Rahman et al. (2025) Machine Learning Approaches for Assessing Avocado Alternate Bearing Using Sentinel-2 and Climate Variables—A Case Study in Limpopo, South Africa

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-12-05
• Authors: Muhammad Moshiur Rahman, Andrew Robson, Theo Bekker
• DOI: 10.3390/rs17243935

Research Groups

• Applied Agricultural Remote Sensing Centre (AARSC), University of New England, Armidale, NSW 2350, Australia
• Westfalia Fruit Estates (Pty) Ltd., R36, Modjadjiskloof 0835, South Africa

Short Summary

This study aimed to assess and predict avocado alternate bearing patterns in commercial orchards using satellite remote sensing and climatic variables. The TabPFN machine learning model effectively predicted alternate bearing with high accuracy, demonstrating that a combination of Sentinel-2 vegetation/flowering indices and key climatic factors during the flowering period can support proactive orchard management.

Objective

• To assess the potential of satellite remote sensing and climatic variables to characterize and predict alternate bearing patterns in commercial avocado orchards in Tzaneen, Limpopo Province, South Africa.
• To develop and validate a remote sensing-based framework that integrates Sentinel-2 vegetation and flowering indices with climate variables and multiple machine learning algorithms to detect alternate bearing in avocado orchards.

Study Configuration

• Spatial Scale: 46 'Hass' avocado orchard blocks within Belvedere Farm, Tzaneen, Limpopo Province, South Africa. Satellite imagery at 10–20 meter spatial resolution.
• Temporal Scale: Historical yield data from 2018 to 2024. Sentinel-2 imagery from January 2016 to December 2024. Monthly climate data from 2016 to 2024 (TerraClimate dataset, originally from 1958 onward). Vegetation and flowering indices aggregated into eight quarterly averages from the two years preceding each yield year. Climate variables focused on June–October, particularly July–September (flowering period).

Methodology and Data

• Models used: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LightGBM), Tabular Prior-Data Fitted Network (TabPFN).
• Data sources:
  • Satellite: Harmonized Sentinel-2 Level-2A surface reflectance imagery (Google Earth Engine).
  • Observation/Field: Historical avocado block yield data (tonnes per hectare) from 2018 to 2024 for 46 orchard blocks, detailed farm maps, and general avocado phenological information (all from Westfalia Fruit Estates (Pty) Ltd.).
  • Reanalysis/Climate: TerraClimate dataset (Google Earth Engine) for mean monthly maximum temperature (Tmax, degrees Celsius), mean monthly minimum temperature (Tmin, degrees Celsius), mean monthly vapor pressure deficit (VPD, kilopascals), and mean monthly precipitation (millimeters).
  • Derived: Sentinel-2 derived Vegetation Indices (NDVI, GNDVI, NDRE, CIG, CIRE, EVI2, LSWI) and Flowering Indices (WYI, NDYI, MTYI). Savitzky–Golay filter applied for smoothing time series.

Main Results

• Alternate bearing in avocado can be effectively assessed and predicted using a combination of Sentinel-2 vegetation indices, flowering indices, and key climatic variables (VPD, Tmin, Tmax, precipitation) during the flowering period.
• The TabPFN model significantly outperformed other machine learning algorithms, achieving an overall Accuracy of 0.88, F1-score of 0.88, and Area Under the Curve (AUC) of 0.95 across Leave-One-Year-Out (LOYO) cross-validation folds.
• Climatic variables, specifically vapor pressure deficit (VPD), minimum temperature (Tmin), and maximum temperature (Tmax) during the flowering period (July–September), were identified as the most influential factors affecting subsequent yields.
• Previous year's yield and bearing index were among the most important predictors in the TabPFN model.
• Chlorophyll indices (CIG and NDRE) were strong predictors, reflecting their link to canopy health and photosynthetic activity.
• Spectral gradients between flowering and early fruit drop were lower during "on" years for VIs (indicating stable canopy vigor) and steeper for FIs (indicating less flower/fruit abscission), serving as early indicators for predicting alternate bearing.
• TabPFN demonstrated superior temporal stability and generalization capacity across all test years (2020–2024), maintaining high predictive accuracy even during pronounced "off" years.

Contributions

• Development of a high-resolution, remotely sensed framework that integrates Sentinel-2 spectral indices (VIs and FIs) with climate variables to characterize canopy dynamics associated with alternate bearing.
• Application and comparison of multiple machine learning algorithms to detect and classify alternate bearing behavior in commercial avocado orchards.
• Demonstration of the feasibility of remote sensing-driven alternate bearing detection in a data-limited African production context, addressing a longstanding knowledge gap.
• Provision of a scalable, data-driven approach that supports improved orchard-level management and long-term production planning in major avocado-growing regions.

Funding

• Westfalia Fruit Estates (Pty) Ltd. (Project number: TRIM A23/3798)

Citation

@article{Rahman2025Machine,
  author = {Rahman, Muhammad Moshiur and Robson, Andrew and Bekker, Theo},
  title = {Machine Learning Approaches for Assessing Avocado Alternate Bearing Using Sentinel-2 and Climate Variables—A Case Study in Limpopo, South Africa},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17243935},
  url = {https://doi.org/10.3390/rs17243935}
}