Crop Classification on Hyperspectral Images Using Residual Network and Multi-Head Self-Attention Feature Extraction

Nowadays, hyperspectral imaging has empowered classification and prediction in many fields such as agriculture, land usage, tourism, etc. due to advancements in deep learning. This research focuses on obtaining efficient crop classification using Residual Network, Multi-Head Self- Attention, Convolu...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2024 International Conference on Data Science and Network Security (ICDSNS) s. 01 - 04
Hlavní autoři: Naga Saranya, N., Vivekanandhan, V., Al-Farouni, Mohammed, Shreyas, A V, Ranjith kumar, Gotte
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 26.07.2024
Témata:
Accuracy
Convolutional neural networks
Crops
Degradation
Error analysis
Feature extraction
Filters
high-resolution
hyperspectral images
Hyperspectral imaging
multi-head self-attention
Noise
residual learning
residual network
Residual neural networks
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Nowadays, hyperspectral imaging has empowered classification and prediction in many fields such as agriculture, land usage, tourism, etc. due to advancements in deep learning. This research focuses on obtaining efficient crop classification using Residual Network, Multi-Head Self- Attention, Convolutional Neural Network (ResNet-MHSA- CNN) approach. The Adaptive Wiener Filtering (AWF) is applied in preprocessing to eliminate noise, handles blurred parts of images effectively and decreases error rate in inverse filtering. The Residual Network (ResNet) method is employed by implement residual learning to eliminate degradation issues of deep neural networks. The high-resolution depth features of hyperspectral images are extracted by residual learning ResNet network. The Multi-Head Self-Attention (MHSA) utilized to capture long-range dependencies and global context information and helps the model in learning diverse patterns. Finally, the extracted features are given to Convolutional Neural Network (CNN) with complex pattern learning ability that excels in image classification with multiple layers of filters. The proposed ResNet-MHSA-CNN approach is evaluated with Indian Pines dataset and achieved 99.74%, 99% accuracy and kappa value respectively. The proposed method outperformed state-of-art methods such as Cross-Mixing Residual Network of Convolutional Neural Network (CMR-CNN), Deep Stacked Denoising Autoencoder (DSDA).
DOI:10.1109/ICDSNS62112.2024.10691054