The huge wastes on Nitrogen fertilizer for rice cultivation and the consequent pollution of the environment, created the need to map the rice paddies’ needs in fertilizer with more efficient ways. With the rapid development of UAVs (unmanned aerial vehicle) equipped with high-tech multispectral sensors, the collection of field data, even from flooded paddies, became way easier. Most of the models created so far for this cause, are empirical linear regression models, that make use of vegetation and hyperspectral indices. This study implements machine learning algorithms based in Support Vector Regression (SVR) and multikernel learning models. The final adjustment of the kernels was based in a similarity measurement with the ideal output kernel. Two (2) main models were developed for this purpose and each architecture was tested in three (3) different scenarios. All aimed to predict each of the eighteen (18) output field traits combining the input data in different ways. The inputs were the one hundred seventeen (117) vegetation indices, which were collected during the three main growth stages of the plant. Two different preprocessing methods were applied in the input data and, depending on the model’s architecture, they created a final combined kernel for each output variable. The different approaches and scenarios are then compared to find the optimal one. For each output value that the models are able to predict, the most important features are selected, for the regarding prediction. Finally, some suggestions for further improvement are stated.
Key words: support vectors, regression, multikernel, kernel alignment, vegetation indices
There are two main models created for the purpose of my Master thesis. Their goal was to train few multikernel models (described on the individual readme files in each specific forlder) in order to predict the output agronomic traits. The models differ in:
- Scaling
- Input and/or output
- Robust Scaler
- Standard Scaler
- Input and/or output
- Type of splitting in training and test set
- (a), 80% train set - 20% test set (regarding the year)
- (b), 80% train set - 20% test set (regarding the year & the type of treatment)
- LOOT, Leave One Out Testing
- RBF Kernels' parameters selection
- static
- dynamic
- Grid Search - Cross validation
- LOOCV, Leave One Out Cross Validation
- Holdout
