• Rodrigo Ortega-Polo

• Etienne Lord, Agriculture and Agri-Food Canada, Canada

Substainable agriculture faces many challenges since it must reconcile both agricultural productivity, while maintaining social, environmental and economic outcomes. Digital agriculture uses the technique of precision agriculture, allowing precise management decisions, while adding some further constraints and algorithmic developments to identify the right crop and the right data needed to take action. Thus, technology opens some solutions to balance those imperatives, a necessary step to the durability of the agricultural sector.

Some of those innovations are the introduction of machine learning, deep learning and now quantum machine learning to extract more information from the now available multilayered datasets. In this talk, we will be revisiting some concepts of digital agriculture and problematic associated with this domain. Then, we will present some of our ongoing machine learning research and robotic platform that could be useful in linking phenology to genetic information. We will also present some imagery datasets developed to help in creating the next generation of machine learning models in agriculture.

• Michael Van de Voorde, VIB-UGent Center for Plant Systems Biology, Belgium
• Sam De Meyer, VIB-UGent Center for Plant Systems Biology, Belgium
• Stijn Hawinkel, VIB-UGent Center for Plant Systems Biology, Belgium
• Daniel Felipe Cruz, VIB-UGent Center for Plant Systems Biology, Belgium
• Tom De Swaef, Flanders Research Institute for Agriculture, Fisheries and Food (ILVO), Belgium
• Peter Lootens, Flanders Research Institute for Agriculture, Fisheries and Food (ILVO), Belgium
• Jolien De Block, VIB-UGent Center for Plant Systems Biology, Belgium
• Heike Sprenger, VIB-UGent Center for Plant Systems Biology, Belgium
• Tom Van Hautegem, VIB-UGent Center for Plant Systems Biology, Belgium
• Dirk Inzé, VIB-UGent Center for Plant Systems Biology, Belgium
• Hilde Nelissen, VIB-UGent Center for Plant Systems Biology, Belgium
• Isabel Roldán-Ruiz, Flanders Research Institute for Agriculture, Fisheries and Food (ILVO), Belgium
• Steven Maere, VIB-UGent Center for Plant Systems Biology, Belgium

Historically, processes influencing plant phenotypes have been studied intensively under controlled laboratory conditions. However, the results of such controlled lab studies often do not translate well to more complex field settings. To help close this lab-field gap, we developed a new experimental setup to study the wiring of plant traits directly in the field, based on omics profiling, micro-environmental profiling and phenotyping of individual plants of the same genetic background grown in the same field. We used this single-plant omics strategy on winter-type rapeseed (Brassica napus) and built models predicting yield phenotypes of field-grown rapeseed plants from their autumnal leaf gene expression and environmental data layers such as soil nutrient profiles and microbiomes at single-plant resolution. Many of the top yield predictors are linked to developmental processes known to occur in autumn in winter-type B. napus accessions, such as the floral transition. We applied methods from the single-cell field on our single-plant data to further unravel these developmental effects.

• Hayda Almeida, Université du Québec a Montréal, Canada
• Nicolas Barbeau Grégoire, Université de Montréal, Canada
• Muhammad Bilal, McGill University, Canada
• Maxime Leduc, McGill University, Canada
• Younes Chorfi, Université de Montréal, Canada
• Xin Zhao, McGill University, Canada
• Jocelyn Dubuc, Université de Montréal, Canada
• Abdoulaye Baniré Diallo, Université du Québec a Montréal, Canada

Metabolic diseases have great impact on dairy production and animal welfare [1, 2]. Metabolomic profiling has helped identify biomarkers to predict disease risk in dairy cows [3, 4]. Previous studies tend to overlook other biomarkers, like from milk production, which could help predict diseases in cows [5]. Our ensemble learner supports predicting disease risk based on heterogeneous biomarkers from metabolomic and health profiles, milk production history, and herd history.
Our datasets contain biomarkers for over 13,700 health events of 1,200 cows from 50 dairy farms in Canada. Biomarkers are captured for a health event e at timepoint t. Given an upcoming lactation Ln, base predictions are obtained for all health events et occurring during lactation Ln−1.
The ensemble learner averages base predictions for an animal and outputs disease probabilities for lactation Ln. Binary classes are disease or non-disease, based on a curated set of nine most common diseases in dairy cows. Classification performance was evaluated for multiple combinations of biomarkers and classifiers.
Classification models based on Logistic Regression and Random Forest classifiers yield best performances, with an average of 0.6 and 0.77 F-measure for disease and non-disease respectively.

