation. Additionally, two types of residual physical hybrid networks (ResPHN) and two types of residual physical fusion supervisory networks (ResPFSN) are introduced in this study. The performance of these four network models is meticulously compared and evaluated. The findings reveal that the most effective feature fusion monitoring model is the ResPFSN-type2, achieving an impressive accuracy of 99.4% and displaying consistent performance across varying input image sizes and training data volumes. It underscores its potential for real-time process control applications. Furthermore, the interpretability of the model is scrutinized, with results indicating that the ResPFSN-type2 model adeptly identifies the contour texture and local features of the laser-induced melt pools.

HCFormer: A Lightweight Pest Detection Model Combining CNN and ViT

Pests are widely distributed in nature, characterized by their small size, which, along with environmental factors such as lighting conditions, makes their identification challenging. A lightweight pest detection network, HCFormer, combining convolutional neural networks (CNNs) and a vision transformer (ViT) is proposed in this study. Data preprocessing is conducted using a bottleneck-structured convolutional network and a Stem module to reduce computational latency. CNNs with various kernel sizes capture local information at different scales, while the ViT network’s attention mechanism and global feature extraction enhance pest feature representation. A down-sampling method reduces the input image size, decreasing computational load and preventing overfitting while enhancing model robustness. Improved attention mechanisms effectively capture feature relationships, balancing detection accuracy and speed. The experimental results show that HCFormer achieves 98.17% accuracy, 91.98% recall, and a mean average precision (mAP) of 90.57%. Compared with SENet, CrossViT, and YOLOv8, HCFormer improves the average accuracy by 7.85%, 2.01%, and 3.55%, respectively, outperforming the overall mainstream detection models. Ablation experiments indicate that the model’s parameter count is 26.5 M, demonstrating advantages in lightweight design and detection accuracy. HCFormer’s efficiency and flexibility in deployment, combined with its high detection accuracy and precise classification, make it a valuable tool for identifying and classifying crop pests in complex environments, providing essential guidance for future pest monitoring and control.

Exploring Factors Affecting the Performance of Neural Network Algorithm for Detecting Clouds, Snow, and Lakes in Sentinel-2 Images

Detecting clouds, snow, and lakes in remote sensing images is vital due to their propensity to obscure underlying surface information and hinder data extraction. In this study, we utilize Sentinel-2 images to implement a two-stage random forest (RF) algorithm for image labeling and delve into the factors influencing neural network performance across six aspects: model architecture, encoder, learning rate adjustment strategy, loss function, input image size, and different band combinations. Our findings indicate the Feature Pyramid Network (FPN) achieved the highest MIoU of 87.14%. The multi-head self-attention mechanism was less effective compared to convolutional methods for feature extraction with small datasets. Incorporating residual connections into convolutional blocks notably enhanced performance. Additionally, employing false-color images (bands 12-3-2) yielded a 4.86% improvement in MIoU compared to true-color images (bands 4-3-2). Notably, variations in model architecture, encoder structure, and input band combination had a substantial impact on performance, with parameter variations resulting in MIoU differences exceeding 5%. These results provide a reference for high-precision segmentation of clouds, snow, and lakes and offer valuable insights for applying deep learning techniques to the high-precision extraction of information from remote sensing images, thereby advancing research in deep neural networks for semantic segmentation.

Automatic Foreign Matter Segmentation System for Superabsorbent Polymer Powder: Application of Diffusion Adversarial Representation Learning

In current industries, sampling inspections of the quality of powders, such as superabsorbent polymers (SAPs) still are conducted via visual inspection. The size of samples and foreign matter are around 500 μm, making them difficult for humans to identify. An automatic foreign matter detection system for powder has been developed in the present study. The powder samples can be automatically delivered, distributed, and recycled, and images of them are captured through the hardware of the system, while the identification software of this system was developed based on diffusion adversarial representation learning (DARL). The background image is a foreign-matter-free powder image with an input image size of 1024 × 1024 × 3. Since DARL includes adversarial segmentation, a diffusion process, and synthetic image generation, the DARL model was trained using a diffusion block with the employment of a U-Net attention mechanism and a spatial-adaptation de-normalization (SPADE) layer through the adoption of a loss function from a vanilla generative adversarial network (GAN). This model was then compared with supervised models such as a fully convolutional network (FCN), U-Net, and DeepLABV3+, as well as with an unsupervised Otsu threshold segmentation. It should be noted that only 10% of the training samples were utilized for the DARL to learn and the intersection over union (IoU) of the DARL can reach up to 80.15%, which is much higher than the 59.00%, 53.47%, 49.39%, and 30.08% for the Otsu threshold segmentation, FCN, U-Net, and DeepLABV3+ models. Therefore, the performance of the model developed in the present study would not be degraded due to an insufficient number of samples containing foreign matter. In practical applications, there is no need to collect, label, and design features for a large number of foreign matter samples before using the developed system.

