Desarrollo de un modelo basado en algoritmos de deep learning para la detección y cuantificación de estado de apertura en imágenes microscópicas de estomas.
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Esta memoria de título aborda la necesidad de automatizar la identificación y medición de los estomas en imágenes microscópicas. Este proceso es fundamental para el análisis hídrico de cultivos, especialmente relevante ante el creciente impacto del cambio climático y la escasez de agua en la agricultura.
Los estomas son estructuras clave en el tejido epidérmico vegetal, esenciales para el intercambio de gases y la fotosíntesis, esto los convierte en indicadores vitales del estado hídrico de las plantas. El estudio propone el uso de metodologías de deep learning y computer vision para superar las limitaciones de los métodos manuales y semiautomatizados actuales, que son ineficientes, costosos y propensos a errores. La investigación esta centrada en dos objetivos principales:
Identificación del estoma, que consiste en la implementacion de un modelo de deep learning especializado en computer vision para la tarea de detección de objetos, basado en una arquitectura YOLO, se desarrolla para localizar estomas. Este modelo alcanza un alto rendimiento en métricas que evaluan distintas capacidades de los modelos. En precisión (0,957), recall (0,956), mAP@.50 (0,983) y mAP@.50:.95 (0,676), demostrando una capacidad robusta y precisa para identificar estomas. La determinación del estado de apertura, se logra mediante un modelo de segmentación de instancias que clasifica directamente el estado (abierto o cerrado). Los resultados del modelo tambien evidencian un buen rendimiento. Con mAP@.50 (0,889)mAP@.50:.95 (0,806), recall (0,906) y precisión (0,741), lo que indica su fiabilidad para la delimitación precisa de los estomas y su estado.
La metodología incluye la recolección y composición de diversos conjuntos de datos, provenientes de distintas fuentes, incorporando imágenes de especies como Cicer arietinum y Arabidopsis thaliana. Se realiza un preprocesamiento y re-etiquetado manual para asegurar la calidad y precisión de las anotaciones, utilizando máscaras de segmentación para delimitar las clases “stomata”, “pore-open” y “pore-closed”. Los resultados demuestran que los modelos desarrollados representan un avance significativo en la automatización del análisis estomático, proporcionando una herramienta fiable para el fenotipado y la gestión agrícola. Además, se explora el impacto de los mecanismos de atención en las arquitecturas del modelo, confirmando su potencial para mejorar el rendimiento de la segmentación. En resumen, esta memoria de título contribuye al campo del computer vision en la agricultura al ofrecer soluciones basadas en deep learning para la identificación y cuantificación del estado de apertura estomática, facilitando así la toma de decisiones informadas para el manejo hídrico de los cultivos.
This thesis addresses the need to automate the identification and measurement of stomata in microscopic images. This process is fundamental for the water analysis of crops, especially relevant in the face of the growing impact of climate change and water scarcity in agriculture. Stomata are key structures in the plant epidermal tissue, essential for gas exchange and photosynthesis. This makes them vital indicators of the water status of plants. The study proposes the use of deep learning and computer vision methodologies to overcome the limitations of current manual and semiautomated methods, which are inefficient, costly, and error-prone. The research is focused on two main objectives: Stomata identification, which consists of the implementation of a deep learning model specialized in computer vision for the task of object detection, based on a YOLO architecture, developed to locate stomata. This model achieves high performance in metrics that evaluate different model capabilities: precision (0.957), recall (0.956), mAP@.50 (0.983), and mAP@.50:.95 (0.676), demonstrating robust and accurate capacity to identify stomata. The determination of the opening state is achieved through an instance segmentation model that directly classifies the state (open or closed). The model’s results also show good performance, with mAP@.50 (0.889), mAP@.50:.95 (0.806), recall (0.906), and precision (0.741), indicating its reliability for the precise delimitation of stomata and their state. The methodology includes the collection and composition of diverse datasets from different sources, incorporating images of species such as Cicer arietinum and Arabidopsis thaliana. Manual preprocessing and re-labeling are performed to ensure the quality and accuracy of the annotations, using segmentation masks to delimit the classes “stomata,” “pore-open,” and “pore-closed.” The results demonstrate that the developed models represent a significant advance in the automation of stomatal analysis, providing a reliable tool for phenotyping and agricultural management. In addition, the impact of attention mechanisms in the model architectures is explored, confirming their potential to improve segmentation performance. In summary, this thesis contributes to the field of computer vision in agriculture by offering deep learning–based solutions for the identification and quantification of stomatal opening states, thus facilitating informed decision-making for crop water management.
This thesis addresses the need to automate the identification and measurement of stomata in microscopic images. This process is fundamental for the water analysis of crops, especially relevant in the face of the growing impact of climate change and water scarcity in agriculture. Stomata are key structures in the plant epidermal tissue, essential for gas exchange and photosynthesis. This makes them vital indicators of the water status of plants. The study proposes the use of deep learning and computer vision methodologies to overcome the limitations of current manual and semiautomated methods, which are inefficient, costly, and error-prone. The research is focused on two main objectives: Stomata identification, which consists of the implementation of a deep learning model specialized in computer vision for the task of object detection, based on a YOLO architecture, developed to locate stomata. This model achieves high performance in metrics that evaluate different model capabilities: precision (0.957), recall (0.956), mAP@.50 (0.983), and mAP@.50:.95 (0.676), demonstrating robust and accurate capacity to identify stomata. The determination of the opening state is achieved through an instance segmentation model that directly classifies the state (open or closed). The model’s results also show good performance, with mAP@.50 (0.889), mAP@.50:.95 (0.806), recall (0.906), and precision (0.741), indicating its reliability for the precise delimitation of stomata and their state. The methodology includes the collection and composition of diverse datasets from different sources, incorporating images of species such as Cicer arietinum and Arabidopsis thaliana. Manual preprocessing and re-labeling are performed to ensure the quality and accuracy of the annotations, using segmentation masks to delimit the classes “stomata,” “pore-open,” and “pore-closed.” The results demonstrate that the developed models represent a significant advance in the automation of stomatal analysis, providing a reliable tool for phenotyping and agricultural management. In addition, the impact of attention mechanisms in the model architectures is explored, confirming their potential to improve segmentation performance. In summary, this thesis contributes to the field of computer vision in agriculture by offering deep learning–based solutions for the identification and quantification of stomatal opening states, thus facilitating informed decision-making for crop water management.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Industrial.
Keywords
Estomas, Fotosíntesis, Cambios climáticos