Estudo comparativo entre redes neurais artificiais e interpolação de dados por krigagem aplicados a altitude

Abstract

This work investigates the potential of Artifcial Neural Networks (ANNs) as an alternative to ordinary kriging interpolation. Using altitude attributes, normalization routines (based on mean and standard deviation) and data pre-processing (considering nearby sample points) were developed and ANN architecture was optimized by conducting tests to evaluate topologies, activation functions, learning rates, and number of epochs. It was found that topologies with four hidden layers, eight to two neurons per layer, and the Mish activation function showed the lowest squared error in the testing and validation procedures. The developed pre-processing routines were inspired by classical interpolation methods, primarily utilizing nearby sample points to make predictions. Regarding normalization processes, those related to the mean and standard deviation of the analyzed dataset’s altitude resulted in improved performance. After the network’s development, its execution resulted in a mean squared error of less than three in the validation sets, indicating the possibility of applying the developed model in precision agriculture. Finally, in the cross-validation comparison, the confgured ANN model performed inferiorly to kriging. Despite this, the generalization capacity demonstrated by the ANN with altitude data in preliminary tests suggests that future optimizations in the pre-processing methods or the use of newer network architectures could make ANNs viable alternatives for data interpolation.