fruit flies detection

Project Spotlight: AI-Powered Biosecurity via AutoTraps

Project Objective

To deploy a network of fully automated digital traps (AutoTraps) equipped with computer vision and machine learning algorithms to detect, identify, and count destructive fruit flies—specifically targeting high-risk species such as Anastrepha—delivering real-time population diagnostics directly to growers.

Project Background & Motivation

Fruit flies belonging to the Tephritidae family, including the aggressive Anastrepha, Bactrocera and Ceratitis genus, represent one of the single greatest biosecurity threats to global fruit production. Attacking high-value export crops like mangoes, melons, citrus, and stone fruits, these pests cause billions of dollars in direct agricultural losses and enforce strict international trade quarantine barriers on growers.

Monitoring these insect populations traditionally relies on a highly manual, flawed workflow:

The Labor Bottleneck: Farmers place hundreds of physical plastic traps across large acreage. Field technicians must manually walk the orchards week after week to inspect traps, count dead insects, and log paper data. This process is incredibly slow and expensive.
The Delayed Response Gap: Because human inspection only happens every few days or weeks, a minor pest infestation can balloon into an uncontrollable population explosion before anyone notices.
The Species Identification Problem: Telling different fruit fly species apart (or distinguishing them from harmless insects) requires trained entomological expertise. Mistaken identities lead to delayed or incorrect chemical treatments.

Our Solution: Real-Time Edge Analytics with AutoTraps

Our center bridges the gap between field agronomy and artificial intelligence by digitizing orchard biosecurity. The AutoTrap system completely automates pest management using a multi-layered computer vision pipeline:

Autonomous Capture & Imaging: Attracted by targeted food baits or synthetic pheromones, incoming fruit flies enter the low-power smart trap enclosure, where a miniaturized, high-resolution internal camera automatically takes their picture.
Artificial Neural Network (ANN) Identification: Each trap is equipped with an integrated Raspberry Pi 4 that securely stores the captured images and transmits them to a centralized server. On the cloud server, an advanced machine learning classification model is deployed. The AI analyzes highly precise morphological markers—such as the unique, intricate wing patterns of the fruit fly—to immediately identify the exact pest species
Instant GIS Mapping: The trap automatically tallies the population and sends a daily, real-time data update containing exact GPS coordinates and infestation counts straight to a web dashboard or farmer mobile application.

By transforming biosecurity into a continuous digital feed, our system allows agricultural teams to spot the very first fly of a major outbreak. Growers can apply localized, target-specific pesticide treatments only where and when they are truly needed, drastically lowering chemical usage, protecting the environment, and saving massive amounts of operational capital.

Project Collaborators

Agricultural Research Organization, Volcani Institute

Department of Entomology and Nematology

UF/IFAS Tropical Research & Education Center

University of Florida

Cervantes Agritech

Project Spatial Scope & Field Deployments

The automated traps (AutoTraps) utilized in these treatments are currently deployed across key global agricultural regions, including Israel, South Africa, and Florida (USA).

Israel Deployments: The captured datasets focus primarily on two distinct genera of destructive fruit flies: Bactrocera and Ceratitis. Within these, the most ecologically dominant species are the Peach Fruit Fly (Bactrocera zonata) and the Mediterranean Fruit Fly (Ceratitis capitata).
South Africa Deployments: Field monitoring heavily targets the highly invasive Oriental Fruit Fly (Bactrocera dorsalis), which represents the dominant species in this region.
Florida Deployments: Computer vision tracking specifically monitors native and invasive species within the Anastrepha genus.

Expert Domain Supervision & Annotations

To train our neural networks with absolute precision, the ground-truth image annotations were supervised by leading entomological authorities:

Dr. David Nestel, Senior Scientist Emeritus from the Department of Entomology within the Institute of Plant Protection at the Volcani Institute (ARO), Israel.
Prof. Daniel Carrillo, Associate Professor in the Department of Entomology and Nematology at the University of Florida, USA.

The Machine Learning Pipeline & Architecture

Our core computer vision framework relies on a state-of-the-art Deep Neural Network (DNN) utilizing the Faster R-CNN architecture built upon a robust ResNet-50-FPN backbone for simultaneous object detection and classification.

We fine-tuned and configured this network using a proprietary architectural approach. This specialized training enables the AI to detect and classify target flies with exceptionally high performance metrics (including localized Average Recall and Average Precision). The model maintains its high accuracy even under extreme field conditions, such as high-density imagery (traps completely saturated with flies) or visual environments cluttered with non-target mimic insects. Furthermore, a unique variant of our model successfully distinguishes the gender of specific species with elite structural accuracy.

Page updated

Google Sites

Report abuse