We propose a computer vision and deep learning-based method of detecting sugar beet leaf damage rates using a convolutional neural network that can be incorporated into a smartphone application. We adopt a transfer learning mechanism by pre-training the model on plant-related images for plant classification task to improve the performance of the damage regression model. We visualize the learned features of our trained model to see how it recognizes the pattern of the damaged sugar beets. Our trained model achieves a RMSE of 1.71 on our held-out test set, hence demonstrating the feasibility of a deep learning-based approach. Our work can play an important role in providing a clear path toward automated and efficient crop damage detection system which can bring lots of benefits to both farmers and agriculture insurers.

Create conda environment with packages to be installed.

conda env create --file=idp_yoonha.yaml

To pre-train the plant classification model:

python pretrain_plant.py --epochs NUM_OF_EPOCHS

To train the damage regression model:

python train_regression.py --epochs NUME_OF_EPOCHS

To fit the linear damage regression model:

python linear_generate_training_test.py
python linear_fitting.py

To evaluate the trained damage regression model:

python test_regression.py

To evaluate the linear damage regression model:

python test_linear.py

To visualize the learned feature map of damage regression model:

python featurevisualizer.py