UAV Flight Height Impacts on Wheat Biomass Estimation via Machine and Deep Learning

Abstract

Optical unmanned aerial vehicle (UAV) remote sensing is widely prevalent to estimate crop aboveground biomass (AGB). Nevertheless, limited knowledge of the UAV flight height (mainly characterized by different image numbers and spatial resolutions) influences the crop AGB estimation accuracy across diverse sensing datasets and machine-/deep-learning models. This article assessed the impacts of flight height and integration of multiscale sensing information on wheat AGB estimation. The multispectral UAV flight missions with 30, 60, 90, and 120 m heights were conducted at the wheat grain filling phase in 2018 and 2019. To estimate AGB, we used the UAV-based crop surface model (CSM), spectral, texture indices, and their combinations along with a deep convolutional neural network (DCNN with AlexNet architecture), random forest, and support vector machine models. Results showed the CSM and textures exhibit sensitivity to flight height, with estimation accuracy declining by 48% and 41%, respectively, as the flight height increased from 30 to 120 m. Spectral indices displayed lesser sensitivity with accuracy decrease of 25%. Integrating data from different heights exhibited better performances in texture and spectral indices while reducing performance when CSM was input. The DCNN performed best particularly at high spatial image scales, whereas more sensitive to flight height, as the AGB estimation accuracy decreased by 30% and 47% from 30 to 120 m for machine learning and DCNN, respectively. Integrating texture and spectral information derived from images with moderate spatial resolutions (4–6 cm), and the integration of multiscale textures, are optimal for grain-filling wheat AGB estimation.

Keywords

• Agriculture
• image resolution
• image texture analysis
• remote sensing
• spectral analysis