Novel identification algorithm for plasma boundary gap based on visible endoscope diagnostic on EAST tokamak

The precise plasma boundary gap identification at the midplane is a prerequisite for achieving controlled plasma positioning and holds a significant importance for the stable operation of tokamak devices. This study proposes a plasma boundary gap at the midplane recognition algorithm based on visual endoscopy diagnostic. The model is an end-to-end one that uses a convolutional neural network that does not require manual data labeling. The model performance is improved by experimentally comparing different convolutional layers and input image sizes. The model is validated using a testing dataset comprising 400 plasma discharge moments. The model has average errors of 3.7 and 4 mm for gap-in and -out, respectively, when compared to those obtained by equilibrium fitting. The proposed approach offers a convenient and effective means of obtaining the boundary gap value and is particularly suited for future fusion experimental devices, such as BEST and ITER tokamak.

An improved 3D-UNet-based brain hippocampus segmentation model based on MR images

ObjectiveAccurate delineation of the hippocampal region via magnetic resonance imaging (MRI) is crucial for the prevention and early diagnosis of neurosystemic diseases. Determining how to accurately and quickly delineate the hippocampus from MRI results has become a serious issue. In this study, a pixel-level semantic segmentation method using 3D-UNet is proposed to realize the automatic segmentation of the brain hippocampus from MRI results. Methods: Two hundred three-dimensional T1-weighted (3D-T1) nongadolinium contrast-enhanced magnetic resonance (MR) images were acquired at Hangzhou Cancer Hospital from June 2020 to December 2022. These samples were divided into two groups, containing 175 and 25 samples. In the first group, 145 cases were used to train the hippocampus segmentation model, and the remaining 30 cases were used to fine-tune the hyperparameters of the model. Images for twenty-five patients in the second group were used as the test set to evaluate the performance of the model. The training set of images was processed via rotation, scaling, grey value augmentation and transformation with a smooth dense deformation field for both image data and ground truth labels. A filling technique was introduced into the segmentation network to establish the hippocampus segmentation model. In addition, the performance of models established with the original network, such as VNet, SegResNet, UNetR and 3D-UNet, was compared with that of models constructed by combining the filling technique with the original segmentation network. Results: The results showed that the performance of the segmentation model improved after the filling technique was introduced. Specifically, when the filling technique was introduced into VNet, SegResNet, 3D-UNet and UNetR, the segmentation performance of the models trained with an input image size of 48 × 48 × 48 improved. Among them, the 3D-UNet-based model with the filling technique achieved the best performance, with a Dice score (Dice score) of 0.7989 ± 0.0398 and a mean intersection over union (mIoU) of 0.6669 ± 0.0540, which were greater than those of the original 3D-UNet-based model. In addition, the oversegmentation ratio (OSR), average surface distance (ASD) and Hausdorff distance (HD) were 0.0666 ± 0.0351, 0.5733 ± 0.1018 and 5.1235 ± 1.4397, respectively, which were better than those of the other models. In addition, when the size of the input image was set to 48 × 48 × 48, 64 × 64 × 64 and 96 × 96 × 96, the model performance gradually improved, and the Dice scores of the proposed model reached 0.7989 ± 0.0398, 0.8371 ± 0.0254 and 0.8674 ± 0.0257, respectively. In addition, the mIoUs reached 0.6669 ± 0.0540, 0.7207 ± 0.0370 and 0.7668 ± 0.0392, respectively. Conclusion: The proposed hippocampus segmentation model constructed by introducing the filling technique into a segmentation network performed better than models built solely on the original network and can improve the efficiency of diagnostic analysis.

Investigation of Appropriate Scaling of Networks and Images for Convolutional Neural Network-Based Nerve Detection in Ultrasound-Guided Nerve Blocks.

