Artificial intelligence (AI) is the trending topic of 2024. Even though it’s not the main focus of popular culture, researchers from agriculture, biology, and technology are increasingly utilizing AI. They work together to find effective ways for these algorithms and models to sift through data to comprehend and forecast a world affected by climate change.
A recent article published in Frontiers in Plant Science showcases the work of Claudia Aviles Toledo, a geomatics PhD candidate at Purdue University, along with her faculty advisors Melba Crawford and Mitch Tuinstra. They have showcased the capabilities of a recurrent neural network—a model that trains computers to process information using long short-term memory—to predict maize yields using various remote sensing technologies combined with environmental and genetic data.
Plant phenotyping involves examining and detailing plant characteristics, which can be a labor-intensive process. For example, measuring plant height using a tape, assessing reflected light across different wavelengths with cumbersome handheld equipment, and harvesting individual plants for chemical analysis are all demanding and costly tasks. Fortunately, remote sensing techniques harnessing uncrewed aerial vehicles (UAVs) and satellites are making it easier to access field and plant information.
According to Tuinstra, who holds the Wickersham Chair of Excellence in Agricultural Research and is a professor in plant breeding and genetics at the agronomy department, “This study showcases how progress in UAV-based data gathering and advanced deep-learning networks can assist in predicting complex traits in food crops like maize.”
Crawford, a distinguished professor in Civil Engineering and Agronomy, acknowledges Aviles Toledo and her team’s diligent efforts to collect phenotypic data both directly in the field and through remote sensing. Projects like this demonstrate how remote sensing-based phenotyping can simultaneously minimize labor needs and gather unique plant data that would normally be undetectable by human senses.
Advanced hyperspectral cameras capable of capturing precise reflectance measurements across wavelengths beyond the visible spectrum can now be mounted on robots and UAVs. Light Detection and Ranging (LiDAR) tools emit laser pulses and track the time it takes for reflections to return to the sensor, creating “point clouds” that map the geometric structure of plants.
“Plants tell their own stories,” Crawford stated. “They respond to stress in their environment. Observing these reactions can help relate them to various traits, environmental factors, and management practices like fertilizer usage, irrigation, or pest control.”
As engineers, Aviles Toledo and Crawford develop algorithms that manage substantial datasets to uncover patterns for predicting various outcomes, including the yield of diverse hybrids bred by professionals like Tuinstra. These algorithms can identify healthy and stressed plants before they are observable to farmers or scouts, and they offer insights into the effectiveness of different farming strategies.
Bringing a biological perspective to the research, Tuinstra notes that plant breeders use data to pinpoint specific genes responsible for certain crop traits.
“This represents one of the pioneering AI models that incorporates plant genetics into the yield dynamics observed over multiple years and extensive plots,” Tuinstra explained. “This enables plant breeders to understand how various traits respond to different settings, assisting them in selecting traits for new, more resilient crop varieties. Farmers can also utilize this information to determine which varieties are likely to perform best in their regions.”
The neural network was developed by merging remote-sensing hyperspectral and LiDAR data from corn, genetic markers of well-known corn varieties, and environmental data sourced from weather stations. This advanced deep-learning approach, a category of AI, learns from spatial and temporal data patterns to forecast future outcomes. Once educated in a specific location or time frame, the network can adapt to new geographic areas or time intervals with minimal retraining, thereby reducing reliance on reference data.
Crawford remarked, “Previously, we employed conventional machine learning techniques focusing heavily on statistics and mathematics. We couldn’t fully leverage neural networks due to insufficient computational power.”
Neural networks resemble a mesh of chicken wire, where connections link various points, allowing every point to communicate with all others. Aviles Toledo adjusted this model with long short-term memory, which keeps previous data constantly at the forefront alongside current data while making future predictions. This long short-term memory model, enhanced by attention mechanisms, also highlights key physiological stages in the growth cycle, such as flowering time.
Although the remote sensing and meteorological data are integrated into this innovative structure, Crawford notes that genetic data is still processed to extract “aggregated statistical features.” Together with Tuinstra, Crawford aims to more meaningfully integrate genetic markers into the neural network and include more complex traits in their dataset. Achieving this would not only reduce labor costs but also provide farmers with valuable insights for making optimal decisions regarding their crops and land.
