A survey of the vision transformers and their CNN-transformer based variants

Artificial Intelligence Review, Год журнала: 2023, Номер 56(S3), С. 2917 - 2970

Опубликована: Окт. 4, 2023

Язык: Английский

---

Multiclass Land Use and Land Cover Classification of Andean Sub-Basins in Colombia with Sentinel-2 and Deep Learning

Remote Sensing, Год журнала: 2023, Номер 15(10), С. 2521 - 2521

Опубликована: Май 11, 2023

Land Use and Cover (LULC) classification using remote sensing data is a challenging problem that has evolved with the update launch of new satellites in orbit. As are launched higher spatial spectral resolution shorter revisit times, LULC to take advantage these improvements. However, advancements also bring challenges, such as need for more sophisticated algorithms process increased volume complexity data. In recent years, deep learning techniques, convolutional neural networks (CNNs), have shown promising results this area. Training models complex architectures require cutting-edge hardware, which can be expensive not accessible everyone. study, simple CNN based on LeNet architecture proposed perform over Sentinel-2 images. Simple CNNs less computational resources compared more-complex architectures. A total 11 classes were used training validating model, then classifying sub-basins. The analysis showed achieved an Overall Accuracy 96.51% kappa coefficient 0.962 validation data, outperforming traditional machine methods Random Forest, Support Vector Machine Artificial Neural Networks, well state-of-the-art ResNet, DenseNet EfficientNet. Moreover, despite being trained seven million images, it took five h train, demonstrating our only effective but efficient.

Язык: Английский

---

A Multispectral Remote Sensing Crop Segmentation Method Based on Segment Anything Model Using Multistage Adaptation Fine-Tuning

IEEE Transactions on Geoscience and Remote Sensing, Год журнала: 2024, Номер 62, С. 1 - 18

Опубликована: Янв. 1, 2024

Язык: Английский

---

Unveiling the potential of diffusion model-based framework with transformer for hyperspectral image classification

Scientific Reports, Год журнала: 2024, Номер 14(1)

Опубликована: Апрель 10, 2024

Abstract Hyperspectral imaging has gained popularity for analysing remotely sensed images in various fields such as agriculture and medical. However, existing models face challenges dealing with the complex relationships characteristics of spectral–spatial data due to multi-band nature redundancy hyperspectral data. To address this limitation, we propose a novel approach called DiffSpectralNet, which combines diffusion transformer techniques. The method is able extract diverse meaningful features, leading improvement HSI classification. Our involves training an unsupervised learning framework based on model high-level low-level followed by extraction intermediate hierarchical features from different timestamps classification using pre-trained denoising U-Net. Finally, employ supervised transformer-based classifier perform We conduct comprehensive experiments three publicly available datasets validate our approach. results demonstrate that significantly outperforms approaches, achieving state-of-the-art performance. stability reliability are demonstrated across classes all datasets.

Язык: Английский

---

Joint superpixel and Transformer for high resolution remote sensing image classification

Scientific Reports, Год журнала: 2024, Номер 14(1)

Опубликована: Март 1, 2024

Abstract Deep neural networks combined with superpixel segmentation have proven to be superior high-resolution remote sensing image (HRI) classification. Currently, most HRI classification methods that combine deep learning and use stacking on multiple scales extract contextual information from segmented objects. However, this approach does not take into account the dependencies between each object. To solve problem, a joint Transformer (JST) framework is proposed for In JST, first objects as input, used model long-range dependencies. The relationship input object obtained class of analyzed output by designing an encoding decoding Transformer. Additionally, we explore effect semantic range accuracy. JST also tested using two datasets overall accuracy, average accuracy Kappa coefficients 0.79, 0.70, 0.78 0.91, 0.85, 0.89, respectively. effectiveness method compared qualitatively quantitatively, results achieve competitive consistently better than benchmark comparison method.

