Keywords. Computer vision, drone imaging, pattern recognition, semi-supervised deep learning, time series, data fusion, digital agriculture.
Weeds compete directly with crops for moisture, nutrients, and sunlight. As such, they have a significant negative impact on crop yields if their presence is not adequately controlled. According to the New Zealand environmental research organization, Land Care, weeds have been responsible for approximately $95 billion in food production losses worldwide. The Food and Agriculture Organization of the United Nations (FAO) considers that weeds should be recognized as the number one natural enemy for agricultural production (FAO 2009). Weed control has always been considered a major challenge for agricultural production. Current approaches and trends in perception (image sensor and localization technologies, image acquisition and processing methodologies), robotics, and artificial intelligence (AI) in general, are paving the way for promising new advances. Today there is a trend towards digital and robotic agriculture to solve different problems and improve working conditions in agricultural fields.
The detection and recognition of plants from optical sensors are among the major challenges for automatic weed control. Despite the considerable progress made in recent years in automatic visual recognition methods, especially through deep learning, the ability to discriminate between vegetation remains limited. This is the case when variations in lighting conditions, occlusion between plants, changes in terrain, as well as crop development stages are observed… One of the key points for the success of a deep learning algorithm is the abundance of labeled data, which is a limitation in digital agriculture. The labeling process usually requires the intervention of experts, which is one of the main limitations to build good robust and generalizable models with deep neural networks. The work of the thesis will address this type of problem by developing new methods to map weeds from high resolution drone images, with image analysis and machine learning algorithms (semi-supervised learning, weak learning, generative learning, attention mechanism, transformers…). This mapping will be improved by considering different types of data such as plant biology, weather, … In particular, the predictive model fed by heterogeneous multisource data (regional weather, local sensors, history of mapping …) should allow to predict the probability of presence and growth of plants locally and thus improve the detection. Whether it is on plant recognition or prediction, new contributions in machine learning will be studied.
The supervision and the direction of the thesis will be ensured by researchers of the laboratories PRISME (INSA CVL-University of Orléans) and LIFAT (University of Tours). The thesis is financed by a regional research project. It will take place at INSA CVL in Bourges.
- Master 2 and/or engineering school
- Skills in computer vision and machine learning
- Knowledge in robotics will be appreciated but not essential.
- Skills in Python, C/C++ development, …
- Good level in English language
How to apply
Send by email to the contacts below: a CV, cover letter and transcripts before July 08, 2021.
 A S M Mahmudul Hasan, Ferdous Sohel, Dean Diepeveen, Hamid Laga, Michael G.K. Jones, A survey of deep learning techniques for weed detection from images, Computers and Electronics in Agriculture, V 184, 2021.
 M-D. Bah, A. Hafiane, R. Canals. Deep Learning with Unsupervised Data Labeling for Weed Detection in Line Crops in UAV Images, Remote Sensing, Volume 10, 2018.
 M-D. Bah, A. Hafiane, R. Canals, B. Emile, Deep Features and One-class Classification with Unsupervised Data for Weed Detection in UAV Images, International Conference on Image Processing Theory, Tools and Applications (IPTA), 2019
 M-D. Bah, E. Dericquebourg, A. Hafiane, R. Canals Deep Learning based Classification System for Identifying Weeds using High-Resolution UAV Imagery, IEEE Computing Conference, 2018
 Ishan Misra, Abhinav Shrivastava, Martial Hebert. Watch and Learn: Semi-Supervised Learning of Object Detectors from Videos, https://arxiv.org/abs/1505.05769
 Hieu Pham, Zihang Dai, Qizhe Xie, Minh-Thang Luong, Quoc V. Le. Meta Pseudo Labels, https://arxiv.org/abs/2003.10580
 Luyu Yang, Yan Wang, Mingfei Gao, Abhinav Shrivastava, Kilian Q. Weinberger, Wei-Lun Chao, Ser-Nam Lim. Deep Co-Training with Task Decomposition for Semi-Supervised Domain Adaptation, https://arxiv.org/abs/2007.12684