Early-Season forecasting of citrus block-yield using time series remote sensing and machine learning: A case study in Australian orchards

Abstract

This study presents a comprehensive evaluation of seasonal, locational, and varietal variations in canopy reflectance responses in 315 commercial citrus blocks from three major growing regions in Australia. The dataset includes three different citrus types (Mandarin, Navel, Valencia) and 26 varieties. The aim is to utilize this combined information to better understand yield variation and develop improved forecasting models. Landsat satellite data spanning from October 2006 to February 2021 (1419 tiles) were used to derive reflectance values, and calculate four vegetation indices (NDVI, GNDVI, LSWI, and GCVI), for each citrus block. These indices were then analyzed alongside corresponding yield data, which consisted of 3660 individual yield records dating back to 2007. Two temporal resolutions were incorporated as predictors: spatio-temporal vegetation index time series (TS) aggregated every two months and annual time series of historical block-yield records. Six statistical and machine learning algorithms were calibrated using a leave-one-year-out cross-validation approach (LOYO CV) and validated for one-year forward prediction over a five-year period (2017–2021). The results highlight significant yield variations across years, alternate bearing patterns, and spatio-temporal changes in reflectance profiles influenced by seasonal conditions, varietal characteristics, and locations. The support vector machine (SVM) algorithm with a radial basis function kernel consistently outperformed other algorithms, indicating a non-linear relationship between citrus yield and predictors. The SVM model achieved an RMSE of 15.5 T ha−1, R2 of 0.88, MAE of 12.1 T ha−1, and MAPE of 29% in predicting block-yield across farms, varieties, and seasons. These prediction accuracy metrics demonstrate an improvement over current forecasting methods. Notably, the proposed approach utilizes freely available imagery, provides forecasts between two to nine months before harvest, and eliminates the need for infield counting of fruit load for image calibration. This approach provides an improved method for understanding seasonal yield variation and quantifying citrus block-yield, offering valuable insights for growers in harvest logistics, labor allocation, and resource management.

Keywords

• Citrus
• Yield forecasting
• Time series
• Remote sensing
• Statistical models
• Machine learning