Integrating Hyperspectral, Thermal, and Ground Data with Machine Learning Algorithms Enhances the Prediction of Grapevine Yield and Berry Composition

Accurately predicting grapevine yield and quality is critical for optimising vineyard management and ensuring economic viability. Numerous studies have reported the complexity in modelling grapevine yield and quality due to variability in the canopy structure, challenges in incorporating soil and microclimatic factors, and management practices throughout the growing season. The use of multimodal data and machine learning (ML) algorithms could overcome these challenges. Our study aimed to assess the potential of multimodal data (hyperspectral vegetation indices (VIs), thermal indices, and canopy state variables) and ML algorithms to predict grapevine yield components and berry composition parameters. The study was conducted during the 2019/20 and 2020/21 grapevine growing seasons in two South Australian vineyards. Hyperspectral and thermal data of the canopy were collected at several growth stages. Simultaneously, grapevine canopy state variables, including the fractional intercepted photosynthetically active radiation (fiPAR), stem water potential (Ψstem), leaf chlorophyll content (LCC), and leaf gas exchange, were collected. Yield components were recorded at harvest. Berry composition parameters, such as total soluble solids (TSSs), titratable acidity (TA), pH, and the maturation index (IMAD), were measured at harvest. A total of 24 hyperspectral VIs and 3 thermal indices were derived from the proximal hyperspectral and thermal data. These data, together with the canopy state variable data, were then used as inputs for the modelling. Both linear and non-linear regression models, such as ridge (RR), Bayesian ridge (BRR), random forest (RF), gradient boosting (GB), K-Nearest Neighbour (KNN), and decision trees (DTs), were employed to model grape yield components and berry composition parameters. The results indicated that the GB model consistently outperformed the other models. The GB model had the best performance for the total number of clusters per vine (R2 = 0.77; RMSE = 0.56), average cluster weight (R2 = 0.93; RMSE = 0.00), average berry weight (R2 = 0.95; RMSE = 0.00), cluster weight (R2 = 0.95; RMSE = 0.13), and average berries per bunch (R2 = 0.93; RMSE = 0.83). For the yield, the RF model performed the best (R2 = 0.97; RMSE = 0.55). The GB model performed the best for the TSSs (R2 = 0.83; RMSE = 0.34), pH (R2 = 0.93; RMSE = 0.02), and IMAD (R2 = 0.88; RMSE = 0.19). However, the RF model performed best for the TA (R2 = 0.83; RMSE = 0.33). Our results also revealed the top 10 predictor variables for grapevine yield components and quality parameters, namely, the canopy temperature depression, LCC, fiPAR, normalised difference infrared index, Ψstem, stomatal conductance (gs), net photosynthesis (Pn), modified triangular vegetation index, modified red-edge simple ratio, and ANTgitelson index. These predictors significantly influence the grapevine growth, berry quality, and yield. The identification of these predictors of the grapevine yield and fruit composition can assist growers in improving vineyard management decisions and ultimately increase profitability.