Macadamia trees produce a high value crop. In recent years in Australia, the industry has seen significant expansion. However, there have also been challenges such as drought and storms that have reduced production in some areas.

Producers want early yield forecasts, as it aids in harvest planning, insurance, finance and contracting. Block-level or even tree-level forecasts can inform management decisions aimed at reducing variability across orchards. The whole industry desires yield forecasts to assist in marketing and sales planning.

Common with other horticultural tree crops, forecasting macadamia yields is challenging. Trees often exhibit phenomena such as irregular bearing, making yields inconsistent from one year to the next, even for trees with healthy canopies.

Estimation by observation of nut count and volume on individual trees are challenging as trees are large and nuts are obscured by other tree components (leaves and branches). The effect of weather conditions on yields are not well understood, particularly understanding the extent to which previous seasons’ conditions affect the current season’s productivity.

Information sources

Fifteen orchards have provided tree and productivity data. These typically describe tree-planting year, and yield per year at either the block or farm level, many with yield records covering more than 10 years. Orchards cover the whole east coast growing regions, from the mid-NSW coast to north Queensland. This large dataset has enabled investigation of the benefits that incorporation of satellite and weather data into yield forecasting methodologies using machine-learning techniques.

Automatic tree age estimation

The first step was to develop a methodology that can predict macadamia tree age. We correlated the planting year data supplied by the growers with time-series satellite imagery from 1988 to the present (from the NASA/USGS Landsat program). The normalised difference vegetation index (NDVI) was derived from the imagery.

We determined that on average, a three-year-old orchard has an NDVI of 0.62. This algorithm produced average planting year production errors of less than two years over 95 blocks. We generate area by year estimations for the macadamia industry annually, by scaling this algorithm across all orchards in Australia (using the national macadamia map from a related project, also delivered by AARSC).

Yield forecast models

Satellite and weather data was accumulated for each orchard block. Weather variables include minimum and maximum temperatures, evapotranspiration, vapour pressure deficit and rainfall.

Satellite data included many vegetation indices, which indicate factors such as canopy biomass, nutrient status and water stress. Each of these variables was grouped into quarterly and yearly periods.

The yield was modelled using the satellite and weather variables from two years before each harvest year, as well as tree age. A variety of statistical and machine learning algorithms produce the forecast models, which can then be applied to forecast yield for each macadamia orchard block. Forecasts are generated in January, well before harvest (April-September).

We are in our third year of supplying forecasts to our collaborating producers. In 2021, the mean average prediction error for all 160 blocks/orchards was 0.7 tonnes/hectare (average yield 3.1 tonnes/hectare). This is much more accurate than simply using previous year’s yields to predict current yield.

Conclusion

There are challenges that remain in forecasting macadamia yield, such as predicting irregular bearing and differences caused by local effects such as pollination and tree canopy density. However, this work has shown the potential for free public data (satellite and weather) to forecast tree crop yields. The models can be scaled from small orchards to the whole industry, representing new opportunities for data-driven planning and decision making.

*Written by James Brinkhoff an associate professor with the Applied Agricultural Remote Sensing Centre (AARSC), University of New England, Armidale. His research focusses on using spatial and temporal data to derive useful decision support information for growers and agricultural industries.