Ultrasound imaging is an essential tool in anesthesiology, particularly for ultrasound-guided peripheral nerve blocks (US-PNBs). However, challenges such as speckle noise, acoustic shadows, and variability in nerve appearance complicate the accurate localization of nerve tissues. To address this issue, this study introduces a deep convolutional neural network (DCNN), specifically Scaled-YOLOv4, and investigates an appropriate network model and input image scaling for nerve detection on ultrasound images. Utilizing two datasets, a public dataset and an original dataset, we evaluated the effects of model scale and input image size on detection performance. Our findings reveal that smaller input images and larger model scales significantly improve detection accuracy. The optimal configuration of model size and input image size not only achieved high detection accuracy but also demonstrated real-time processing capabilities.

Impact of Design Decisions in Scanpath Modeling

Modeling visual saliency in graphical user interfaces (GUIs) allows to understand how people perceive GUI designs and what elements attract their attention. One aspect that is often overlooked is the fact that computational models depend on a series of design parameters that are not straightforward to decide. We systematically analyze how different design parameters affect scanpath evaluation metrics using a state-of-the-art computational model (DeepGaze++). We particularly focus on three design parameters: input image size, inhibition-of-return decay, and masking radius. We show that even small variations of these design parameters have a noticeable impact on standard evaluation metrics such as DTW or Eyenalysis. These effects also occur in other scanpath models, such as UMSS and ScanGAN, and in other datasets such as MASSVIS. Taken together, our results put forward the impact of design decisions for predicting users' viewing behavior on GUIs.

Implementing Inception v3, VGG-16 and VGG-19 Architectures of CNN for Medicinal Plant leaves Identification and Disease Detection

Medicinal plants have been a valuable source of healing and healthcare for centuries, serving as a crucial component of traditional medicine systems worldwide. The identification of these plants is a fundamental step in harnessing their therapeutic potential and ensuring their sustainable use. Accurate identification of these plants and the early detection of diseases affecting their leaves are crucial for ensuring a consistent supply of high-quality medicinal resources. In this paper, three types of CNN architectures of deep learning are used to develop robust classification models for distinguishing between healthy and diseased medicinal plant leaves. Inception v3 is known for its versatility in handling various input image sizes without the need for extensive pre-processing, while VGG-19 exhibits high accuracy, robust feature extraction, and suitability for complex image patterns. VGG-16, known for its simplicity and ease of training, provided competitive results, particularly where computational resources were limited. These networks have been pre-trained on large-scale image datasets and fine-tuned using them. These different approaches provide a varied solution to this problem by comparing their accuracy levels, feasibility and to select which one suits the best in this study.

A novel Fourier neural operator framework for classification of multi-sized images: Application to three dimensional digital porous media

Fourier neural operators (FNOs) are invariant with respect to the size of input images, and thus images with any size can be fed into FNO-based frameworks without any modification of network architectures, in contrast to traditional convolutional neural networks. Leveraging the advantage of FNOs, we propose a novel deep-learning framework for classifying images with varying sizes. Particularly, we simultaneously train the proposed network on multi-sized images. As a practical application, we consider the problem of predicting the label (e.g., permeability) of three-dimensional digital porous media. To construct the framework, an intuitive approach is to connect FNO layers to a classifier using adaptive max pooling. First, we show that this approach is only effective for porous media with fixed sizes, whereas it fails for porous media of varying sizes. To overcome this limitation, we introduce our approach: instead of using adaptive max pooling, we use static max pooling with the size of channel width of FNO layers. Since the channel width of the FNO layers is independent of the input image size, the introduced framework can handle multi-sized images during training. We show the effectiveness of the introduced framework and compare its performance with the intuitive approach through the example of the classification of three-dimensional digital porous media of varying sizes.

Vision-based seam tracking for GMAW fillet welding based on keypoint detection deep learning model

Pre-programmed welding robots significantly improved the efficiency and quality of the welds in large batch production. In small and medium batch production, the robots need appropriate sensors to perform well and adapt to the changes and uncertainties in a noisy welding environment. Vision-based sensors enabled by machine learning are making it possible to sense in process previously not measurable. One challenge is developing artificial intelligent models capable of real-time seam tracking, particularly in fillet joints where visual analysis is hindered by non-perpendicular camera angles and arc reflections. In this paper, we propose a vision system that enables automated seam tracking with a collaborative robot. The vision-based deep learning classification model detects the tacks, where the seam is not visible. It is based on a keypoint detection deep learning model that addresses the challenges in distorted and noisy images of fillet joints between the pipes and flanges during the real-time Gas Metal Arc Welding to track the location of the seam in non tack images. The system is optimized for real time seam tracking by proposing the appropriate input image size. Multiple images and multiple points are also considered to provide a controllable signal of the location of the seam with less errors and outliers. Our proposed model can track the seam with more than 80 percent accuracy for errors less than 0.3 mm in fillet joints. The high accuracy of the proposed method would result in fewer flaws and defects and reduced rework, resulting in significant cost saving in manufacturing. The real-time monitoring also enables the adaptability to slight variations in gap.

