Machine learning based plot level rice lodging assessment using multi-spectral UAV remote sensing

Rice plant lodging leads to change in canopy structure, yield loss and creates a menace in harvest operations. In situ assessment of lodging is time consuming, labour intensive, inefficient and inaccurate. Its assessment contributes greatly in in-situ field management and damage analysis. In this study, imaging observations from ten-band (within 444–842 nm) multispectral camera at 0.06 m Ground Sampling Distance (GSD) on-board an unmanned aerial vehicle (UAV) were acquired over a rice research farm (22.7930 N and 72.57140 E), Anand, Gujarat in western part of India. A set of features such as spectral reflectance, vegetation indices, colour coordinates and index, textural parameters and combination of all these were used for discriminating lodged rice crop from standing ones. All these features were extracted and analysed to optimize the sensitive features followed by discrimination of these two classes of rice using ensemble learning based Random Forest (RF) classifier. The analysis revealed that Green, Red-edge and Near-infrared (NIR) bands showed most optimal spectral features for lodging detection. The mean texture of these bands was also found to be sensitive indicators for rice lodging. Combined features with RF classifier produced an overall accuracy of 96.1% with kappa coefficient (κ) of 0.92 followed by textural features with an overall accuracy of 93.5 % and κ of 0.86. Plot level lodging assessment revealed that lodged area varied from 0.1 % to 15.5 % of the cropped area over different plots. The results were validated with the visually interpreted lodged areas using RGB image that resulted into R2 of 0.97 with relative root mean square error (rRMSE) of 0.02 %. Our results conclude that multispectral UAV based remote sensing can help in rapid damage assessment and plot-level field management with high precision.