Thomas Oberthür

Using Bayesian networks to predict future yield functions with data from commercial oil palm plantations: A proof of concept analysis

Bayesian networks were used to predict yield functions from three commercial oil palm estates. The networks were trained using a range of environmental, agronomic and management data routinely collected during plantation management. The Bayesian networks predicted fruit yield (FFB), average weight of fruit bunches (ABW) and average bunch number per hectare (BUNCH_HA). Comparing the predictions of most probable yield against observed data showed the Bayesian networks were highly accurate, with r2 values between 0.6 and 0.9. Predictions for attaining specific yield targets exceeded 75% accuracy for the FFB, 85% for the BUNCH_HA, and 90% for the ABW function. Supplementary analysis compared the precision of the Bayesian networks with artificial neural networks (ANNs), and demonstrated that the Bayesian networks gave equivalent or superior accuracy for every test. The utility of the networks were demonstrated by predicting the probability of achieving above average yield functions for each block across the three estates using a set of hypothetical rainfall and fertiliser input scenarios during the year prior to harvest. For the majority of blocks, the probability of exceeding the yield target depended on the level of fertiliser and rainfall inputs received, indicating that production from these blocks is greatly influenced by prior rainfall and fertilizer. However, some blocks in favourable areas showed a very high probability of exceeding the mean yields at all rainfall and fertiliser inputs, while a number of other blocks showed a consistently low probability of achieving the same productivity; production from these blocks will be resistant to the effects of historic rainfall and fertiliser inputs. The ability of Bayesian networks to represent future yield expectations will greatly assist managers under pressure to improve the economic and environmental sustainability of plantations. The demonstration that machine learning can extract important insight from complex datasets will have broad application in the analysis of big data collected from oil palm as well as other agricultural industries.

Digital Soil Mapping of Soil Properties in Honduras Using Readily Available Biophysical Datasets and Gaussian Processes

Creating detailed soil maps is an expensive and time consuming task that most developing nations cannot afford. In recent years, there has been a significant shift towards digital representation of soil maps and environmental variables and the associated activity of predictive soil mapping, where statistical analysis is used to create predictive models of soil properties. Predictive soil mapping requires less human intervention than traditional soil mapping techniques, and relies more on computers to create models that can predict variation of soil properties. This paper reports on a multi-disciplinary collaborative project applying advanced data-mining techniques to predictive soil modelling for Honduras. Gaussian process models are applied to map continuous soil variables of texture and pH in Honduras at a spatial resolution of 1,km, using 2472 sites with soil sample data and 32 terrain, climate, vegetation and geology related variables. Using split sample validation, 45% of variability in soil pH was explained, 17% in clay content and 24% in sand content. The principle variables that the models selected were climate related. Gaussian process models are shown to be powerful approaches to digital soil mapping, especially when multiple explanatory variables are available. The reported work leverages the knowledge of the soil science and computer science communities, and creates a model that contributes to the state of the art for predictive soil mapping.