Orthogonal Capsule Network with Meta-Reinforcement Learning for Small Sample Hyperspectral Image Classification

Remote Sensing, Journal Year: 2025, Volume and Issue: 17(2), P. 215 - 215

Published: Jan. 9, 2025

Most current hyperspectral image classification (HSIC) models require a large number of training samples, and when the sample size is small, performance decreases. To address this issue, we propose an innovative model that combines orthogonal capsule network with meta-reinforcement learning (OCN-MRL) for small HSIC. The OCN-MRL framework employs Meta-RL feature selection CapsNet data sample. module through clustering, augmentation, multiview techniques enables to adapt new HSIC tasks limited samples. Learning meta-policy Q-learner generalizes across different effectively select discriminative features from data. Integrating orthogonality into reduces complexity while maintaining ability preserve spatial hierarchies relationships in 3D convolution layer, suitably capturing complex patterns. Experimental results on four rich Chinese datasets demonstrate model’s competitiveness both higher accuracy less computational cost compared existing CapsNet-based methods.

Language: Английский

---

One to All: Towards a Unified Model for Counting Cereal Crop Heads Based on Few-Shot Learning

Plant Phenomics, Journal Year: 2024, Volume and Issue: 6

Published: Jan. 1, 2024

Accurate counting of cereals crops, e.g., maize, rice, sorghum, and wheat, is crucial for estimating grain production ensuring food security. However, existing methods cereal crops focus predominantly on building models specific crop head; thus, they lack generalizability to different varieties. This paper presents Counting Heads Cereal Crops Net (CHCNet), which a unified model designed multiple heads by few-shot learning, effectively reduces labeling costs. Specifically, refined vision encoder developed enhance feature embedding, where foundation model, namely, the segment anything (SAM), employed emphasize marked while mitigating complex background effects. Furthermore, multiscale interaction module proposed integrating similarity metric facilitate automatic learning crop-specific features across varying scales, enhances ability describe various sizes shapes. The CHCNet adopts 2-stage training procedure. initial stage focuses latent mining capture common representations crops. In subsequent stage, inference performed without additional training, extracting domain-specific target from selected exemplars accomplish task. extensive experiments 6 diverse datasets captured ground cameras drones, substantially outperformed state-of-the-art in terms cross-crop generalization ability, achieving mean absolute errors (MAEs) 9.96 9.38 13.94 7.94 15.62 mixed A user-friendly interactive demo available at http://cerealcropnet.com/, researchers are invited personally evaluate CHCNet. source code implementing https://github.com/Small-flyguy/CHCNet.

Language: Английский

---

Perfect Labelling: A Review and Outlook of Label Optimization Techniques in Dynamic Earth Observation

Remote Sensing, Journal Year: 2025, Volume and Issue: 17(7), P. 1246 - 1246

Published: April 1, 2025

Advances in Artificial Intelligence (AI) and Machine Learning (ML) have significantly enhanced the practice of Earth Observation (EO), enabling complex analyses such as land cover change detection, vegetation monitoring, disaster response. However, while model architectures matured, refinement reference data remains a major challenge. Accurate dynamic multi-temporal labelling is essential for capturing evolving ground conditions high-dimensional EO datasets, yet key challenges persist, including spatiotemporal inconsistencies, heterogeneous integration, multi-resolution harmonization. Without robust preprocessing, labels may introduce biases, resulting reduced reliability generalizability. This review tackles four core aspects preprocessing EO: (i) steps producing consistent high-quality particularly data; (ii) best practices guidelines that enable scalable accurate workflows across diverse applications; (iii) introduction HELIX framework, unified approach standardizing, enhancing, automating label preprocessing; (iv) forward-looking discussion on future features, next-generation techniques integration. By synthesizing existing methodologies, highlighting emerging approaches, addressing current gaps, this underscores how well-engineered are fundamental to advancing AI/ML-driven applications.

