Kaaviya is one of the #DigitAg labelled PhDs
Kaaviya defended her thesis on Monday 5th July at 2 pm, at Inrae, Avignon (Alveole room).
Deep learning algorithms for high-throughput cereal plant and organ identification
- Start Date: September 2018
- University: Avignon University
- PhD School: A2E – ED 536 Agrosciences & Sciences
- Field(s): Agronomy, Computer sciences
- Doctoral Thesis Advisor: Frédéric Baret, Inrae, UMR Emmah Avignon et Raul Lopez lozano, Inrae, UMR Emmah Avignon
- Co-supervisors : Frédéric Baret, Inrae, UMR Emmah Capte
- Funding: Cifre Hiphen
- #DigitAg: ,Axis 5:Data Mining, Data Analysis and Knowledge Discovery, Challenge 2: Digital solutions to optimize the genotype in changing production systems and markets
Keywords: IoT, time series, image processing, crop phenology, deep learning, organ detection, spatial representativeness
Abstract: Cereal crops are the most critical source of food for the world population. They are cultivated worldwide for their edible grains which have high nutritional content in terms of energy, protein, carbohydrates, fiber as well as a variety of macronutrients. Cereals thus make an important part of the human diet and livestock feed. The recent advances in plant genomics have generated new opportunities to increase plant genetic variability, with tremendous potentials for crop improvement. However, the effective contribution of these advances to increase crop productivity depends on how tightly genotypic traits can be linked with those eco-physiological mechanisms that produce a distinguishable response of the genotype to the environment. The result of that response is known as phenotype. Plant phenomics –the observation of plant phenotypic traits– is the discipline that must fill the gap between genotype and phenotype. Traditionally, field phenotyping has relied on manual or destructive, low-throughput, observations of phenotypic traits such as plant height, crop development stage, and yield components. The development, in the recent years, of high-throughput field phenotyping platforms and instruments –capable of acquiring and processing efficiently massive volumes of in situ observations over field experiments– has opened a new era of plant phenomics. This has an enormous potential impact on the efficiency of breeding programs, as it would enable plant breeders to phenotype large number of genotypes accurately, thus allowing them to evaluate and identify the best ones. The advances in computer vision and the introduction of deep learning is transforming several traits previously accessible only through manual sampling into high throughput ones. Thanks to their impressive performance, the rapid adoption of these techniques for field plant phenotyping has progressed rapidly in the last five years. The main challenge for the use of deep learning in operational conditions are linked with the lack of generalization where convolutional neural networks are applied over datasets that differ to some extent –i.e. that belong to a different domain– from the dataset used for training them. Compared to the identification of real-world objects, the implementation of deep learning in field phenotyping still has some specific issues that have not been fully addressed by the existing literature. This thesis studies the use of deep learning techniques for the estimation of three essential traits for plant phenotyping: plant density at early stages for maize, wheat head density, and wheat heading date. The thesis is structured into three chapters that take the form of scientific papers, each one dealing with a specific phenotypic trait, and using a specific vector and detection/counting algorithm.
Contact: kaaviya.velumani [AT] inrae.fr
Kaaviya Velumani, SimonMadec, Benoit de Solan, Raul Lopez-Lozano, Jocelyn Gillet, Jeremy Labrosse, Stephane Jezequel, Alexis Comar, Frédéric Baret (2020), An automatic method based on daily in situimages and deeplearning to date wheat heading stage , Field Crops Research – https://www.sciencedirect.com/science/article/abs/pii/S0378429019321604?via%3Dihub