Spatio-Temporal Data Fusion for Early Detection of Phytophthora cinnamomi in Macadamia Orchards using UAV-LiDAR and Multispectral Imagery

Abstract Background Phytophthora cinnamomi (Pc) causes macadamia root rot responsible for 25–40% annual yield losses and over USD 500 million per year in industry-wide costs. The principal challenge for management is a 60–120-day pre-symptomatic latent phase during which subterranean root colonisation proceeds undetected, rendering conventional visual scouting wholly ineffective within the critical intervention window. Methods A longitudinal 18-month field experiment was conducted at three contrasting production sites in Zimbabwe using a DJI Matrice 300 RTK platform equipped with a Zenmuse L2 LiDAR and Zenmuse H30T five-band multispectral camera. Six spectral vegetation indices and eight structural LiDAR metrics were extracted per tree crown for 120 individually labelled trees and fused into 14-dimensional monthly spatio-temporal tensors. Disease ground truth was established via P. cinnamomi-specific TaqMan qPCR assays. A CNN image classifier trained on 12,300 labelled canopy images was developed, and a novel CNN-LSTM hybrid architecture for spatio-temporal fusion was implemented. Results The CNN image classifier achieved 96.16% accuracy (weighted F1 = 0.9618). The CNN-LSTM hybrid architecture achieved F1 = 0.96 and AUC-ROC = 0.99, with a mean pre-symptom detection lead-time of 41 ± 6 days across 84 confirmed infection events, exceeding the critical 30-day phosphonate intervention threshold in every case. Economic modelling yields a benefit-to-cost ratio of approximately 10:1 for orchards with greater than 15% annual infection incidence. Conclusion The proposed spatio-temporal data fusion framework demonstrates the first empirically validated approach for early detection of P. cinnamomi in macadamia orchards, establishing a viable operational deployment pathway with significant implications for yield protection and smallholder livelihoods across Sub-Saharan Africa.