The paper presents MCapsNet, an advanced deep learning model combining a modified Capsule Network with attention mechanisms to improve crop pest recognition, outperforming traditional CNNs and showing strong potential for real-world agricultural use.
This paper proposes a feature fusion network combining CNN and Transformer models with an attention-selection mechanism for robust insect pest recognition, achieving superior performance on the IP102 dataset and offering broader applications in visual classification tasks.
The Sentinel satellites, Sentinel-2A and Sentinel-2B, which are carefully managed by the European Space Agency (ESA) for Earth observation, provide a wealth of data that contributes to the robustness of the EuroSAT dataset.
VGG16 and ResNet50
| Metric | M_MS_VGG16 | M_MS_ResNet50 |
|---|---|---|
| Entropy loss | 0.1767 | 3.0015 |
| Accuracy | 95.78% | 10.50% |
—Building and testing deep learning models to classify EuroSAT satellite images, with an emphasis on the multispectral dataset. Data preprocessing, model construction, and evaluation are the various stages that make up the methodology. Scaling from minimum to maximum and balancing classes are applied to the EuroSAT dataset. The classification process makes use of the VGG16 and ResNet50 models implemented in a convolutional neural network (CNN) architecture. During training, the model is made more resilient by using data augmentation. Criteria like classification accuracy, categorical cross-entropy loss, and class-specific evaluations are all part of the review. As far as accuracy and loss go, the VGG16-based model beats ResNet50, which means it should work well with the RuroSAT multispectral dataset. Class-specific evaluations reveal variations in land cover classification accuracy. Makes a significant contribution to Eu roSAR image classification by providing trustworthy models that highlight the significance of state-of-the-art CNN architecture, careful dataset preparation, and stringent evaluation methods. Visualization is a powerful tool for comprehending the dynamics of model learning and performance during the testing and training phases.