FCA-STNet: Spatiotemporal Growth Prediction and Phenotype Extraction from Image Sequences for Cotton Seedlings.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Existing cotton seedling growth prediction methods struggle with spatiotemporal features and visual fidelity in field settings.
• Challenges include poor feature representation and low texture rendering accuracy.

Purpose of the Study:

• To develop an improved algorithm for predicting cotton seedling growth from images in field environments.
• To enhance spatiotemporal feature representation and visual fidelity in growth prediction models.

Main Methods:

• Proposed FCA-STNet algorithm leveraging historical RGB images for growth prediction.
• Incorporated a novel STNet structure for spatiotemporal dependencies and an Adaptive Fine-Grained Channel Attention (FCA) module.
• FCA module captures global and local features, focusing on individual plants and textures to mitigate field interferences.

Main Results:

• FCA-STNet achieved improved prediction metrics (MSE, MAE, SSIM, PSNR) compared to baseline STNet.
• Outperformed several mainstream spatiotemporal prediction models in experimental tests.
• Predicted phenotypic traits showed high correlation (coefficients > 0.8) with actual measurements.

Conclusions:

• The FCA-STNet model provides visually realistic cotton seedling growth predictions in open-field conditions.
• Offers a robust solution for challenges like lighting variations, leaf flutter, and wind disturbances.
• Represents a significant advancement for research in agricultural growth prediction and phenotyping.