Prediction of Actual Evapotranspiration and Analysis of Air Temperature Trends and Pattern Similarities: A Machine Learning and Statistical Approach

Actual Evapotranspiration (AET) refers to the total amount of water transferred from the Earth’s surface to the atmosphere through the processes of evaporation (from soil, water, and other surfaces) and transpiration (from plants, mainly through their leaves). AET is a crucial element of the hydrological cycle, influencing the energy balance across soil, atmosphere, and vegetation systems. An accurate understanding of AET is essential for effective water resource management, especially within agriculture, hydrology, and climate research, as it directly impacts water availability and ecosystem health. It is influenced by several factors, including temperature, soil moisture, humidity, vegetation cover, and wind speed. AET is one of the meteorological variables for which the installation of eddy covariance systems or lysimeters at remote sites is costly and logistically challenging. It is therefore argued that it should be predicted using machine learning models. Climate change is closely linked to rising air temperatures, which cause global warming. This warming effect alters temperature patterns, resulting in more extreme weather events, shifts in seasonal temperatures, and changes in diurnal and annual temperature trends. As air temperature plays a central role in understanding climate dynamics, it is crucial to assess how temperature patterns evolve over time. Trend analysis and similarity of time patterns provide valuable insights into these changes. Statistical analyses, such as parametric and non-parametric methods, help quantify temperature trends, while machine learning models, such as clustering and time series analysis, capture complex temporal patterns and similarities within and across regions. Both approaches work well together to provide a clear picture of how air temperature is changing over time. This combined understanding helps to develop strategies for adapting to climate change. This thesis focused on predicting AET using machine learning models, enhanced by grid search for hyperparameter tuning. A feature selection technique was used to identify the most relevant variables for the models. The deep learning models outperformed classical and statistical machine learning models in prediction accuracy. To further improve performance, Bayesian optimization was applied to both classical and deep learning models. These optimized models were then evaluated using two variable combinations: the selected variables from feature selection and a set of readily available meteorological variables. In both cases, the deep learning models produced promising results for AET prediction. This demonstrates that deep learning models can effectively capture both linear and complex non-linear relationships within the dataset. This thesis presents statistical methods for detecting and quantifying air temperature trends across various time scales, including annual, monthly, and diurnal intervals (10-, 30-, and 60-day windows). To analyze temperature patterns and similarities, an unsupervised machine learning approach, hierarchical clustering combined with Dynamic Time Warping (DTW), was employed, allowing for the assessment of mean air temperature and trends within and across regions. The similarity between the 32 stations was further examined using a distance correlation matrix. The findings indicate a significant warming trend in most Italian stations, whereas non-significant cooling trends were observed in UK stations. Generally, stations within the same region exhibited similar patterns, while distinct patterns emerged between regions. Overall, the machine learning models have been developed to predict AET and to cluster mean air temperature and its trends. This assists with irrigation scheduling to improve agricultural productivity. Observed temperature trends and regional patterns provide insights into the broader impacts of climate change, emphasising the importance of regional analysis for understanding climate variability.