Science in Action #08: From Orchard Shadows to AI Insights — Manal Mhada’s Citrus Detection Breakthrough

In the vibrant landscape of global agriculture, citrus fruits represent a cornerstone crop—both economically and nutritionally. Yet, detecting citrus fruits during their early, green stage poses a persistent challenge for farmers and researchers alike. These young fruits, often no larger than a walnut and camouflaged within dense foliage, escape traditional human inspection and conventional computer vision tools.

Why does this matter? Because early detection is key to estimating yields, optimizing resources like water and fertilizers, and planning timely harvests. The sooner a farmer knows how much fruit is on the trees, the more precisely they can act. But in real-world orchards, where lighting conditions vary wildly and leaves obscure the view, the task becomes a needle-in-a-haystack problem. The unripe fruits blend in so seamlessly with their leafy background that even sophisticated models struggle.

Artificial Intelligence is reshaping agriculture, empowering farmers not just with tools, but with insights. In this project, AI is not just a buzzword—it’s the brain of the operation. Deep learning models, trained on thousands of annotated images, can recognize complex visual patterns far beyond the capacity of the naked eye. When deployed on citrus trees, AI doesn’t get tired, distracted, or misled by a shadow. It detects, counts, and segments fruit systematically and at scale. What sets this project apart is its dual approach:

Close-up imagery is used to train the AI, while full-tree images—more realistic but also more complex—are used to test it. This method mimics what a farmer might actually see. By combining advanced neural networks with smart slicing techniques, the researchers overcame many challenges posed by cluttered branches and variable lighting. With this system, AI becomes a field assistant—counting fruit clusters, predicting yield, and enabling precision agriculture. And all this with just a camera and a well-trained algorithm.

At the heart of this breakthrough is Manal Mhada, a researcher whose interdisciplinary vision bridges data science, agriculture, and real-world application. A member of the UM6P College of Agriculture and Environmental Sciences, Manal brings to the table not just technical rigor, but a commitment to solving pressing agricultural challenges in Morocco and beyond. Working alongside experts in plant phenotyping and machine learning, her focus is clear: leverage the latest AI technologies to serve farmers, not just papers.

Manal’s role in the project was pivotal—from coordinating field data collection to refining the deep learning pipeline and validating the model’s performance. She believes innovation must meet practicality. Our goal wasn’t to build the most sophisticated model in a lab. It was to build the most reliable solution in a real orchard. Through this work, Manal is not only elevating Moroccan agritech but also mentoring a new generation of researchers determined to make technology work for sustainability.

The project’s technical core is a robust framework that marries two AI powerhouses: Multiscale Vision Transformers (MViTv2) and Cascade Mask R-CNN. Think of MViTv2 as the eyes of the model, scanning the image for subtle differences in scale and color; and Cascade Mask R-CNN as the brain, refining object boundaries across stages until it reaches pixel-level accuracy. But the real innovation lies in the dual-image strategy. The system is trained on detailed close-up photos but tested on wide-angle images of full trees—where fruits are small, sparse, and often hidden.

To handle this, the researchers sliced each full-tree image into tiles, allowing the model to analyze localized sections before reassembling them into a complete fruit map. This technique not only improves detection in complex scenes but also keeps processing time efficient. Add transfer learning from massive image datasets, customized data augmentation, and optimized loss functions—and the result is a citrus-detecting AI that’s smarter, sharper, and more scalable than anything before it.

Where Does the Data Come From?

Any AI model is only as good as the data it learns from. In this case, the research team went to the field—literally. The images were captured in a commercial citrus orchard north of Marrakech, covering three distinct clementine varieties: Nules, Sidi Aissa, and Orogrande. These varieties were selected for their different flowering timelines and fruiting patterns, offering a rich spectrum of citrus phenotypes.

Two protocols guided the image collection: one with close-up shots taken just 50 to 80 cm from the fruit, and another with full-tree images shot from 3 meters away. This meticulous data collection was carried out under natural lighting, accounting for all the visual noise that complicates detection in the real world—shadows, glare, leaf occlusion, and uneven fruit sizes. Each image was manually annotated, including over 1400 segmented fruits. The result? A dataset that not only mirrors actual farming conditions but also trains the model to perform under them.

How Well Does It Work?

Numbers don’t lie—and in this case, they speak volumes. The best-performing model configuration, powered by the MViTv2_L backbone, achieved a mean Average Precision (mAP) of 84.40% for segmentation and 98.60% for bounding box detection on the test dataset. For real-world images of full trees, the slicing technique made all the difference: the model’s fruit count correlated up to R² = 0.81 with expert manual counts. That’s impressive, considering the challenges of visual occlusion and fruit clustering. Without slicing, performance dropped sharply—demonstrating the necessity of segmenting images for detailed analysis. The model also showed adaptability across different varieties and growth stages. Interestingly, the Nules variety, which flowered earliest, had the best detection rate. More slices improved accuracy but also increased the risk of double counting—so the team found the sweet spot at nine tiles per image. In sum, the model works not just in theory, but in the unpredictable, leafy chaos of Moroccan orchards.

What’s the Bigger Picture?

Beyond the pixels and performance metrics, this work has implications that ripple far beyond citrus. In a world grappling with food security, climate unpredictability, and water scarcity, precision agriculture offers a smarter way forward. The ability to count fruits early and accurately supports better yield forecasting, optimized irrigation, and data-driven crop management—all of which translate into economic and environmental gains.

For smallholder farmers, such tools could reduce losses and improve planning. For researchers, the framework opens doors to phenotyping other crops using the same dual-image logic. And for Morocco’s growing agritech ecosystem, it signals the rise of a new breed of tech-enabled solutions rooted in local realities but poised for global relevance.