Classification of maize lodging types using UAV-SAR remote sensing data and machine learning methods

Lodging seriously threatens maize quality and yield and inevitably increases management and harvest costs. Timely collection of crop lodging information plays a pivotal role in the post-disaster assessment and agricultural insurance claims. Although spaceborne radar and optical remote sensing have unparalleled advantages in obtaining large-scale agricultural information, their response capacity to sudden natural maize lodging disasters is insufficient due to the limited spatial-temporal resolution of the satellite data. In recent years, the widespread application of unmanned aerial vehicles (UAVs) based optical remote sensing in precision agriculture has provided an effective alternative to spaceborne remote sensing. However, optical sensing can only effectively reveal the reflectance spectral characteristics of lodging maize under good lighting conditions. This work proposes a novel maize lodging classification method based on UAV synthetic aperture radar (UAV-SAR) and machine learning to circumvent the limitations of spaceborne and UAV-based remote sensing in monitoring maize lodging. Firstly, the raw radar remote sensing data of our study area containing lodging and non-lodging maize plants at the maturity stage is collected by the custom-built X-band and Ku-band UAV-SAR systems. Secondly, the corresponding backscattering coefficients and radar vegetation indices in each lodging type are extracted through radiation calibration and band math. Subsequently, the impacts of radar parameters (bands, polarizations, and observation orientations) and lodging types on backscattering coefficients are comprehensively analyzed. Fourthly, we applied the recursive feature elimination (RFE) algorithm to identify significant feature subsets and constructed multiple datasets using ten filter scales. Finally, five machine learning models (XGBoost, LDA, RF, KNN, and ANN) are trained and tested based on these materials. The classification results under different filter scales and feature combinations show that ANN achieves the best performance with an overall accuracy of 98.26 % and a Kappa coefficient of 0.982. This is the first innovative study successfully introducing cutting-edge UAV-SAR into maize lodging monitoring. Following spaceborne optical, spaceborne radar, and UAV-based optical remote sensing technologies, UAV-SAR holds great potential as the fourth practical means for collecting high-resolution agricultural information.