Estimation of Botanical Composition in Mixed Clover-Grass Fields Using Machine Learning-Based Image Analysis.

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• Accurate botanical composition (BC) estimation is crucial for effective forage management in clover-grass mixtures.
• Traditional methods for BC assessment can be labor-intensive and time-consuming.
• Image analysis offers a potential solution for automated and efficient BC estimation.

Purpose of the Study:

• To develop and evaluate an effective image analysis method for clover detection and BC estimation in clover-grass fields.
• To compare the performance of different transfer learning models for clover fraction detection.
• To build and assess regression models for BC estimation using detected clover fraction and plant height data.

Main Methods:

• Three transfer learning models (DeepLab V3+, SegNet, FCN-8s) were employed for clover fraction detection.
• Multiple linear regression (MLR) and backpropagation neural network (BPNN) models were developed for BC estimation.
• A dataset of 347 clover-grass images, including augmented data, was used for training, validation, and independent testing.

Main Results:

• DeepLab V3+ achieved the highest intersection-over-union (IoU) of 0.73 for clover fraction detection.
• The best transfer learning model (DeepLab V3+) yielded a root mean square error (RMSE) of 8.5% for clover fraction detection.
• The BPNN model incorporating clover fraction, clover height, and grass height achieved the lowest BC estimation RMSE of 6.6%, outperforming MLR and models using only clover fraction.

Conclusions:

• DeepLab V3+ demonstrates superior performance in clover fraction detection compared to SegNet and FCN-8s.
• Backpropagation neural networks provide more accurate BC estimation than multiple linear regression.
• The combined approach of DeepLab V3+ for detection and BPNN for estimation offers a promising method for improving forage management through accurate BC assessment.