A novel heuristic target-dependent neural architecture search method with small samples.

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Crop classification is vital for genetic resources and phenotype development.
• Traditional methods are labor-intensive, while convolutional neural networks offer automatic feature identification.
• Developing universal crop classification methods is challenging due to complex crops and scenarios.

Purpose of the Study:

• To address the limitations of manual design in crop classification.
• To introduce an automated approach for creating network architectures for new species.
• To evaluate the effectiveness of a novel target-dependent search method for crop classification.

Main Methods:

• Developed a rapeseed dataset (RSDS) with eight types of rapeseed images.
• Proposed a novel target-dependent neural architecture search (TD-NAS) method based on VGGNet.
• Validated the TD-NAS method on the RSDS, Pl@ntNet, and ICL-Leaf datasets.

Main Results:

• Test accuracy showed no significant difference between small and large samples, indicating limited influence of dataset size on generalization.
• The TD-NAS method demonstrated stable generalization capabilities.
• The method proved effective even with small sample sizes and was free of unpromising detections.

Conclusions:

• The proposed target-dependent neural architecture search (TD-NAS) method offers an efficient and effective solution for crop classification.
• Automated architecture search is a viable alternative to manual design, especially for new or complex species.
• The method's robustness across different datasets highlights its potential for broad application in agricultural AI.