Descriptor: Dataset of Parasitoid Wasps and Associated Hymenoptera (DAPWH)

This paper introduces the Descriptor Dataset (DAPWH), a curated collection of 3,556 high-resolution images of parasitoid wasps and associated Hymenoptera, including 1,739 COCO-annotated images, to facilitate the development of automated identification systems for these taxonomically challenging groups.

The Gist

Imagine you are trying to teach a robot to recognize different types of bees and wasps. You might think, "Easy! Just show it pictures of bees." But here's the problem: the world of wasps is like a massive, crowded library where most of the books have no titles, and the few that do have titles are written in a language only a handful of experts can read.

This paper introduces a new, massive digital library called DAPWH (Dataset of Parasitoid Wasps and Associated Hymenoptera) designed to solve this problem. Here is a simple breakdown of what the researchers did and why it matters.

1. The Problem: The "Needle in a Haystack" of Nature

There are millions of insect species, but the ones that help farmers by eating pests (called parasitoid wasps) are incredibly hard to identify.

• The Challenge: They look almost identical to the naked eye. It's like trying to tell apart 50 different models of the same car just by looking at a blurry photo.
• The Bottleneck: To identify them, you usually need a human expert with a microscope and a lot of time. But there aren't enough experts, and they are getting older. We need a faster way.

2. The Solution: A "Training Gym" for AI

The researchers built a massive dataset of 3,556 high-resolution photos of these wasps. Think of this dataset as a gym where a computer program (Artificial Intelligence) goes to exercise and get strong.

• The Workout: The dataset includes photos of the main "suspects" (Ichneumonidae and Braconidae wasps) but also includes photos of other wasps and bees (like honeybees and yellow jackets) to teach the AI what not to confuse them with.
• The Views: Just like a car needs to be seen from the front, side, and back to be fully understood, these wasps were photographed from three angles: Lateral (side), Frontal (face), and Dorsal (top). This gives the AI a 3D understanding of the insect.

3. The Secret Sauce: "Highlighting" the Details

The most special part of this dataset is a subset of 1,739 images that have been "highlighted" by humans.

• The Highlighter: Using a tool called CVAT, experts drew precise boxes around the wasp's body, its wings, and even the tiny ruler (scale bar) next to it.
• Why it matters: This is like giving a student a textbook where the teacher has already underlined the most important words. Instead of the AI guessing where the wasp is, it learns exactly where the wings end and the body begins. This helps the AI learn to spot tiny details, like the veins in a wing, which are crucial for identification.

4. The Test Drive: Can the Robot Do It?

The researchers didn't just build the dataset; they tested it to see if it actually works. They taught several different AI models (like YOLO and EfficientNet) using this data.

• The Results: The AI models got really good at the job. Some achieved over 92% accuracy in telling the families of wasps apart.
• The Weakness: The AI sometimes got confused between two very similar-looking families (Andrenidae and Colletidae). Why? Because there were very few examples of those specific wasps in the "gym." It's like trying to learn to recognize a rare bird when you've only seen three photos of it.

5. Why Should You Care?

This isn't just about wasps; it's about biodiversity and food security.

• The Guardians: These wasps are nature's pest control. They eat the bugs that destroy our crops.
• The Future: By giving computers a good "textbook" (this dataset), we can build tools that automatically count and identify these wasps in the wild. This helps scientists monitor ecosystems and helps farmers know if their natural pest control is working, without needing to hire a team of experts to look through microscopes for days.

In a Nutshell

The authors took a chaotic, hard-to-understand group of insects, organized them into a beautiful, high-quality photo album with detailed labels, and used it to teach computers how to recognize them. It's a digital toolkit that turns a super-hard biology problem into a solvable math problem, helping us protect our planet's delicate balance.

Technical Summary

1. Problem Statement

• Taxonomic Challenge: The superfamily Ichneumonoidea (families Ichneumonidae and Braconidae) represents one of the most diverse and ecologically critical groups of insects, playing a vital role in regulating agricultural pests and maintaining biodiversity. However, they are notoriously difficult to identify due to cryptic morphology, a vast number of undescribed species, and a global shortage of expert taxonomists.
• Resource Scarcity: There is a critical lack of robust, high-quality digital resources (specifically annotated image datasets) required to train automated identification systems (Computer Vision/AI) for these groups. Traditional identification relies on slow, manual microscopic examination.

