Improved crop row detection with deep neural network for early-season maize stand count in UAV imagery
Abstract Stand counts is one of the most common ways farmers assess plant growth conditions and management practices throughout the season. The conventional method for early-season stand count is through manual inspection, which is time-consuming, laborious, and spatially limited in scope. In recent years, Unmanned Aerial Vehicles (UAV) based remote sensing has been widely used in agriculture to provide low-altitude, high spatial resolution imagery to assist decision making. In this project, we designed a system that uses geometric descriptor information with deep neural networks to determine early-season maize stands from relatively low spatial resolution (10 to 25 mm) aerial data, which covers a relatively large area (10 to 25 hectares). Instead of detecting individual crops in a row, we process the entire row at one time, which significantly reduces the requirements for the clarity of the crops. Besides, our new MaxArea Mask Scoring RCNN algorithm could segment crop-rows out in each patch image, regardless of the terrain conditions. The robustness of our scheme was tested on data collected at two different fields in different years. The accuracy of the estimated emergence rate reached up to 95.8%. Due to the high processing speed of the system, it has the potential for real-time applications in the future.
Paper
Full text
Improved crop row detection with deep neural network for early-season maize stand count in UAV imagery
Semantic Scholar · Agricultural and Food Sciences · 2020
Abstract
Abstract Stand counts is one of the most common ways farmers assess plant growth conditions and management practices throughout the season. The conventional method for early-season stand count is through manual inspection, which is time-consuming, laborious, and spatially limited in scope. In recent years, Unmanned Aerial Vehicles (UAV) based remote sensing has been widely used in agriculture to provide low-altitude, high spatial resolution imagery to assist decision making. In this project, we designed a system that uses geometric descriptor information with deep neural networks to determine early-season maize stands from relatively low spatial resolution (10 to 25 mm) aerial data, which covers a relatively large area (10 to 25 hectares). Instead of detecting individual crops in a row, we process the entire row at one time, which significantly reduces the requirements for the clarity of the crops. Besides, our new MaxArea Mask Scoring RCNN algorithm could segment crop-rows out in each patch image, regardless of the terrain conditions. The robustness of our scheme was tested on data collected at two different fields in different years. The accuracy of the estimated emergence rate reached up to 95.8%. Due to the high processing speed of the system, it has the potential for real-time applications in the future.