Crops and weeds classification using Convolutional Neural Networks via optimization of transfer learning parameters
Abdel-Aziz Binguitcha-Fare1, Prince Sharma2

1Abdel-Aziz Binguitcha-Fare, Research Scholar, AP Goyal Shimla University, School of Science and Technology, Shimla (H.P), India.
2Prince Sharma, Assistant Professor, AP Goyal Shimla University, School of Science and Technology, Shimla (H.P), India.

Manuscript received on 18 June 2019 | Revised Manuscript received on 25 June 2019 | Manuscript published on 30 June 2019 | PP: 2284-2294 | Volume-8 Issue-5, June 2019 | Retrieval Number: E7727068519/19©BEIESP
Open Access | Ethics and Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (

Abstract: Agriculture remains the backbone of several economies in the world, especially in underdeveloped countries. With the rapid growth of the population and the increasing demand in food, farmers need to maximize the productivity and one possibility is the reduction of losses. Weeds are one of the major dangers in farming.Indeed, they competevigorously with the crop for nutrients and water. As result, they can cause the loss of 10% to 100% of the total harvest. This work aimed at developing a new model tailored to classify crops and weeds images. Using a pubic dataset of 5339 plant images from Aarhus University Signal Processing group in collaboration with University of Southern Denmark, we proposed a methodology based on transfer learning technique to classify 12 species of crops and weeds. Firstly, we converted images to jpeg format in order to accelerate the convergence and data augmentation techniques such as resizing, rotating, flipping, scaling were employed to reduce the chances of overfitting. Then, a model trained on ImageNet dataset with Residual Network 101 architecture was used for performing transfer learning. Finally, the network’s parameters were adjusted through various techniques involving progressive resizing, cyclical learning rate and focal loss function for improving the performance. Our model achieved an overall accuracy of 98,47% during validation and of 96,04% on the test set. We already deployed the model over Internet through a web application and our next step will be to integrate this solution within a mobile application and embedded devices.Future works concerns with the use of more features and descriptors to accurately distinguish two specificclasses of weeds: Black-grass and Loose Silky-Bent, and the possibility to extend our approach to other kinds of plants.
Keywords: Precision Agriculture, Convolutional Neural Networks, Transfer Learning, Residual Networks.

Scope of the Article: Deep Learning