RiceNet Uses Machine Learning and Aerial Drones to Take the Tedium Out of Rice Production Estimation

Researchers from China and Singapore have turned technology onto the problem of quantifying rice production, aiming to do away with laborious and tedious hand-counting in favor of rice-spotting drones backed by a machine learning system.

"The new technique uses UAVs [Uncrewed Aerial Vehicles] to capture RGB images — images composed primarily with red, green, and blue light — of the paddy field," Jianguo Yao, professor at the Nanjing University of Posts and Telecommunications and study lead, explains. "These images are then processed using a deep learning network that we have developed, called RiceNet, which can accurately identify the density of rice plants in the field, as well as provide higher-level semantic features, such as crop location and size."

The idea behind the project is to provide an easier way to quantify production on the estimated 162 million hectares of land used globally for rice growing — and one which is less tedious than the traditional method of having people count the plants by hand. Images captured by commercial off-the-shelf DJI Phantom 4 Advanced drones, which were tested in the skies above Nanchang's rice fields, are analyzed using RiceNet to extract features and decode them to provide estimates of plant density, location, and size.

While technically only plant density is required to provide quantification of expected production rates, the latter two features were added in order to fuel future research projects proposed by the team — including a planned project to automate crop management tasks, including fertilizer spraying.

In real-world testing, the researchers proved that RiceNet is capable of generating density maps which accurately match data provided by hand-counting — and came up with a few recommendations for similar projects, including skipping the image collection on rainy days due to poor results and sending the drones aloft in the four-hour period after sunrise to avoid both obscuring fogs and the tendency for rice plants to curl their leaves later in the day.

"In addition to this, we further validated the performance of our technique using two other popular crop datasets," Yao adds, referring to experiments with the Wheat Ear Dataset (WED) and Maize Tassels Count (MTC) datasets which proved that RiceNet could be generalized to other crops. "The results showed that our method significantly outperforms other state-of-the-art techniques. This underscores the potential of RiceNet to replace the traditional method of manual rice counting."

The team's work has been published in the journal Plant Phenomics under open-access terms.