UTokyo-FieldPhenomics-Lab / heading-date-estimation

Automatic Estimation of Heading Date of Paddy Rice using Deep Learning

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Introduction

This is an implementation of submitted paper:

Desai, S. V., Balasubramanian, V. N., Fukatsu, T., Ninomiya, S., & Guo, W. (2019). Automatic estimation of heading date of paddy rice using deep learning. Plant Methods, 15(1), 76. https://doi.org/10.1186/s13007-019-0457-1

To plan the perfect time for harvest of rice crops, we detect and quantify the flowering of paddy rice. The dataset of rice crop images used in this work is taken from [1].

Methodology

The basic outline of the method is as follows. We use an SVM classifier with a sliding window to detect flowering panicles in the image. To generate the features to be given as input to the SVM classifier, we use two methods of feature extraction and compare their performances.

  • Feature extraction using SIFT Descriptors
  • Feature extraction using a deep neural network (VGG-16 [2])

Training is done using patches sampled from 21 images of Kinmaze dataset. Testing is done on the full 5184x3456 images from Kinmaze dataset. [1].

How to Run

To obtain training accuracy :

python training_accuracy/sift/sift_training_accuracy.py

python training_accuracy/vgg/vgg_training_accuracy.py

To obtain correlation :

python correlation/kinmaze_data/vgg/vgg_testing_on_kinmaze.py path_to_directory_containing_test_images/ output_probs.xlsx

Results

The ROC curves obtained for the SVM classifier after performing cross validation on patches are as follows :

SIFT Features with D=5 Pixels VGG Features

Flowering panicles detected in one of the images of the Kinmaze dataset:

The correlation results obtained on the whole-images are as follows :

Approach Pearson Correlation Coefficient
VGG 0.7495
SIFT 0.5435

The computational cost for VGG based method is as follows :

Task Time Taken
Feature extraction using VGG and training the SVM classifier 42 seconds
Detecting flowering panicles 29 seconds per image

GPU Used : NVIDIA Geforce GTX 1080Ti (11 GB Memory)

Acknowledgements

Paper written under the able guidance of :

Dr. Vineeth N Balasubramanian, IIT Hyderabad, India.

Dr. Wei Guo, University of Tokyo, Tokyo.

Code Contributors

Manasa Kumar [github]

Sai Vikas [github]

References

[1] Guo W, Fukatsu T, Ninomiya S. Automated characterization of flowering dynamics in rice using field-acquired time-series RGB images. Plant Methods. 2015;11:7. doi:10.1186/s13007-015-0047-9

[2] Karen Simonyan, Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv:1409.1556

About

Automatic Estimation of Heading Date of Paddy Rice using Deep Learning


Languages

Language:Python 87.3%Language:Jupyter Notebook 12.7%