• Vahid Naghashi, Universite du Quebec A Montreal, Canada
• Mounir Boukadoum, Universite du Quebec A Montreal, Canada
• Abdoulaye Banire Diallo, Universite du Quebec A Montreal, Canada

In precision livestock, the decision of animal replacement requires an estimation of the lifetime profit of the animal based on multiple factors and operational conditions. In dairy farms, this can be associated with the milk income corresponding to milk production, health condition and herd management costs, which in turn may be a function of other factors including genetics and weather conditions. Estimating the cumulative income from a cow's milk production can be posed as a multivariate time-series prediction task where a late-milk income of a cow can be predicted based on early dairy factors recorded in a sequence of time-steps (lactation months). Furthermore, the predicted milk income would serve as an input to a decision making system for deciding whether to keep or remove an animal in the next lactation period. In this work, a Transformer based model is proposed to predict the cumulative dairy income over the incoming lactation period and further a recommendation procedure is devised for decision making in order to reduce the farmers cost and save their time. In the Transformer, both temporal and inter-variable correlations are captured thanks to the temporal and spatial multi-head attention modules. The proposed framework is assessed using 47749 dairy cows corresponding to more than 5000 herds and the results are compared with the other state-of-the-art models. Our Transformer model outperforms the previous baselines and provides a promising prediction performance with the highest accuracy of 76%, opening the way of better resource management in the dairy industry.

• Cheng-Han Chuang, National Taiwan Ocean University, Taiwan
• Uei-Chen Chiu, National Taiwan Ocean University, Taiwan
• Chang-Wen Huang, National Taiwan Ocean University, Taiwan
• Kuan Y. Chang, National Taiwan Ocean University, Taiwan

Healthy grouper farms thrive on "green water," indicating a balanced ecosystem. A shift to brown water, however, signals potential problems. This study explores how to predict these color changes using water quality data, paving the way for smarter and more sustainable aquaculture.
We analyzed daily monitoring data from six grouper ponds in Fangliao Township, Pingtung, Taiwan, collected from March to December 2018, focusing on water temperature, salinity, and pH. A Long Short-Term Memory (LSTM) model was applied to detect anomalies within these parameters.
Our findings revealed a significant correlation between water quality anomalies and water color changes. Water temperature anomalies were the most effective indicator for early detection of undesired color shifts. Notably, the top 5% of water temperature anomalies successfully predicted over 40% of the water color changes. This means that by targeting the most extreme 5% of water temperature anomalies, farmers can maintain healthier ponds and adopt more sustainable aquaculture methods. Additionally, pH anomalies primarily occurred after the color changes, suggesting a potential consequence rather than a precursor.
The early warning system enables proactive actions like oxygenation and biocontrol, leading to "smart" fish farms with automated water quality management. This will improve efficiency and minimize environmental impact. Future studies will include more factors and data, further refining anomaly detection for sustainable aquaculture.
This study demonstrates the effectiveness of anomaly detection using LSTM models to predict water color changes in grouper ponds. Early detection of water temperature anomalies empowers farmers to manage pond conditions proactively. The prospect of autonomous fish farming systems with real-time monitoring and interventions offers significant potential for sustainable grouper aquaculture. Leveraging innovative technologies and data-driven approaches, we can advance this vital industry sustainably.

• Thomas Gisiger, Université du Québec à Montréal, Canada
• Marjorie Cellier, McGill University, Canada
• Elsa Vasseur, McGill University, Canada
• Abdoulaye Baniré Diallo, Université du Québec à Montréal, Canada

Precision dairy farming is essential to creating a food production system that is durable and respects animal welfare and the environment.

