Simultaneously predicting SPAD and water content in rice leaves using hyperspectral imaging with deep multi-task...

Area of Science:

• Agricultural Science
• Plant Physiology
• Remote Sensing

Background:

• Water and chlorophyll content are key rice growth indicators.
• Phenotypic variations across rice varieties complicate universal model development.
• Hyperspectral imaging offers a non-destructive method for assessing leaf traits.

Purpose of the Study:

• To develop models for simultaneous detection of rice water and chlorophyll content (SPAD values).
• To address challenges in creating universal models due to rice variety differences.
• To evaluate the effectiveness of transfer learning for model adaptation.

Main Methods:

• Utilized hyperspectral imaging to collect data from three rice varieties.
• Developed single-task and multi-task models using partial least squares regression and convolutional neural networks.
• Applied Transfer Component Analysis (TCA) for transfer learning to align feature distributions between varieties.

Main Results:

• Multi-task models showed prediction accuracy comparable to single-task models for individual varieties.
• Transfer Component Analysis (TCA) enabled effective transfer learning, with single-task models performing well across tasks.
• Models utilizing TCA-learned features demonstrated improved and differentiated results for both source and target domains.

Conclusions:

• Multi-task models can simultaneously predict SPAD values and water content.
• Transfer learning via TCA enhances model transferability to new rice varieties.
• This approach improves model construction efficiency and enables rapid detection of rice growth indicators.