Multiview attention networks for fine-grained watershed categorization via knowledge distillation

Area of Science:

• Artificial Intelligence
• Computer Vision
• Environmental Science

Background:

• Village-related modeling faces limitations in fine-grained watershed classification due to data scarcity and inadequate feature extraction in existing networks.
• Current convolutional networks lack global attention capabilities and are computationally intensive, hindering deployment on edge devices.
• Multi-view watershed classification is crucial for environmental modeling but lacks dedicated datasets and efficient AI solutions.

Purpose of the Study:

• To address the challenges in multi-view fine-grained watershed classification by introducing a novel attention mechanism and knowledge distillation framework.
• To develop the first multi-view watershed classification dataset (MVWD) to facilitate research in this area.
• To create an efficient and accurate model for watershed classification suitable for end-device deployment.

Main Methods:

• Developed the Multi-View Watershed Dataset (MVWD), the first of its kind for fine-grained watershed classification.
• Introduced a Cross-View Attention Module (CVAM) for global attention and salient feature extraction across multiple views.
• Proposed a teacher-student network architecture (MANet-Teacher and MANet-Student) combined with Attention Knowledge Distillation (AKD).

Main Results:

• The MANet-Teacher model achieved state-of-the-art accuracy of 78.51% on the MVWD dataset.
• The lightweight MANet-Student model demonstrated comparable performance with only 6.64M parameters and 1.68G computation.
• MANet-KD effectively balances high performance with computational efficiency for multi-view fine-grained watershed classification.

Conclusions:

• MANet-KD offers a significant advancement in multi-view fine-grained watershed classification, overcoming previous limitations.
• The developed MVWD dataset and MANet-KD framework provide valuable resources for future research in AI for environmental modeling.
• The approach enables accurate and efficient watershed classification, paving the way for practical applications on edge devices.