Comparative Evaluation of Convolutional and Transformer-Based Detectors for Automated Weed Detection in Precision Agriculture

Technical Summary: Comparative Evaluation of Convolutional and Transformer-Based Detectors for Automated Weed Detection

Problem Statement
The paper addresses the critical challenge of early-stage weed detection in precision agriculture, a task complicated by the high morphological similarity between crops and weeds, significant intra-class variability, and unstructured environmental conditions (e.g., variable lighting, soil textures, and occlusions). While deep learning has advanced agricultural computer vision, existing solutions often struggle with robustness in real-world scenarios. Specifically, there is a lack of systematic, fair comparisons between Convolutional Neural Network (CNN) architectures and Transformer-based models within a unified experimental framework using realistic, high-resolution datasets. The study aims to determine whether the global contextual modeling capabilities of Transformers offer a tangible advantage over the computational efficiency of CNNs for fine-grained crop-weed discrimination.

Methodology
The authors conducted a comprehensive comparative evaluation using the GROUND-BASED_WEED dataset, a public benchmark collected from commercial tomato fields in Spain. The dataset consists of high-resolution (5184 × 3888 pixels) RGB images captured from a bird's-eye view, featuring seven annotated classes (one crop and six weed species, plus an ambiguous class) and reflecting real-world challenges such as class imbalance and label noise.

The experimental design compared three representative object detection architectures:

1. CNN-based: YOLOv26-nano, selected for its lightweight design and suitability for embedded systems.
2. Transformer-based: RT-DETR (Medium and Large variants) and RF-DETR, chosen for their ability to capture global context via self-attention mechanisms.

All models were trained under a standardized protocol (200 epochs, AdamW optimizer, batch size 4) on an NVIDIA RTX 2080 Super GPU. To analyze the trade-off between accuracy and efficiency, experiments were conducted at three input resolutions: 640, 1080, and 2160 pixels. Performance was evaluated using standard COCO metrics (Precision, Recall, mAP@50, mAP@50–95) and inference efficiency metrics (latency and FPS).

Key Results
The evaluation revealed a clear trade-off between computational efficiency and contextual modeling:

• Performance: YOLOv26-nano and RT-DETR achieved comparable performance in terms of Precision, Recall, and mAP@50, particularly at higher resolutions. While RT-DETR showed marginal improvements in mAP@50 at 2160 pixels, these gains were not statistically significant compared to the CNN-based approach. RF-DETR consistently underperformed across all metrics.
• Resolution Impact: Increasing input resolution from 640 to 2160 pixels yielded only negligible improvements in detection accuracy. The authors attribute this to the intrinsic nature of the dataset, where target objects (early-stage weeds) occupy a very small fraction of the image, meaning higher resolution does not necessarily provide more discriminative features.
• Efficiency: YOLOv26-nano demonstrated significantly lower computational costs and faster inference times compared to the Transformer-based models. The Transformer models required higher memory allocation and exhibited longer inference times, posing challenges for real-time deployment on resource-constrained embedded platforms.

Key Contributions

• Systematic Comparison: The study provides a rigorous, side-by-side evaluation of state-of-the-art CNN and Transformer detectors under identical conditions, addressing a gap in the literature regarding fair comparisons in agricultural contexts.
• Real-World Benchmarking: Unlike many studies relying on controlled environments, this work utilizes a high-resolution, real-field dataset with inherent noise and class imbalance, offering a more realistic assessment of model robustness.
• Resolution Analysis: The paper investigates the diminishing returns of increasing input resolution for small-object detection in agriculture, suggesting that architectural efficiency may be more critical than raw resolution for this specific task.

Significance and Claims
The paper concludes that for early-stage weed detection in precision agriculture, hierarchical CNN optimization (specifically YOLOv26-nano) offers a more robust and viable approach than Transformer-based architectures. While Transformers excel at global context modeling, this capability does not translate into a substantial performance advantage for the specific task of detecting small, dispersed objects in agricultural fields.

The authors claim that YOLOv26-nano provides the most favorable balance between accuracy and computational efficiency, making it the preferred candidate for integration into embedded robotic systems operating under real-world field conditions. The study emphasizes that for this application, the efficiency of single-stage CNNs outweighs the theoretical benefits of global attention mechanisms, particularly when deployment on edge devices is a priority. Future work suggested by the authors includes exploring instance segmentation for better delineation of irregular weed morphologies and using synthetic data augmentation (e.g., GANs or diffusion models) to address dataset imbalance.