Language: Английский

---

Pseudo-class distribution guided multi-view unsupervised domain adaptation for hyperspectral image classification

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2025, Volume and Issue: 136, P. 104356 - 104356

Published: Jan. 15, 2025

Language: Английский

---

Few-shot SAR image classification via multiple prototypes ensemble

Neurocomputing, Journal Year: 2025, Volume and Issue: unknown, P. 129989 - 129989

Published: March 1, 2025

Language: Английский

---

Few-Shot Remote Sensing Scene Classification via Subspace Based on Multiscale Feature Learning

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Journal Year: 2024, Volume and Issue: 17, P. 13292 - 13307

Published: Jan. 1, 2024

Because of the challenges associated with difficulty accurately labeling remote sensing (RS) scene images and need to identify new classes, few-shot learning has shown significant advantages in addressing classification (RSSC) tasks, leading a growing interest. However, due scale variations targets irrelevant complex background images, current methods exist following problems: problem extraction capability feature extractor mechanism; separability RS classifier. To solve above problems, an approach, called RSSC via subspace based on multiscale is introduced this work. We first design technique address images. Concretely, different branches are utilized learn features at various scales. The self-attention mechanism embedded each branch incorporate understanding global information features. After that, fusion operation, incorporating channel attention, will be devised effectively merge features, so as obtain more precise representation Furthermore, capture shared characteristics category, reduce impact backgrounds results our experiments conducted public available datasets demonstrate strong competitiveness approach.

Language: Английский

---

Charting the evolution: bibliometric perspectives on anomaly detection within hyperspectral domains

Published: April 19, 2024

Language: Английский

---

Remote sensing image scene recognition based on densenet-169

Published: June 13, 2024

Scene classification has become an effective technique for classifying high spatial resolution remote sensing images. However, in the traditional deep learning convolutional neural network, as image passes through layers some of features will be gradually lost, resulting a significant decrease accuracy and precision scene recognition, there is problem underutilization features. In addition, images themselves have complexity. To overcome these challenges, we adopt DenseNet network. Specifically, first train from UCMerced dataset network inputs. Then, introduced DenseNet-169 model based on migration learning. Compared with DenseNet-121, more layers, this difference mainly manifested number dense blocks.DenseNet-169 which increases complexity parameters model, bringing following advantages: stronger expressive power, enables extraction complex feature patterns; faster training time, thanks to densely-connected nature, efficiently utilizes gradient flow; better generalization ability, especially large-scale datasets. our experiments, shows excellent performance compared other state-of-the-art networks dataset, 95.14%, 95.31%, Kappa coefficient 94.90%, F1-score 95.11%. The experimental results show that method can make full use good visual effect, providing recognition.

Language: Английский

---

High-Precision Heterogeneous Satellite Image Manipulation Localization: Feature Point Rules and Semantic Similarity Measurement

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(19), P. 3719 - 3719

Published: Oct. 6, 2024

Misusing image tampering software makes it easier to manipulate satellite images, leading a crisis of trust and security concerns in society. This study compares the inconsistencies between heterogeneous images locate tampered areas proposes high-precision manipulation localization (HSIML) framework distinguish from real landcover changes, such as artificial constructions, pseudo-changes, seasonal variations. The model operates at patch level comprises three modules: preprocessing module aligns filters noisy data. feature point constraint mitigates effects lighting variations by performing matching, applying filtering rules conduct an initial screening identify candidate patches. semantic similarity measurement designs classification network assess RS saliency. It determines consistency based on features implements IML using predefined rules. Additionally, dataset for is constructed images. Extensive experiments compared with existing SOTA models demonstrate that our method achieved highest F1 score both accuracy robustness tests demonstrates capability handling large-scale areas.

Language: Английский

---

Satellite remote sensing image classification based on multiple deep learning algorithms

Applied and Computational Engineering, Journal Year: 2024, Volume and Issue: 57(1), P. 218 - 223

Published: April 29, 2024

This paper explores the use of satellite remote sensing technology to observe and measure earth's surface classify images using various deep learning algorithms. By importing AlexNet, VggNet, GoogleNet MobileNet models, 90% data are randomly selected for training 10% testing, changes in loss accuracy validation set during recorded, as well best round epochs training.During process, set's gradually decreased converged, increased stabilised. The results show that all four models able well, among which highest is model, reaches 99.8%, Googlenet model second one, 97.9%.AlexNet VggNet also have higher accuracy, 93.3% 95.4%, respectively. Overall, this provide an effective solution image classification a reference application algorithms sensing.

Language: Английский

---