A novel transfer learning framework for sorghum biomass prediction using UAV-based remote sensing data and genetic...

Area of Science:

• Agricultural Science
• Biotechnology
• Data Science

Background:

• Biomass yield is critical for biofuel crop breeding programs, but traditional measurement methods are destructive, time-consuming, and labor-intensive.
• Modern remote sensing platforms, like unmanned aerial vehicles (UAVs), offer efficient, non-invasive data collection for phenotypic traits.
• Accurately modeling the relationship between remote sensing data and biomass is challenging due to limited ground-truth data for genotypes.

Purpose of the Study:

• To develop an accurate and efficient model for predicting sorghum biomass using time-series remote sensing and weather data.
• To address the challenge of limited ground reference data in breeding experiments by employing advanced machine learning techniques.
• To improve the generalization capability of biomass prediction models through transfer learning strategies.

Main Methods:

• A Long Short-Term Memory (LSTM) based Recurrent Neural Network (RNN) model was designed to integrate time-series remote sensing, weather, and static genotypic data.
• Feature importance analysis was performed to reduce redundant features derived from high-dimensional remote sensing data.
• Transfer learning strategies were implemented to select informative training samples, enabling model refinement with limited data.

Main Results:

• The proposed LSTM-based RNN model achieved high prediction accuracy for sorghum biomass within a single year.
• Feature importance analysis successfully identified and removed redundant remote sensing features.
• Transfer learning strategies allowed a pre-trained model to be refined with limited target-domain samples, yielding accuracy comparable to models trained from scratch.

Conclusions:

• The LSTM-based RNN model effectively predicts sorghum biomass, outperforming traditional methods by utilizing non-invasive, time-series data.
• The developed transfer learning approach significantly enhances model generalization and reduces the dependency on extensive ground-truth data.
• This methodology offers a scalable and accurate solution for biomass estimation in biofuel crop breeding programs, accelerating genetic improvement efforts.