Pre-harvest prediction of apple soluble solids content using UAV based multispectral imagery

초록

This study developed and validated machine learning models for predicting the pre-harvest soluble solids content (SSC) of apples using only vegetation indices derived from UAV-based multispectral imagery. A total of 39 'Fuji' apple trees across 10 orchards in Andong, Gyeongsangbuk-do, Republic of Korea, were surveyed over two growing seasons (2024– 2025). Imagery was acquired at noon on 17 September 2024 and 16 September 2025 using a Trinity F90+ fixed-wing UAV equipped with an Altum-PT multispectral sensor, flown at an altitude of 150 m with 75% overlap, yielding a ground sample distance (GSD) of 3.17 cm/pixel. Following radiometric calibration, mosaicking, and pansharpening, vegetation and background were separated using NDVI, and ARI, GLI, and NDVI rededge were extracted at the individual tree level. Using SSC measured with a handheld refractometer as the dependent variable, multiple linear regression (MLR) and k-nearest neighbor regression (KNR) models were constructed and evaluated by leave-one-out cross-validation (LOOCV) in terms of R2, RMSE, and MAPE. At the validation stage, MLR yielded an R2 of 0.55, a MAPE of 9.21%, and an RMSE of 1.34 °Brix. These results confirm that individual-tree SSC can be estimated non-destructively and without contact from canopy-level spectral information alone, and provide evidence supporting the potential use of this approach for orchard-scale maturity assessment and harvest timing decisions in precision agriculture.

키워드