Machine Learning for Crop Weed Classification

From AIRWiki
Revision as of 12:21, 5 February 2018 by LodewijkVoorhoeve (Talk | contribs) (Results)

Jump to: navigation, search
Machine Learning for Crop Weed Image Classification
Short Description: Automatic detection of crops and weeds by using image data
Coordinator:
Tutor: MatteoMatteucci (matteo.matteucci@polimi.it)
Collaborator:
Students: LodewijkVoorhoeve (lodewijk.voorhoeve@wur.nl)
Research Area: Machine Learning
Research Topic: Image Classification
Start: 2017/09/4
End: 2018/01/31
Status: Active
Level: Ms
Type: Thesis

Project short description

The aim of this project is to transfer the work of Stefano Cereda on crop/weed detection to TensorFlow and to train the model with new data.

Dates

Start date: 2017/09/04

End date: 2018/01/30

Project head(s)

M.Matteucci

Students currently working on the project

Lodewijk Voorhoeve

Laboratory work and risk analysis

This project is related to software developing so there are no dangerous activities

Methods

Crop and weed classification can be accomplished by analyzing an image with the use of Machine Learning. Supervised Machine Learning makes use of data (for example: an image) and an annotation file (an image of the same size, with colours demarcating the area that belongs to a certain class). Therefore, the basis of this project is the availability of data.

Data

Two publicly available datasets are used: Crop/Weed Field Image Dataset (CWFID) and Sugar Beets 2016 (SB2016).

Results

With the modified code, both the CWFID and SB2016 datasets were analysed. For SB2016, a comparison was made with the paper by Milioto, Lottes and Stachniss. The comparison is shown in the image on the right, where the baseline classifiers RFC and SVM perform better than the SegNet architecture of Milioto, Lottes and Stachniss on the crop, but worse on the weed and comparable on the soil.

Comparison of crop, weed and soil with the metrics of IoU, Precision and Recall. The baseline classifiers RFC and SVM outperform the SegNet neural network architecture on crop, but perform worse on weed and similar on soil.

The final report can be found here Media:Final_Internship_Report_-_Machine_Learning_for_Crop_and_Weed_Classification_1.0.pdf