This approach requires gathering many hours of videos with still cameras, which are then used for research, welfare monitoring and tool training. Extracting the video segments with the most meaningful information would allow us to study larger fractions of the recordings taken while gathering the maximum number of observations. This can be framed as a movement detection problem, or, alternatively, a detection problem using traditional or deep learning techniques. However, the latter process of training in a cluttered farm environment might prove challenging.

Here, we propose an algorithm that estimates cow movement in a robust manner without the need for object detection or training. The resulting movement indices, paired with an independently set movement threshold, can then be used to partition videos into episodes where the cow is either immobile or displaying relevant movements and behaviours. This approach takes advantage of typical cow behaviour features and allows for factoring out video sections with repetitive or little/no movement.

The experimental setting consists of five 15-minute videos and focuses on measuring the extent to which discarding episodes with little to no movements speeds up the process of labelling behaviours by animal science experts.

This approach will allow for more complex experiments, novel angles of investigation and larger data-sets to study cow behaviour and interaction with their environment as well as monitoring for welfare status.

• Amanda Ashley Boatswain Jacques, Université du Québec à Montréal, Canada
• Etienne Lord, Agriculture and Agri-Food Canada, Canada
• Vladimir Reinharz, Université du Québec à Montréal, Canada
• Abdoulaye Baniré Diallo, Université du Québec à Montréal, Canada

Implementing site-specific management practices like profitability zones aids in stabilizing long-term profits while conserving the environment. Profitability maps standardize data by converting yield maps (kg/hectare) into $/hectare for different cropping systems. Consequently, they help identify regions in a field which remain more profitable for production, or those that can be transformed into conservation zones. These conservation zones can mitigate economic losses while simultaneously restoring soil health.

Various studies have explored methods for generating management zones using empirical thresholding, fuzzy clustering, algorithmic approaches, and machine learning. The Integer Linear Programming Management Zone delineation method (ILPMZ) has shown promise since it can solve the problem of generating homogenous and rectangular management zones to optimality. However, ILPMZ application in the literature remains limited to soil property maps like pH, Nitrogen, Phosphorus, and Potassium.

In this study, we use the ILPMZ method to generate homogeneous rectangular management zones from profitability maps. Yearly profitability maps are derived from multi-year yield and economic data (2016 to 2021) from a field in Quebec, Canada. This data includes high-resolution yield maps, grain prices, and yearly costs and revenue summaries.

The study underscores the ILPMZ method's potential in optimizing both field profitability and sustainability. It offers valuable insights for future site-specific management practices driven by both ROI and soil conservation.

• Voncarlos Marcelo De Araujo, UQAM, Canada
• Thomas Gisiger, UQAM, Canada
• Sebastien Gambs, UQAM, Canada
• Elsa Vasseur, MCGILL, Canada
• Abdoulaye Baniré Diallo, UQAM, Canada

In modern agricultural management, ensuring accurate estimations of livestock population density within farm environments is essential. Our focus lies in leveraging AI-powered image analysis to detect cows efficiently and accurately within farm environments. Traditional cow detection methods are labor-intensive and invasive, frequently depending on manual labor for image processing. These methods encounter notable challenges, including variations in camera perspectives, object occlusion, and environmental factors. To address these challenges and develop a generalized detection model capable of identifying cows across different camera types, we propose a novel approach. Our research utilizes a dataset of cow images collected from six distinct types of cameras, with three placed inside barns and three positioned outdoors. From this dataset, we selected 2,000 images of cows, representing diverse camera perspectives. The primary challenge lies in creating a model that can effectively detect cows across these varied camera setups. To tackle this, we devised a semi-manual strategy for annotating cow candidates in images captured by different cameras, considering both single and multiple cow individuals. We evaluated the performance of two state-of-the-art object-detection algorithms, Mask-RCNN and YOLOv5, on this diverse dataset. Our fine-tuned CNN-based detector, trained with semi-manual annotations, achieves robust performance in cow detection across diverse camera setups with an mAP@0.5 of 91.10%. It excels in detecting overlapping or occluded cow individuals, crucial for accurate livestock monitoring, boasting a precision of 93.29% and a recall of 88.14%. This approach streamlines farm management and enhances livestock monitoring and resource allocation in agriculture.