Язык: Английский

---

An efficient parallel DCNN algorithm in big data environment

PeerJ Computer Science, Год журнала: 2025, Номер 11, С. e2871 - e2871

Опубликована: Май 20, 2025

Big data plays a vital role in developing remote sensing, landslide prediction, and enabling applications, the integration of deep convolutional neural networks (DCNN) has significantly improved its prediction accuracy. However, several challenges remain processing vast satellite imagery other geospatial data. These include excessive redundant features, slow convolution operation, poor loss function convergence. An efficient parallel DCNN algorithm (PDCNN-MI), combined with MapReduce Im2col algorithms, is introduced to address these challenges. First, feature extraction strategy based on Marr-Hildreth operator (PFE-MHO) proposed extract target features from as inputs network, effectively solving problem high redundancy. Next, model training method (PMT-IM) designed remove kernels by designing center value distance, improving operation speed. Finally, small batch gradient descent (IMBGD) presented exclude influence anomalous nodes solve convergence function. By utilizing enhancements, experimental results indicate that PDCNN-MI outperforms existing algorithms classification accuracy well-suited for fast large-scale image dataset processing.

Язык: Английский

---

Land Cover Classification of Remote Sensing Images Based on Hierarchical Convolutional Recurrent Neural Network

Forests, Год журнала: 2023, Номер 14(9), С. 1881 - 1881

Опубликована: Сен. 15, 2023

Convolutional neural networks (CNNs) and recurrent (RNNs) have gained improved results in remote sensing image data classification. Multispectral classification can benefit from the rich spectral information extracted by these models for land cover This paper proposes a model called hierarchical convolutional network (HCRNN) to combine CNN RNN modules pixel-level of multispectral images. In HCRNN model, original 13-band Sentinel-2 is transformed into 1D sequence using fully connected layer. It then reshaped 3D feature matrix. The 2D-CNN features are used as inputs corresponding RNN. at each level adapted same convolution size. structure leverages advantages CNNs RNNs extract temporal spatial data, leading high-precision experimental demonstrate that overall accuracy on dataset reaches 97.62%, which improves performance 1.78% compared model. Furthermore, this study focused changes forest area Laibin City, Guangxi Zhuang Autonomous Region, was 7997.1016 km2, 8990.4149 8103.0020 km2 2017, 2019, 2021, respectively, with an trend small increase covered.

Язык: Английский

---

A Fourier Frequency Domain Convolutional Neural Network for Remote Sensing Crop Classification Considering Global Consistency and Edge Specificity

Remote Sensing, Год журнала: 2023, Номер 15(19), С. 4788 - 4788

Опубликована: Сен. 30, 2023

The complex remote sensing image acquisition conditions and the differences in crop growth create many classification challenges. Frequency decomposition enables capture of feature information an that is difficult to discern. domain filters can strengthen or weaken specific frequency components enhance interclass among different crops reduce intraclass variations within same crops, thereby improving accuracy. In concurrence with Fourier learning strategy, we propose a convolutional neural network called (FFDC) net, which transforms maps from spatial spectral domain. this network, dynamic filtering are used separate into low-frequency high-frequency components, strength distribution automatically adjusted suppress crop, enhancing overall consistency crops. Simultaneously, it also widen achieve high-precision classification. test areas, randomly selected multiple farms located far sampling area, compare our method other methods. results demonstrate frequency-domain approach better mitigates issues, such as incomplete extractions fragmented boundaries, leads higher accuracy robustness. This paper applies deep classification, highlighting novel effective solution supports agricultural management decisions planning.

Язык: Английский

---

A Social Community Sensor for Natural Disaster Monitoring in Indonesia Using Hybrid 2D CNN LSTM

Опубликована: Окт. 24, 2023

One of the most notable advantages utilizing social media is its ability to aid communication during natural disasters, specifically in regards disseminating information. Social users serve as community monitors by collecting public messages from a multitude platforms. The implementation artificial intelligence can facilitate automatic recognition concerning disasters. Deep learning-based models for text classification, such Convolutional Neural Network (CNN) and Long Short-Term Memory networks (LSTM), each possess their own unique strengths weaknesses. Nevertheless, utilization word padding techniques presents further challenge accurately classify texts, may negatively impact classification performance. In this study, feature extraction based on embedding were employed, maximum number words be processed hybrid 2D CNN LSTM model. results indicate highest level accuracy cases floods, with an 81.27% rate, forest fires 86.14% earthquakes 80.16% rate. This outcome represents significant advancement over attained model constructed mean using solely CNN.

Язык: Английский

---

Integrating Transformer-Based Deep Learning and Vegetation Indices for Accurate Mapping of Rainfed Smallholder Farms in Semi-Arid Regions of South Africa Using Google Earth Engine

Опубликована: Янв. 1, 2024

Язык: Английский

---