Ardan Wiratmoko a  , Andri Prima Nugroho a, Mutiara Alifia Ramadhanty a, Fahmi Arsyad a, Fadel Arya Pradana a, Bondan Satria Pamungkas a, Lilik Sutiarso a, Takashi Okayasu b

• ^a Smart Agriculture Research Center, Department of Agricultural and Biosystems Engineering, Faculty of Agricultural Technology, Universitas Gadjah Mada, Jl. Flora No. 1 Bulaksumur, Yogyakarta 55281, Indonesia

• b Department of Agro-Environmental Sciences, Faculty of Agriculture, Kyushu University, 744 Motooka Nishi-ku, Fukuoka 810-0395, Japan

Abstract

Spatial heterogeneity in crop physiological performance remains a major constraint to efficient management in tropical horticultural systems. This study evaluated the capability of UAV-based multispectral imagery to predict intrinsically derived physiological health states of tropical papaya by integrating canopy spectral predictors with multivariate plant functional measurements. Field observations were conducted on 103 papaya trees in a commercial plantation in Yogyakarta, Indonesia. Six physiological indicators, namely SPAD chlorophyll meter value, stomatal conductance (gsw), electron transport rate (etr), maximum fluorescence (F_m), steady-state fluorescence (F_s), and effective quantum yield of photosystem II (ΦPSII), were used to derive intrinsic health states through K-means clustering after z-score standardization. Candidate cluster numbers (K = 2–5) were evaluated using inertia, silhouette coefficient, Calinski–Harabasz index, and Davies–Bouldin index. Although K = 2 yielded the most compact statistical partition, K = 3 was retained to preserve agronomically interpretable Healthy, Moderate, and Stressed physiological states. A total of 124 UAV-derived spectral predictors were constructed at the plant level and sequentially reduced to 21 predictors through correlation pruning and recursive feature elimination with cross-validation (RFECV). Random Forest classification using the selected predictors achieved a training accuracy of 0.958, testing accuracy of 0.903, and 5-fold cross-validation accuracy of 0.96 ± 0.02. Model interpretability identified GBNDVI_max, GBNDVI_mean, rho_green_min, MGRVI_std, RTVI_max, and MGRVI_min as key spectral features, while partial dependence analysis showed stronger threshold-like relationships with stomatal conductance than with ΦPSII. These findings indicate that UAV multispectral sensing can capture physiologically meaningful within-field heterogeneity in tropical papaya and support management-priority monitoring in precision agriculture.

Keywords

UAV multispectral imagery

Physiological health states

Tropical papaya

Random forest classification

Precision agriculture

https://doi.org/10.1016/j.ages.2026.100023