Evaluación de técnicas de aprendizaje profundo para la detección de productos fitosanitarios en la industria del olivar
Fecha
2024-12-05
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Jaén: Universidad de Jaén
Resumen
El objetivo del TFM es estudiar la capacidad de los métodos de aprendizaje profundo para la clasificar tipos
y niveles de concentración de fitosanitarios a partir de imágenes hiperespectrales de aceitunas y hojas.
Para ello, primero se obtuvieron las áreas de interés para aceitunas y hojas mediante segmentación
semántica y umbralización, respectivamente. A continuación, se implementó una red neuronal
convolucional 1D (CNN-1D) con la que se clasificaron las imágenes obteniendo exactitudes superiores al
90%, sobre todo en la clasificación por niveles, en algunos casos tras optimizar la red.
Los resultados se compararon con los obtenidos con métodos de aprendizaje automático, siendo los
resultados también ajustados, aunque ligeramente inferiores en general a los de la red CNN-1D.
Adicionalmente, se han aplicado métodos de selección de características, obteniendo bandas con
significación física que permiten disminuir la cantidad de datos.
Se comprueba así la utilidad de estas técnicas, como método no invasivo para detectar fitosanitarios en
las aceitunas y hojas.
The objective of the Master's Thesis (TFM) is to study the ability of deep learning methods to classify types and concentration levels of pesticides based on hyperspectral images of olives and leaves. To achieve this, areas of interest for olives and leaves were first obtained using semantic segmentation and thresholding, respectively. Subsequently, a one-dimensional convolutional neural network (1D-CNN) was implemented to classify the images, achieving accuracies above 90%, particularly in level classification, in some cases after optimizing the network. The results were compared with those obtained using machine learning methods, which were also consistent, though generally slightly inferior to those of the 1D-CNN. Additionally, feature selection methods were applied, identifying bands with physical significance that allowed for a reduction in the amount of data. These findings demonstrate the utility of these techniques as a non-invasive method for detecting pesticides in olives and leaves.
The objective of the Master's Thesis (TFM) is to study the ability of deep learning methods to classify types and concentration levels of pesticides based on hyperspectral images of olives and leaves. To achieve this, areas of interest for olives and leaves were first obtained using semantic segmentation and thresholding, respectively. Subsequently, a one-dimensional convolutional neural network (1D-CNN) was implemented to classify the images, achieving accuracies above 90%, particularly in level classification, in some cases after optimizing the network. The results were compared with those obtained using machine learning methods, which were also consistent, though generally slightly inferior to those of the 1D-CNN. Additionally, feature selection methods were applied, identifying bands with physical significance that allowed for a reduction in the amount of data. These findings demonstrate the utility of these techniques as a non-invasive method for detecting pesticides in olives and leaves.
Descripción
Palabras clave
General