Remember in Interstellar how crops were failing everywhere on Earth, with dust storms and droughts making farming nearly impossible? While the rest of this movie is science fiction, climate change is turning that agricultural crisis into reality. Over seven hundred million people suffer from hunger today, and this number continues to rise as climate change reduces both food quality and quantity. Drought stress is one of the most pressing threats to agriculture, especially in low-and middle-income countries (LMICs), where it is the leading environmental cause of crop failure.
Measuring the effects of drought stress on plants can help us face this situation. For example, farmers can optimise crop watering by measuring how much water they need. Researchers can also use these measurements to evaluate which crop varieties show better drought resistance or what management practices can help plants fare better with less water. However, precisely measuring drought stress is often a complicated process that requires expensive, specialised equipment, expertise, and time. Researchers and farmers alike need efficient and affordable new tools to tackle this challenge.
Seeing Plants with AI
The boom of AI in recent years has brought innovative solutions in many fields, including digital agriculture. For instance, chatbots like Darli AI are now being deployed to help farmers learn about regenerative farming. Meanwhile, companies such as SenzAgro use Internet of Things (IoT) sensors to monitor environmental conditions and help farmers make informed decisions. AI is quite powerful for all tasks related to images. New models can detect subtle patterns and automatically measure or compare a wide range of parameters, making them ideal for applications in agriculture. These techniques are exciting when considering affordable research, as mobile phone cameras are widely accessible to people in almost every country.
For example, detecting diseases from plant images is now an active avenue of research for many species, with many applications in apps and on-field cameras. However, image-based drought detection has mostly focused on costly approaches that do not always scale up for smallholder farms, such as drone and satellite pictures or hyperspectral images.
How Do You Measure Drought Stress in Plant Pictures?
As drought detection still lacks a fast, simple, and cost-efficient tool, I am using AI to create a tool to detect drought stress from standard images. For that, I am using a technique called supervised machine learning. This means that I am training an algorithm with plant pictures taken at various stages of drought, each paired with a label that quantifies stress levels. This project involves four key challenges:
• Building a dataset: The algorithm must be trained with a dataset of images and labels. For that, I collected a dataset with plant images and precise drought stress markers that reflect a plant’s internal status (eco-physiological markers).
• Defining drought stress labels: Drought stress is a continuum that requires numerical values to be estimated. However, it cannot be measured with a single instrument, and there are no established classes of drought stress using eco-physiological markers that can be used as labels. Therefore, the labelling process must integrate multiple eco-physiological markers into an objective label that accurately represents plant status.
• Training AI with limited data: Most machine learning models require thousands of images, but this project works with a much smaller dataset—just a few hundred images. Thus, I am using specialised AI strategies designed for low-data environments.
• Ensuring explainability: AI decisions should not be a black box. To build trust in the model’s predictions, I need to develop a way to highlight which parts of an image the algorithm relies on when assessing drought stress.
Machine Learning Drought Detection
The first step in developing an AI-powered drought detection tool was to build a dataset of plant images paired with eco-physiological markers. I grew plants under controlled conditions, then withheld water and took measurements throughout the drying process. I captured images of the plants along with key drought response indicators, such as relative water content in leaves and gene expression levels.
I then used these physiological measurements to create objective drought stress categories to serve as training labels. This was done through HDBSCAN, a clustering algorithm that can group points objectively based on how far they are from each other when looking at their values for a range of parameters – here, physiological measurements. This approach resulted in stress classes ranging from low to high drought stress, each associated with mean values of physiological parameters.
We can then use these labels to train a supervised machine learning model, such as a convolutional neural network, which will learn to recognise drought stress levels from images alone. To overcome the lack of data for such a method, I am implementing two key strategies:
• Data augmentation: I artificially increase the number of training images without altering their content by applying simple transformations to the pictures, like horizontal and vertical flips, while keeping their label. This expands the training dataset by the number of transformations applied and allows the model to generalise better.
• Transfer learning: I am using a pre-trained model instead of training one from scratch. This means I am only training the last layers of a model already trained to classify images in a different domain. This technique significantly decreases both training time and the amount of data required.
Finally, I will ensure that the model makes decisions based on meaningful features rather than irrelevant visual cues by highlighting which parts of an image contribute most to the algorithm’s prediction in a visualisation technique called saliency maps.
A Deeper, Rigorous Study of Drought
This project has been the occasion to create an original, valuable dataset that bridges a critical gap in drought research. Most studies focus either on eco-physiological markers or on imaging-based morphological traits, rarely integrating both in a single dataset. Additionally, many assume a uniform drying rate for all plants—an unrealistic simplification. Instead, I measured environmental variability —soil moisture—and plant responses over time, grounding the dataset in real-world conditions. As such, the dataset could yield insights beyond the original project scope and will be an asset for the community once published.
This project has already expanded our understanding of drought stress. The original clustering approach helped establish unbiased stress levels with associated Eco physiological marker levels. These findings could inform future studies and help compare results from different drought types and intensities.
A Project Turned Towards the Future
This project aims to develop a user-friendly online platform where users can simply upload photos of their crops and receive an AI-powered drought stress assessment, backed by scientific markers.
This tool could help farmers optimise irrigation by providing insights into how their crops respond to drought under their specific environmental conditions. This would support agroecological practices, for more resilient and sustainable farming systems in the face of climate change. Crucially, this tool is designed to be accessible and affordable for smallholder farmers in low-and middle-income countries (LMICs), who withstand the worst of climate change.
Beyond farming, the platform includes scientific markers of drought which would benefit agriculture scientists with decision-making. This could open new possibilities for data-driven research and enable more informed, evidence-based agricultural practices and more resistant crop varieties worldwide. Beyond this project, the clustering methodology can be applied in most situations with a heterogeneous environment. For example, I am also working on a project on mango tree resilience to environmental stress. Controlling the environment is impossible in field conditions but I have access to a wide variety of environmental conditions, providing a natural experimental setup. Once again, I will use eco-physiological markers to cluster plants, which will allow me to categorise different tree responses beyond drought and link them to local climate conditions. This new project will help understand the environmental factors affecting mango production for climate change previsions.
* * *
Dr Alice Malivert has also been recently featured in the Schmidt Science’s AI in Science video: