Sony/Hokkaido_Agriculture_Image_Dataset

Hokkaido Agriculture Image Dataset

Introduction

In the face of declining numbers and aging of agricultural workers, image AI undoubtedly represents a crucial technology for smart agriculture. For instance, it enables accurate monitoring of farmland and crop conditions through counting crops and detecting pests, while precise shape recognition allows robots to reduce heavy labor, with many such applications anticipated. Image AI algorithms and models continue to advance rapidly, not only achieving more accurate recognition capabilities but also progressing in research and development of systems that can handle language, such as answering questions about images.

Meanwhile, datasets, along with models, are essential for image AI development. No matter how capable a model is, without datasets tailored to the environment where the AI will be used, achieving sufficient performance becomes difficult, making datasets extremely important.

However, in agricultural AI, even data collection must be described as high-cost. This is because it requires gathering comprehensive data on crops and surrounding environments that change daily, sometimes necessitating multiple visits to vast agricultural lands.

To help lower these barriers in agricultural image AI development, Hokkaido University and Sony Group Corporation have jointly constructed an agricultural image dataset, Hokkaido Agriculture Image Dataset. This dataset consists of approximately 2,000 images and 5 crops (grapes, wheat, onions, apples, and haskap) collected from various agricultural lands in Hokkaido, tailored for DIVERSE and REAL agricultural AI application scenarios including crop counting, heading detection and harvesting. Also, The photographs were taken from various growth stages, lighting conditions, and viewpoints to create a dataset that supports image recognition and machine learning research in actual agricultural environments.

What kind of data is included ?

Grape instance segmentation

In wine grape cultivation, harvesting is extremely labor-intensive tasks. To reduce this workload, research and development of robots for harvesting and pruning is being actively conducted. In this context, precise recognition and instance segmentation to identify pruning and harvesting points are crucial. This dataset includes data for these applications. For harvesting, the classes are grapes (fruit), branches, cut points (parts of branches connected to fruit), and wires.

Object detections

While not as precised as instance segmentation, object detection is also very useful for accurately determining crop quantities. This dataset includes object detection data for grapes (fruit), wheat, onions, apples, and haskap. Particularly for grapes, wheat, apples, and haskap berries, images were captured over extended periods, resulting in data covering various weather conditions and quantities. Note that for grapes, apples, and haskap, annotation in this dataset is focused to the foremost trees or plants, excluding fruits visible in the background. This is because counting fruits by individual tree or plant is better for management purposes. However, these background elements may serve as negative samples in AI training. If concerned about their impact, please consider removing parts of images or cropping them.

Here is the dataset specifications and sample images. The annotation format of all images is COCO format.

Crop	Number of Images	classes	resolution	Annotation Format
Grape	1061	grape, wire, branch, cutpoint	4896 x 3672, 1024 x 768, 640 x 480	Object Detection, Instance Segmentation
Wheat	255	wheat	4056 x 3040, 4000 x 3000, 640 x 480	Object Detection
Onion	336	onion	600 x 600	Object Detection
Apple	171	apple	4896 x 3672	Object Detection
Haskap	51	haskap, flower	4896 x 3672	Object Detection

Dataset Structure

The dataset is organized as follows:

dataset/
├── README
├── grape_instance_segmentation/          
│   ├── images/     
│   └── annotation/annotation_coco.json  
├── grape_object_detection/          
│   ├── images/     
│   └── annotation/annotation_coco.json   
├── wheat/          
│   ├── images/     
│   └── annotation/annotation_coco.json  
├── onion/          
│   ├── images/     
│   └── annotation/annotation_coco.json  
├── apple/          
│   ├── images/     
│   └── annotation/annotation_coco.json  
└── haskap/         
    ├── images/     
    └── annotation/annotation_coco.json  

Licence

The datasetis relreased under CC-BY-4.0

Acknowledgements

This dataset is the result of joint research between Hokkaido University and Sony Group Corporation, Social Innovation Division for Planetary Boundary. We appreciate everyone for their support and effort in this joint research. Especially, We would like to express our deep gratitude to Vehicle Robotics Laboratory at Hokkaido University for their comprehensive cooperation in dataset Construction.

We also extend our sincere thanks to all those who agreed to the collection and publication of crop images.

• HOKKAIDO WINE CO.,LTD.
• MIURA FARM CO.,LTD.
• Kitami Institute of Technology, Laboratory of Bio-Mechatronics
• The Field Science Center for Northern Biosphere, Hokkaido University