In one of Matogen Applied Insights’ (MAI) flagship agricultural analytics projects, the data science team examined how modelling leaf analysis data can enable crop nutrient deficiency prediction.
Plant leaf analysis
Plant leaf analysis is the only truly accurate means to determine whether plants are suffering from nutrient deficiencies and what the consequent nutritional requirements are. It is a critical component of the farm management process to inform decision-making surrounding the dosage and frequency of fertiliser application, in order to optimise efficiency and environmental sustainability.
The leaf analysis process is labour and time intensive, mainly due to the physical leaf sampling process as well as transportation to laboratories. Once the leaf sample arrives at the laboratory, a variety of tests are conducted to obtain as much information as possible from the sample. A wide range of nutrients can be measured including macronutrients such as nitrogen, potassium and phosphate, major nutrients such as calcium, sulphur and magnesium, as well as heavy metals such as lead, nickel, arsenic, cadmium, chromium, mercury, copper and zinc.
Leaf sample analysis and modelling
A client, a major player in the South African agricultural sector, approached Matogen Applied Insights (MAI) to conduct a pilot study to analyse a vast amount of leaf analysis data to identify trends and examine the feasibility of “virtual sampling”. The primary objective of analysing the data was to determine in which areas sufficient numbers of samples were required to conduct the next phase of the exercise, “virtual sampling”.
For the initial stage of the project the focus was on soybeans and maize, as well as winter wheat crops in KwaZulu-Natal.
Data analysis and modelling
Analysing the data revealed that the number of samples taken varied greatly across province and growth stage. MAI developed a simple model based on the frequency of the nutrient levels of interest and the number of samples for specific areas to produce a matrix highlighting sampling priority per crop, location and growth phase.
In areas where sufficient numbers of samples were taken, leaf sample data was combined with NDVI and weather station data to examine correlations. New variables were created to measure specific relationships. Statistical modelling techniques were applied to predict likelihood of nutrient deficiencies at specific locations. The most effective method produced a model for plant nutrient deficiency prediction with 95% precision on test data.
Virtual sampling involves making inferences about a sample using historical leaf samples and other data. The model produced by MAI sets the stage to explore the extent to which virtual sampling can be used to decrease the number of physical plant leaf samples required. As this aspect of leaf analysis is the most costly, as mentioned above, reducing the number of samples required without decreasing information derived would lead to significant cost and efficiency gains in this crucial aspect of agriculture management. Prioritising physical leaf sample-taking at the locations and during the growth stages highlighted in the data analysis component of the project would enable the implementation of “virtual sampling”.