Objective To design a hyperspectral detection system for nitrogen and phosphorus contents in citrus leaves.

Method Based on SR-GRU network training, an inversion model of nitrogen and phosphorus contents was constructed to obtain the spectral data of citrus leaves and corresponding nitrogen and phosphorus contents. A detection system of nitrogen and phosphorus contents in citrus leaves based on cloud edge-to-end architecture was designed. An improved iForest-SAM algorithm was proposed for outlier spectra test and rejection of spectral signals that were easily disturbed by outdoor light. The sparse LoRa message based on over-complete learning dictionary was proposed for fast transmission of spectral data with multiple bands, large size and slow transmission. The edge end of the system was acted as a Lora gateway in the citrus orchard, and at the mobile terminal end, the sparse Lora messages were sent to the cloud end via the edge end to load the inversion model for prediction.

Result The SR-GRU inversion model had the best inversion effect on the contents of nitrogen and phosphorus in citrus leaves, with determination coefficients of 0.929 and 0.865, respectively, and the normalized root mean square error of 0.083 and 0.079, respectively. The system took less than 1 s to detect the nitrogen and phosphorus contents of citrus leaves once, and the LoRa node was connected stably. The Web program based on the Internet ran stably, and the average page loading time was less than 0.5s.

Conclusion The system meets the practical application requirements for timely detection of nitrogen and phosphorus contents in citrus leaves.