Ai-driven weed detection system for selective herbicide spraying in carrot crops

Overview

This research-driven solution addresses the challenge of automated weed detection and selective herbicide application in carrot plantations. By combining deep learning and computer vision methods with dedicated computing hardware and acceleration, the technology overcomes computational cost barriers that traditionally limit AI adoption in agriculture. The system enables precise identification of weeds versus crops under challenging field conditions including high pattern heterogeneity, noise, rotation, and translation variations. The ultimate goal is to support selective herbicide fumigation that improves crop yields while minimizing chemical usage, using accessible, low-cost equipment suitable for field deployment.

Technical specifications

• Sparse-based feature selection approach that reduces redundant and noisy features to improve crop/weed discrimination accuracy
• Convolutional neural network (CNN) extraction methods for more discriminative and robust features in image-based phenotyping tasks
• HOG descriptor-based morphological plant descriptors feeding a classifier stage for performance assessment
• Incremental learning framework combined with ten-fold cross-validation to ensure affordable computational burden
• U-net type CNN architecture planned for fast and accurate image segmentation across varying crop densities
• Transfer learning strategy to address limited training data by leveraging existing labeled datasets from related crop/weed discrimination tasks
• IoT device implementation with moderate cost hardware to support field-ready selective spraying operations

Technology readiness level

The technology is currently at a validation stage with proof-of-concept demonstrated on the Crop/Weed Field Image Database, achieving competitive accuracy on 60 top-down field images of organic carrots in BBCH 10-20 growth stage. The dataset includes 162 crop plants and 332 weed plants with intra-row and close-to-crop weed appearances. Future validation efforts will focus on full CNN-based segmentation implementation, transfer learning integration, and complete deployment on IoT devices for selective herbicide spraying. The research team at Universidad Nacional de Colombia is actively seeking partnerships for project development support, field test equipment provision, and funding for personnel to advance the technology toward commercial readiness.