INTRODUCTION
Sugar beet is one of the more technified agro industries in Spain. In the last years, it has leaded as well the digital transformation with the objective of maintaining sugar beet competitivity both national and internationally. Among other lines, very high potential has been identified in determining the sucrose content using a combination of Artificial Intelligence and Remote Sensing. This work presents the conclusions of an extensive data acquisition task, creation of a model capable of estimating the moment of maximum sucrose content in the sugar beet and its validation.
DATA
Root and foliar samples have been gathered during the complete crop cycle to capture and trace sugar beet behaviourduring the last four seasons (2018-2021) in a broad scale containing all the Spanish sugar beet geography. Samples have been analysed in laboratory obtaining, in the root samples and among others, polarization and net weight parameters, needed to estimate sucrose content.
In addition, data such as geographical values, season-related and time-series of remote sensing have been used.
METHODOLOGY
A complete KDD process has been performed. Starting with a depurative process of the variables. Dependent variables: weight and polarization, were addressed with a population study, measuring its polynomic trend, and analysing from an agronomical perspective those with more than two standard deviations. With independent variables, correlation and distribution techniques were applied. To work in a global scale, PCA techniques were applied to reduce the dimensionality of the fifty-five variables and detect, with the help of dendrograms the outliers that were not detected in the univariable processes.
MODEL
80% of the population was used for training, and 20% for validation. For the segmentation of the train/test, a stratified sampling of histogram of the variable to predict was used. This was performed so the data set and the test could be included in a similar histogram. And to ensure that the model has all the needed values to train and test. To create the model, we worked with a MLP Neural Network, studying different configurations for a better result. Obtaining a model of 3.748 samples, with R2 of 96,03 and MAE of 0,42 Sucrose Tonnes per Hectare.
VALIDATION
Externally to the team involved in the project, a sampling campaign has been carried out at two dates: the date decided with the conventional methodology and the date proposed by the model. A total of thirteen fields have been involved in this validation. With this campaign, an increase of 9% in sucrose has been measured. Potential increment in ideal conditions has been established at 20%.
CONCLUSIONS
A model has been generated and put in place operatively with a neural network structure, formed by 3.748 final samples, obtaining a precision measured by an R2 of 96,03 and MAE of 0,42.
In its more operative stage, the sucrose yield increment has been demonstrated, determining the harvest based on the optimal moment described by the model. The validation carried out in thirteen fields has measured an increment of 9%.
