Resumen de Análisis multitemporal de la degradación del páramo ecuatoriano mediante redes neuronales artificiales: Influencia de variables meteorológicas y detección de cambios

Marco Javier Castelo Cabay

• español

  Esta tesis doctoral explora cómo la inteligencia artificial y el procesamiento de imágenes satelitales pueden ayudar a monitorear y entender la degradación de los páramos en Ecuador, un ecosistema esencial para el ciclo del agua y la captura de carbono. La investigación se enfoca en los páramos del cantón Quero, en Tungurahua, desarrollando una metodología que también puede aplicarse a otras zonas montañosas con condiciones ambientales similares.

  Utilizando imágenes satelitales procesadas a través de Google Earth Engine (GEE), se comparan tres metodologías de clasificación del suelo: el análisis basado en píxeles (PBIA), el análisis basado en objetos (GEOBIA) y redes neuronales profundas (DNN). Los resultados muestran que las DNN permiten clasificaciones más detalladas, ya que captan características complejas de las imágenes. Incorporar variables como altura, pendiente, índices espectrales (NDVI y BSI) y la textura del suelo también mejoró notablemente la precisión, creando un modelo adaptado a la compleja topografía y clima del páramo. Este enfoque proporciona una herramienta de monitoreo más precisa para la conservación de estos ecosistemas.

  Para evaluar los cambios en la vegetación del páramo, la investigación utiliza métodos tradicionales (PELT, DYNP, WINDOW) y redes neuronales avanzadas (LSTM, GRU), enfocándose en el análisis de series temporales del Índice de Vegetación de Diferencia Normalizada (NDVI). Los resultados indican que las redes neuronales superan en precisión y sensibilidad a los métodos tradicionales, especialmente en la detección de cambios abruptos en la vegetación. Esta capacidad de detectar puntos de cambio es valiosa, pues permite tomar decisiones rápidas ante alteraciones del ecosistema, como las causadas por la agricultura y el cambio climático.

  Adicionalmente, se implementaron redes neuronales multivariantes (LSTM y GRU, junto con CNN-LSTM y LSTM con mecanismos de atención) para predecir la temperatura en el páramo, tomando en cuenta variables climáticas como la precipitación, humedad y velocidad del viento. Los modelos mostraron una precisión superior a la de métodos convencionales como ARIMA, permitiendo anticipar cambios que podrían afectar los ciclos hidrológicos y la biodiversidad, factores que influyen en la disponibilidad de agua y en la agricultura de las comunidades locales.

  En conclusión, los resultados validan la hipótesis central de la investigación: la combinación de teledetección y aprendizaje automático mejora el monitoreo del páramo ecuatoriano, facilitando la conservación y restauración del ecosistema. La tesis destaca la importancia de tecnologías como GEE y las redes neuronales para abordar las limitaciones de acceso y análisis de datos en áreas montañosas complejas. Además, este enfoque metodológico es replicable y adaptable, representando un marco de trabajo para la gestión ambiental y la restauración de ecosistemas montañosos en riesgo. La investigación contribuye al campo de la ecología computacional y la conservación, promoviendo la sostenibilidad ambiental y la resiliencia frente al cambio climático en escenarios similares a nivel global.

• English

  The p´aramo is a high-altitude ecosystem crucial for water regulation and carbon sequestration, playing a vital role in biodiversity and environmental balance.

  However, these ecosystems are under threat due to climate change and human intervention, which has led to significant degradation. The main causes of this degradation include the expansion of the agricultural frontier, vegetation burning for grazing, and climate change, resulting in biodiversity loss, disruption of hydrological cycles, and increased soil erosion. This doctoral thesis, titled “Multitemporal Analysis of the Degradation of the Ecuadorian P´aramo using Artificial Neural Networks: Influence of Meteorological Variables and Change Detection”, investigated the degradation of the Ecuadorian p´aramos, specifically in the area of Quero, Ecuador, using advanced remote sensing technologies and machine learning.

  The research focused on three main areas: land use and cover classification, the detection of temporal disturbances, and the prediction of critical climate variables. To achieve this, a land cover classification of the Ecuadorian p´aramo was developed using satellite images processed through Google Earth Engine. Three approaches were compared: Pixel-Based Image Analysis, Geographic Object-Based Image Analysis, and Deep Neural Networks (DNN).

  The study used Sentinel-2A images, utilizing indices such as NDVI and BSI, as well as additional parameters such as texture, altitude, and slope. The data were obtained with the collaboration of local experts and georeferencing systems.

  The evaluated approaches included variants of Object-Based Image Analysis, using both spectral indices and topographic parameters, and the DNNs that leveraged all spectral bands from Sentinel-2A along with the NDVI and BSI indices. The results showed that Deep Neural Networks achieved the highest accuracy, reaching 87.43 %.

  The study also focused on the detection of breakpoints in time series of the Normalized Difference Vegetation Index (NDVI) in the Ecuadorian p´aramo, evaluating traditional methods such as PELT, DYNP, and WINDOW, comparing them with advanced techniques based on neural networks such as LSTM and GRU. Simulated and real NDVI data were used, processed to have a daily frequency, and interpolation and normalization techniques were applied to prepare the data. Advanced neural network techniques were shown to be more effective than traditional methods in detecting breakpoints. Initially, simulated data were used to calibrate the accuracy of the methods, and then the methodology was applied to real NDVI data in the study area. The results showed that LSTM and GRU neural networks offered greater effectiveness, with notable precision and sensitivity, achieving an F1 Score of 60 %. In contrast, traditional methods performed poorly. This research highlighted the importance of adapting breakpoint detection algorithms to the specific characteristics of the data and application context, providing more precise and efficient tools for monitoring and conserving the Ecuadorian p´aramo.

  The study also aimed to predict temperature using multivariate time series in two locations in the province of Tungurahua, Ecuador: Mula Corral in the p´aramo ecosystem and Ambato airport in the city center, with the goal of conducting a comparative analysis and highlighting the advantages of predicting temperature in the planning of p´aramo restoration and adaptation to climate change. Daily data collected between March 12, 2013, and March 31, 2023, from both meteorological stations were processed through Google Earth Engine. Six neural network architectures were evaluated: LSTM, bidirectional LSTM, attention-based LSTM, GRU, CNN, and CNN-LSTM, with LSTM and GRU networks being more accurate in Mula Corral, while the attention-based LSTM was the most effective in Ambato.

  The results showed significant climate differences between the p´aramo and the city, emphasizing the importance of precise models for environmental management and climate change adaptation. The research concluded that temperature predictions using neural networks are essential for the effective regeneration of p´aramos, facilitating the optimal implementation of conservation techniques and adaptation to climate fluctuations.

  P´aramos are not only essential for ecology and the environment but also support local human communities by providing water resources and contributing to agriculture. The degradation of these ecosystems puts both biodiversity and the quality of life of local inhabitants at risk. In this context, the application of advanced technologies such as machine learning and neural networks offers new opportunities to monitor, predict, and manage changes in the p´aramo more precisely and efficiently. This research not only provides new methodologies for remote sensing and change analysis in the Ecuadorian p´aramo but also provides a solid scientific basis for informed decision-making in the conservation and restoration of these valuable ecosystems.