Improving harvesting operations in an oil palm plantation

Abstract

Oil palm agricultural systems involve large extensions of land that demand careful planning of the harvesting operations. The plantation manager is in charge of synchronizing resources (i.e., crews and a complex cableway network) to harvest at the right time to maximize palm oil yield in latter stages of the value chain. In particular, in this type of crop it is ideal to visit each palm every eight to twelve days to avoid loose fruit picking, over-ripeness, or rotten fruit harvesting. To optimize harvest operations, we propose an end-to-end analytics solution involving data treatment, descriptive (simulation), and prescriptive models (optimization) in this agricultural system. At the core of our approach lies a set of interconnected models that use optimization, heuristic techniques, and simulation. These models cover strategic (harvest cycle), tactical (resource allocation), and operational (transport allocation) decisions. We present a case study of a 2000-ha plantation located in the Colombian Orinoquia. The results show a strong potential for improving yield by reducing the harvest cycle length from 19.6 to 8.3 days.

Appendices

Appendix A: Notation tables

See Table 2, 3 and 4.

Table 2 Notation table for the Harvest Cycle Model (Sect. 5.1)

Table 3 Notation table for the Resource Allocation Model (Sect. 5.2)

Table 4 Notation table for the Transportation Allocation Model (Sect. 5.3)

Appendix B: Data preparation

2.1 Ripeness cost function

Table 5 shows the oil yield for each ripeness state based on Durán (2004). We normalized the costs based on the optimal ripeness cost per FFB (unit cost), so that the other ripeness states decrease the extraction rate and affect the cost relative to the optimal ripeness. Then, we estimated the individual cost for each of the land plots, based on their number of FFB. As a result, we have the costs associated with visiting a land plot within each ripeness state.

Table 5 Palm oil yield

We used a similar methodology to estimate the in-field logistics costs. These costs refer to the extra labor hours that a worker spends manually picking spilled fruitlets while harvesting FFB at each state of ripeness. Table 6 shows the extra time to harvest a palm in each of the ripeness states with respect to the optimal ripeness time, where fruit spill is minimal.

Table 6 FFB picking times

Finally, we estimated a weighted average total cost based on the two piecewise functions with the ripeness states domain.

2.2 Yield forecast

After processing over 30,000 data entries, we clustered the plantation into 13 zones, adjusted a time series, and forecasted one period for each zone using R (RStudio Team 2016). Due to data quality at the source, some zones did not have the best adjustment in terms of the mean absolute percentage error (MAPE). The best fit accounts for a MAPE of 20%. For this reason, we incorporated uncertainty in the forecast through a log-normal distribution (which had minimum BIC against Weibull, normal and gamma distributions) and conducted a Monte Carlo simulation to estimate the yield for each land plot. Figure 34 illustrates an example of this process for land plot 52, located at zone 9. For this particular plot, the zone uncertainty was modeled with a log-normal distribution with $\hat{\mu } \left( {{\text{kg}}/{\text{ha}}} \right) = 722.3$ and $\sigma = 0.816$.

Fig. 34

Ensemble simulation of time series, forecast, and yield variability (plot 52, zone 9)

Lastly, we explored the best representation of yield growth given a harvest cycle length between 8 and 20 days. The resultant model was $\hat{Y} = 0.0104X^{2} - 0.0003X^{3}$ with R² of 0.22. Thus, the quadratic effect of harvest cycle length increases by 1% the yield and the cubic effect regulates the growth by 0.03%. Aside from its coefficient of determination, this model depicts the reality of the harvesting process, that is, the yield may increment after the optimal cycle length, but eventually fruit starts to rot.