2. Methodology

The authors developed a comprehensive dataset and a validation pipeline to address these gaps.

• Data Collection:

  • Source: Specimens were retrieved from four major university-affiliated repositories in Brazil (UFSCar, USP, UNIP, and Spencer Collection), representing years of field research.
  • Sampling: Specimens were originally captured via Malaise traps and identified by experts using morphological keys.
  • Imaging: Images were captured using a Leica M205C stereomicroscope with a Leica K5C camera, controlled by LAS X software. Image stacking was performed using Helicon Focus to ensure high resolution and depth of field.
  • Scope: The dataset includes 3,556 high-resolution images covering 11 Hymenoptera families. The primary focus is on Ichneumonidae (786 images) and Braconidae (648 images), with supplementary families including Apidae, Vespidae, Pompilidae, and others to improve model robustness.
  • Views: Images are organized into three morphological perspectives: Dorsal/Ventral, Frontal, and Lateral.

• Annotation Process:

  • A subset of 1,739 images was annotated in the COCO (Common Objects in Context) format.
  • Tools: Annotations were generated using the Computer Vision Annotation Tool (CVAT) assisted by the Segment Anything Model (SAM).
  • Classes: Three distinct classes were annotated:
    1. Full insect body (segmentation masks).
    2. Wings (segmentation masks).
    3. Scale bars (bounding boxes).
  • Validation: Annotations were reviewed by computer vision specialists and finalized by a lead taxonomist to ensure accuracy.

• Experimental Setup:

  • Split: The dataset was stratified into Training (70%), Validation (15%), and Testing (15%) sets.
  • Hardware: Experiments were run on a workstation with an NVIDIA GeForce RTX 4090 GPU.
  • Models:
    • Classification: Tested EfficientNetV2, ConvNeXt, ViT-B/16, VGG16, ResNet50, and YOLOv12.
    • Object Detection: Benchmarked YOLOv8 and YOLOv12.

3. Key Contributions

• DAPWH Dataset: The creation of the first large-scale, high-resolution image dataset specifically for Neotropical Ichneumonoidea and associated Hymenoptera.
• Multi-Modal Annotation: Unlike standard datasets, this includes precise segmentation masks for wings (critical for taxonomic identification) and scale bars (enabling biometric measurements), alongside full-body segmentation.
• Open Access: The dataset is publicly available via Zenodo (DOI: 10.5281/zenodo.18501018) under a CC BY-NC 4.0 license, accompanied by a GitHub repository containing the full computational pipeline (preprocessing, training, and evaluation scripts).
• Benchmarking: Established baseline performance metrics for both image classification and object detection on this specific taxonomic group.

4. Results

• Image-Level Classification:
  • All tested architectures achieved high accuracy.
  • EfficientNetV2 and ConvNeXt achieved the highest Top-1 accuracy at 92.28%.
  • YOLOv12 demonstrated the highest robustness with an F1-score of 95.59%.
• Object Detection (COCO Subset):
  • YOLOv12 outperformed YOLOv8, achieving an mAP@50 of 90.53%, Precision of 90.23%, and Recall of 82.54%.
  • Feature Detection: The model showed exceptional accuracy in detecting wings (96%), which are crucial for species differentiation.
  • Challenges: Lower performance was observed for the Colletidae family due to severe class imbalance (only 25 annotations) and for scale bars (0.61% false negative rate), where high-resolution background noise occasionally interfered with localization.
• Confusion Analysis: The model successfully distinguished key families like Ichneumonidae, Apidae, Vespidae, and Bethylidae but struggled with morphologically similar families (e.g., Andrenidae vs. Colletidae) due to limited training data for the latter.

5. Significance

• Accelerating Taxonomy: This resource provides the foundational data necessary to develop AI tools that can assist taxonomists in identifying parasitoid wasps, potentially mitigating the "taxonomic impediment" caused by a lack of experts.
• Ecological & Agricultural Impact: By enabling automated monitoring of parasitoid populations, the dataset supports better biological control strategies for agricultural pests and aids in biodiversity monitoring.
• Future-Proofing: While current annotations are at the family level, the dataset preserves higher-resolution taxonomic data in filenames, allowing for future extension to genus or species-level classification as more annotations become available.
• Phenotypic Analysis: The inclusion of wing segmentation and scale bars facilitates the automated extraction of phenotypic features and biometric measurements, opening new avenues for evolutionary biology research.