Inequality-Constrained 3D Morphable Face Model Fitting.

3D morphable model (3DMM) fitting on 2D data is traditionally done via unconstrained optimization with regularization terms to ensure that the result is a plausible face shape and is consistent with a set of 2D landmarks. This paper presents inequality-constrained 3DMM fitting as the first alternative to regularization in optimization-based 3DMM fitting. Inequality constraints on the 3DMM's shape coefficients ensure face-like shapes without modifying the objective function for smoothness, thus allowing for more flexibility to capture person-specific shape details. Moreover, inequality constraints on landmarks increase robustness in a way that does not require per-image tuning. We show that the proposed method stands out with its ability to estimate person-specific face shapes by jointly fitting a 3DMM to multiple frames of a person. Further, when used with a robust objective function, namely gradient correlation, the method can work "in-the-wild" even with a 3DMM constructed from controlled data. Lastly, we show how to use the log-barrier method to efficiently implement the method. To our knowledge, we present the first 3DMM fitting framework that requires no learning yet is accurate, robust, and efficient. The absence of learning enables a generic solution that allows flexibility in the input image size, interchangeable morphable models, and incorporation of camera matrix.

Training a convolutional neural network for real–bogus classification in the ATLAS survey

Abstract We present a convolutional neural network (CNN) for use in the real–bogus classification of transient detections made by the Asteroid Terrestrial-impact Last Alert System (ATLAS) and subsequent efforts to improve performance since initial development. In transient detection surveys, the number of alerts made outstrips the capacity for human scanning, necessitating the use of machine learning aids to reduce the number of false positives presented to annotators. We take a sample of recently annotated data from each of the three operating ATLAS telescope with $\sim$340 000 real (known transients) and $\sim$1030 000 bogus detections per model. We retrained the CNN architecture with these data specific to each ATLAS unit, achieving a median false positive rate (FPR) of 0.72 per cent for a 1.00 per cent missed detection rate. Further investigations indicate that if we reduce the input image size it results in increased FPR. Finally architecture adjustments and comparisons to contemporary CNNs indicate that our retrained classifier is providing an optimal FPR. We conclude that the periodic retraining and readjustment of classification models on survey data can yield significant improvements as data drift arising from changes in the optical and detector performance can lead to new features in the model and subsequent deteriorations in performance.

Application of lightweight YOLOv5 for walnut kernel grade classification and endogenous foreign body detection

The appearance quality grading of walnut kernels is extremely important in the deep processing of walnuts. Foreign body pollution in food is also a challenging problem that has been troubling food companies. Therefore, it is necessary to propose an efficient and non-destructive detection method for walnut kernels and their endogenous foreign bodies. To improve the quality and safety of walnut deep-processed products, this paper proposes an improved detection model based on the original YOLO V5. The improved model achieves a reduction in memory consumption by 38% and computational workload by 34%. Additionally, it increases the average precision of the model by 1.1% and improves CPU speed by 52%. When compared to other lightweight networks, our model demonstrates optimal overall performance. Moreover, our algorithm significantly enhances the precision of detecting small targets and target selection, as compared to the original model.By reducing the input image size of the model, it is able to meet the requirements of real-time detection of walnut kernels and their endogenous foreign bodies on the edge device Raspberry Pi. This provides a technical reference for the non-destructive detection of food quality and its endogenous foreign bodies.

