The emergence of data-driven methods, especially deep learning, as the state-of-the-art computer vision problems in several areas has created a necessity for annotated data for training such models. In the domain of agriculture, such techniques are being applied in applications such as phenotyping, yield prediction, disease detection, weed identification, and management, etc. Each of these specific use cases requires training data in formats that may range from a simple image-level label to complex pixel level or instance-level labels.
Creating realistic synthetic datasets can help apply artificial intelligence to agriculture because data annotation, especially at the pixel and instance level, is a costly and time-consuming process, thereby leading to quicker deployment. Transfer learning on such synthetic datasets may also be able to improve the algorithm performance, due to the possibility to model variations in environmental settings. Further, working with synthetic data allows for easier switching between sensors/platforms, for example between cameras, camera perspectives, RGB to/from multispectral, and even close range versus drones.
Most standard approaches to generating synthetic datasets in agriculture or plant science domains involve using manual inputs such as plant characteristics or scribbled sketches, followed by rendering tools to incorporate settings such as illumination and shading. More recently, generative models, the most well-known of which are Generative Adversarial Networks (GANs), provide a way to generate realistic data automatically. GANs are versatile in learning patterns of the target dataset and produce a similar image given a new image from the source dataset.
The increasing use of digital twins in agriculture is potentially a source of base datasets from which to generate desired synthetic data, as is already being done in other domains such as biomedicine and healthcare.
This Research Topic focuses on new work and results on the use of synthetic data in the agricultural and plant sciences. The topic welcomes, but is not limited to, original research and reviews related to the generation and use of synthetic datasets in precision agriculture, horticulture, and plant phenotyping areas:
-Crop/weed detection and identification
-Plant part segmentation
-Fruit/flower detection and counting
-Multi-modality data
-Plant growth models
-Generative models
The emergence of data-driven methods, especially deep learning, as the state-of-the-art computer vision problems in several areas has created a necessity for annotated data for training such models. In the domain of agriculture, such techniques are being applied in applications such as phenotyping, yield prediction, disease detection, weed identification, and management, etc. Each of these specific use cases requires training data in formats that may range from a simple image-level label to complex pixel level or instance-level labels.
Creating realistic synthetic datasets can help apply artificial intelligence to agriculture because data annotation, especially at the pixel and instance level, is a costly and time-consuming process, thereby leading to quicker deployment. Transfer learning on such synthetic datasets may also be able to improve the algorithm performance, due to the possibility to model variations in environmental settings. Further, working with synthetic data allows for easier switching between sensors/platforms, for example between cameras, camera perspectives, RGB to/from multispectral, and even close range versus drones.
Most standard approaches to generating synthetic datasets in agriculture or plant science domains involve using manual inputs such as plant characteristics or scribbled sketches, followed by rendering tools to incorporate settings such as illumination and shading. More recently, generative models, the most well-known of which are Generative Adversarial Networks (GANs), provide a way to generate realistic data automatically. GANs are versatile in learning patterns of the target dataset and produce a similar image given a new image from the source dataset.
The increasing use of digital twins in agriculture is potentially a source of base datasets from which to generate desired synthetic data, as is already being done in other domains such as biomedicine and healthcare.
This Research Topic focuses on new work and results on the use of synthetic data in the agricultural and plant sciences. The topic welcomes, but is not limited to, original research and reviews related to the generation and use of synthetic datasets in precision agriculture, horticulture, and plant phenotyping areas:
-Crop/weed detection and identification
-Plant part segmentation
-Fruit/flower detection and counting
-Multi-modality data
-Plant growth models
-Generative models