Time-series UAV multispectral imaging for HLB detection via improved U-Net under dynamic orchard environment

Abstrac t Citrus Huanglongbing (HLB) is the most devastating disease threatening citrus growth, posing a significant challenge to citrus cultivation due to its difficult cure and high transmissibility. Therefore, rapid detection and then removal of HLB infected trees is essential for orchard management. However, the complex and dynamic orchard environment often blocks HLB detection. Using time series analysis methods for large-scale monitoring of citrus orchards is uncommon. This study proposed a two-stage method for HLB detection utilizing time-series multispectral unmanned aerial vehicle (UAV) remote sensing images. The citrus canopy segmentation was carried out firstly using a proposed Canopy-Net with MIoU of 95.4%, which minimized background interference due to dynamic orchard conditions caused by weed. To improve labeling efficiency of UAV-based HLB datasets, an auxiliary labeling strategy was proposed for 551 trees which only costed approximately 30 minutes by combining previous canopy segmentation outputs and plant polymerase chain reaction (PCR) results. Furthermore, RS-HLBNet was established for HLB detection in different seasons. It introduced a 2D-3D parallel convolution block in head to enhance spectral feature extraction. To resolve potential semantic mismatches in the original UNet, a CBDCA block was proposed to integrate multi-level skip connection features. The EUCB block further improved feature representation during upsampling. Finally, an MIoU of 88.5% was achieved during the asymptomatic period (1 st August) and 92.2% in the symptomatic period (22 nd November), respectively. In conclusion, this study demonstrated the capability of a time-series UAV multispectral imaging for the rapid HLB detection across orchard scale under dynamic orchard environment. • Analysis of UAV time-series spectral features for citrus Huanglongbing. • Canopy segmentation model reduces the effects of complex environment variations. • Auxiliary labeling strategy significantly reduces annotation time. • The Citrus Huanglongbing detection model is efficient for multi-period monitoring.