Deep Learning Network Comparison and Validation in Landsat Imagery Cloud Detection

Cloud coverage poses a prevalent challenge in optical remote sensing image processing, significantly hindering the visibility and interpretability of ground‐level information. The application of deep learning technology in remote sensing image cloud detection has witnessed a remarkable advancement, rendering it an increasingly pervasive and potent solution to the challenge of cloud coverage. To thoroughly investigate the effectiveness of diverse deep learning architectures in cloud detection tasks, this research selected five seminal models: AlexNet, VGG16, GoogLeNet, ResNet34, and the cutting‐edge Swin Transformer, and performed a comprehensive comparative analysis of their application in identifying clouds within medium‐resolution Landsat 8 satellite imagery. During the training phase, each model distinctly demonstrated unique strengths in enhancing accuracy, refining loss functions, optimizing resource utilization, and accelerating training processes. In the test phase, the models’ performance was rigorously evaluated using overall accuracy (OA) and other quantitative metrics, providing solid empirical evidence for the variations in their performance. Furthermore, an insightful analysis of the visualization outcomes and a meticulous comparison of the nuanced differences between model predictions and actual cloud images underscored the individual models’ exceptional abilities in capturing intricate cloud details and executing precise edge processing. To alleviate the cumbersome process of pixel‐level labeling, this study adopts the effective block‐level labeling approach. This approach drastically streamlines annotation processes, ultimately translating into substantial cost savings and a marked reduction in time required for data preparation. The experimental results demonstrate that the Swin Transformer network performs particularly well in the cloud detection task of Landsat 8 images, achieving an OA of 90.73% under block‐level annotation, significantly higher than other comparison models, showcasing excellent cloud detection capabilities. Furthermore, the study also found that the size of input image blocks has a significant impact on detection results. Under the same network architecture, using 32 × 32 image blocks for cloud detection improves the OA by at least 10 percentage points compared to using 64 × 64 image blocks. This finding provides important insights for optimizing cloud detection algorithms. In summary, this study not only verifies the superiority of Swin Transformer in cloud detection in remote sensing images but also reveals the impact of image block size on detection performance, providing new ideas and directions for future advancements in remote sensing image processing and cloud detection technology.

City Scale Traffic Monitoring Using WorldView Satellite Imagery and Deep Learning: A Case Study of Barcelona

Accurate traffic data is crucial for a range of different applications such as quantifying vehicle emissions, and transportation planning and management. However, the availability of traffic data is geographically fragmented and is rarely held in an accessible form. Therefore, there is an urgent need for a common approach to developing large urban traffic data sets. Utilising satellite data to estimate traffic data offers a cost-effective and standardized alternative to ground-based traffic monitoring. This study used high-resolution satellite imagery (WorldView-2 and 3) and Deep Learning (DL) to identify vehicles, road by road, in Barcelona (2017–2019). The You Only Look Once (YOLOv3) object detection model was trained and model accuracy was investigated via parameters such as training data set specific anchor boxes, network resolution, image colour band composition and input image size. The best performing vehicle detection model configuration had a precision (proportion of positive detections that were correct) of 0.69 and a recall (proportion of objects in the image correctly identified) of 0.79. We demonstrated that high-resolution satellite imagery and object detection models can be utilised to identify vehicles at a city scale. However, the approach highlights challenges relating to identifying vehicles on narrow roads, in shadow, under vegetation, and obstructed by buildings. This is the first time that DL has been used to identify vehicles at a city scale and demonstrates the possibility of applying these methods to cities globally where data are often unavailable.

Comparison between U-shaped structural deep learning models to detect landslide traces

Landslides endanger lives and public infrastructure in mountainous areas. Monitoring landslide traces in real-time is difficult for scientists, sometimes costly and risky because of the harsh terrain and instability. Nowadays, modern technology may be able to identify landslide-prone locations and inform locals for hours or days when the weather worsens. This study aims to propose indicators to detect landslide traces on the fields and remote sensing images; build deep learning (DL) models to identify landslides from Sentinel-2 images automatically; and apply DL-trained models to detect this natural hazard in some particular areas of Vietnam. Nine DL models were trained based on three U-shaped architectures, including U-Net, U2-Net, and U-Net3+, and three options of input sizes. The multi-temporal Sentinel-2 images were chosen as input data for training all models. As a result, the U-Net, using an input image size of 32 × 32 and a performance of 97 % with a loss function of 0.01, can detect typical landslide traces in Vietnam. Meanwhile, the U-Net (64 × 64) can detect more considerable landslide traces. Based on multi-temporal remote sensing data, a different case study in Vietnam was chosen to see landslide traces over time based on the trained U-Net (32 × 32) model. The trained model allows mountain managers to track landslide occurrences during wet seasons. Thus, landslide incidents distant from residential areas may be discovered early to warn of flash floods.