The first Agricultural Revolution happened some 12,000 years ago, when humans settled down and started cultivating crops. We have since dramatically improved the art and science of agriculture, increasing scale and yield and shaping human civilization in the process. Has a new, AI-powered agricultural revolution now begun?
AI systems are already helping farmers with soil analysis, planting, animal husbandry, water conservation and more. Now a team of researchers from the University of Illinois at Urbana-Champaign (UIUC), Intelinair, and University of Oregon have introduced Agriculture-Vision, a large aerial image dataset for agricultural pattern analysis.
Accurate and real-time visual pattern recognition on farmland has enormous economic value. Aerial image semantic segmentation for detecting field conditions for example can help farmers avoid losses and increase yields throughout the growing season. Progress on visual pattern recognition in agriculture however has been slow, hindered in particular by a dearth of relevant datasets.
One of the challenges in developing aerial image datasets for agricultural pattern analysis is image size. Semantic segmentation for detecting field conditions over aerial farmland images requires “inference over extremely large-size images with extreme annotation sparsity,” the researchers explain in the paper Agriculture-Vision: A Large Aerial Image Database for Agricultural Pattern Analysis.
The researchers collected 94,986 aerial images from 3,432 farmlands across the US. Each image consists of RGB and Near-infrared (NIR) channels with resolution as high as 10 cm per pixel. Researchers identified nine of the most important types of field anomaly patterns for annotation. Five annotators trained by expert agronomists did the labeling, which was reviewed by the experts to ensure accuracy.
The research team captured the farmland images using specialized cameras on aerial vehicles flown over the fields during the 2107-19 growing seasons. The unprocessed farmland images had extremely large image sizes (up to 33571 x 24351 pixels) making end-to-end segmentation difficult due the high compute and memory requirements, and so researchers cropped the annotations with a window size of 512 x 512 pixels.
The researchers used a series of popular object semantic segmentation models to explore possible applications for the Agriculture-Vision dataset. They chose the DeepLab V3 and DeepLab V3+ for comparative evaluations and introduced a specialized FPN-based model. In the experiments, the FPN-based model outperformed the common object semantic segmentation models.
Although the extremely large aerial farmland images created unique challenges, the Agriculture-Vision dataset offers the research community an opportunity to explore the field with abundant data resources. The research team plans to expand the dataset with more images and additional modalities such as thermal images, soil maps, and topographic maps.
The other Agricultural images datasets referenced in the study, which may be of interest to readers:
- Sensefly Agriculture Dataset
- A crop/weed field image dataset for the evaluation of computer vision based precision agriculture tasks
- DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning
The paper Agriculture-Vision: A Large Aerial Image Database for Agricultural Pattern Analysis is on arXiv.
Journalist: Fangyu Cai | Editor: Michael Sarazen