• Houda Orchi, Computer Science Department, Université du Québec à Montréal, Montreal, QC H3C 3P8, Canada, Canada
• Abdoulaye Baniré Diallo, Computer Science Department, Université du Québec à Montréal, Montreal, QC H3C 3P8, Canada, Canada
• Elsa Vasseur, Department of Animal Science, McGill University, Sainte-Anne-de-Bellevue, QC H9X 3V9, Canada, Canada
• Halima Elbiaze, Computer Science Department, Université du Québec à Montréal, Montreal, QC H3C 3P8, Canada, Canada
• Essaid Sabir, Computer Science Department, Université du Québec à Montréal, Montreal, QC H3C 3P8, Canada, Canada
• Mohammed Sadik, Department of Electrical Engineering, NEST Research Group ENSEM,Hassan II University, Casablanca, Morocco, Canada

Synchronizing and aligning multi-angle video footage is an intricate task in computer vision, especially in complex environments like barn settings, where observing key behavioral events is critical for understanding dairy cow behavior. Various methods exist to address this challenge, but each has limitations. Frame-by-frame analysis, for instance, is too slow and error-prone for real-time use. Automated techniques like SIFT and SURF are ineffective in low-light or cluttered settings. Additionally, advanced SSM methods falter with dynamic textures and complex movements. To tackle these issues, we introduce a novel synchronization framework consisting of three main components.

First, it leverages YOLOv8-Oriented Bounding Boxes to extract bounding boxes that identify cow movements across videos, forming the cornerstone to construct self-similarity matrices that enable sophisticated comparisons of video sequences, by analyzing similarities through the dynamic positions and trajectories of the identified cows. Second, we enrich these matrices by incorporating advanced descriptors like the Histogram of Oriented Gradients (HOG) and optical flow. HOG aids in differentiating cow behaviors by analyzing edge directions and strengths while optical flow tracks pixel motion across frames. Third, we employ Dynamic Time Warping to identify minimal distances.

Our experiments involve real-time videos captured on the farm, comprising 50 recordings totaling nearly 1500 hours, meticulously curated and annotated by animal scientists. The synchronization quality is examined using assessment metrics like Earth Mover's Distance, Matched Frame Rate, and Mean Temporal Error. This evaluation confirms our proposed framework's effectiveness, demonstrating its potential to revolutionize livestock management and enhance animal welfare through synchronized multi-camera analysis.

• Kush Narang, College of Biological Sciences, University of California, Davis, United States
• Abhigyan Nath, Pt. J.N.M Medical College, India
• William Hemstrom, Department of Biological Sciences, Purdue University, United States
• Kit Sang Chu, Biophysics Graduate Program, University of California, Davis, United States

Across the globe, increasing soil salinity poses a unique and dangerous threat to food supplies. Past work has suggested that modifying critical enzymes is an important approach to confer salt-tolerance adaptations into vital crops. However, most proteins are destabilized by high salinity, and it is difficult to evaluate stability without slow and expensive experimental testing. Therefore, leveraging computational tools to engineer salt-tolerance into existing proteins could serve as a highly effective approach to create more resilient plant populations.

Here, we present HaloClass, an algorithm that utilizes features from ESM-2, a protein language model (pLM), to accurately classify whether a protein is salt-tolerant or not. On multiple benchmark datasets, HaloClass establishes a new state-of-the-art. HaloClass significantly outperforms existing approaches in classification performance metrics and generalization. On 8 pairs of homologous salt-tolerant and non-salt-tolerant proteins independent from the training data, HaloClass classifies only 1 pair incorrectly. Closer structural analysis suggests that surface-exposed charged residues might help HaloClass’s prediction.

Finally, we simulated a guided mutation study using previously published experimental data. The experimental study reported changes in salt-tolerance of nine multiple-point mutants ranging from 2 to 8 sites, HaloClass accurately predicted all changes in salt-tolerance in congruence with experimental data and is the first protein classifier to evaluate mutants accurately on near-identical AlphaFold structures (RMSD < 0.6Å).

This data suggests that HaloClass can facilitate the discovery and guided design of salt-tolerant enzymes, for industrial and agricultural applications. All code for HaloClass will be publicly available on GitHub and Google Colab in our preprint.

• Rodrigo Ortega-Polo