Mapping almond stem water potential using machine learning and multispectral imagery

• Published in: Irrigation science Vol. 43; no. 1; pp. 105 - 120
• Main Authors: Savchik, Peter, Nocco, Mallika, Kisekka, Isaya
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.01.2025; Springer Nature B.V
• Subjects: Accuracy; Agriculture; algorithms; almonds; Aquatic Pollution; Artificial neural networks; Biomedical and Life Sciences; California; canopy; Climate Change; Crops; data collection; Decompression chambers; Drought; Environment; Environmental regulations; Evapotranspiration; Groundwater; Irrigation; Irrigation efficiency; Learning algorithms; Life Sciences; Machine learning; Moisture content; multispectral imagery; Neural networks; Orchards; Original Paper; prediction; Pressure chambers; Reflectance; Root-mean-square errors; Soil moisture; soil water; Spatial variations; Spectral reflectance; Stems; Sustainable Development; Waste Water Technology; Water; Water Industry/Water Technologies; Water Management; Water Pollution Control; Water potential; Water use; Water use efficiency; Workflow; California
• ISSN: 0342-7188; 1432-1319; 1432-1319
• DOI: 10.1007/s00271-024-00932-8

Almonds are a major crop in California which produces 80% of all the world’s almonds. Widespread drought and strict groundwater regulations pose significant challenges to growers. Irrigation regimes based on observed crop water status can help to optimize water use efficiency, but consistent and accurate measurement of water status can prove challenging. In almonds, crop water status is best represented by midday stem water potential measured using a pressure chamber, which despite its accuracy is impractical for growers to measure on a regular basis. This study aimed to use machine learning (ML) models to predict stem water potential in an almond orchard based on canopy spectral reflectance, soil moisture, and daily evapotranspiration. Both artificial neural network and random forest models were trained and used to produce high-resolution spatial maps of stem water potential covering the entire orchard. Also, for each ML model type, one model was trained to predict raw stem water potential values, while another was trained to predict baseline-adjusted values. Together, all models resulted in an average coefficient of correlation of R 2  = 0.73 and an average root mean squared error (RMSE) of 2.5 bars. Prediction accuracy decreased significantly when models were expanded to spatial maps (R 2  = 0.33, RMSE = 3.31 [avg]). These results indicate that both artificial neural networks and random forest frameworks can be used to predict stem water potential, but both approaches were unable to fully account for the spatial variability observed throughout the orchard. Overall, the most accurate maps were produced by the random forest model (raw stem water potential R 2  = 0.47, RMSE = 2.71). The ability to predict stem water potential spatially can aid in the implementation of variable rate irrigation. Future studies should attempt to train similar models with larger datasets and develop a simpler faster workflow for producing stress predictions from field measurements.