Javier González-Enrique

Oil spills represent a severe threat to aquatic ecosystems, requiring rapid and reliable detection methods to support environmental response. Hyperspectral imaging (HSI) offers high spectral resolution for distinguishing hydrocarbon types, but its effective use depends on the performance and robustness of deep learning (DL) models, especially under data-limited conditions. This study presents a systematic evaluation of convolutional neural network (CNN) configurations for oil spill classification in visible-near-infrared (VNIR) hyperspectral data, examining the influence of architectural depth and hyperparameters such as the number of convolutional kernels, neuron density, and dropout rate. Two architectures were tested across 54 configurations and two training set sizes (259 and 518 samples). Results show that a compact architecture with an additional max pooling layer achieved near-perfect accuracy (>0.99) with reduced complexity and greater robustness, outperforming its deeper counterpart. Importantly, this study reveals that under small-sample scenarios, optimal performance can still be achieved by carefully balancing model capacity, favouring moderate convolutional depth and high neuron density, while avoiding over-regularisation. These findings provide practical guidance for designing efficient CNNs for UAV-based oil spill monitoring and lay the groundwork for future integration into local real-time processing pipelines and transfer learning applications.

The detection of hydrocarbon spills in marine environments is a critical challenge due to their severe impact on aquatic ecosystems. Hyperspectral imaging technology has emerged as a promising tool for detecting such spills, offering high-resolution spectral data for accurate identification. This work aims to develop an automated procedure for detecting hydrocarbon spills in local marine areas using a compact hyperspectral camera installed on a drone. The methodology integrates machine learning with hyperspectral technology, employing hyperspectral signatures of various hydrocarbons and water to create a spectral knowledge base for classification. A novel approach is used in the classifier training, where hyperspectral signatures are directly utilized as input to an artificial neural network, bypassing the need for extensive image labelling, eliminating the requirement for a large image dataset, and significantly reducing the computational costs associated with deep learning models. The trained classifiers, using autoencoders for encoding hyperspectral signatures, were applied to classify hyperspectral profiles (hyperpixels) in hyperspectral images. The results were highly promising, with the classifiers achieving near-perfect accuracy in detecting contaminated water and distinguishing between different hydrocarbons. Recall and F-score values above 0.90 demonstrate the effectiveness of the proposed approach, offering an efficient, scalable solution for real-time monitoring of marine spills.

The objective of this research is to develop accurate forecasting models for chlorophyll-α concentrations at various depths in El Mar Menor, Spain. Chlorophyll-α plays a crucial role in assessing eutrophication in this vulnerable ecosystem. To achieve this objective, various deep learning forecasting techniques, including long short-term memory, bidirectional long short-term memory and gated recurrent uni networks, were utilized. The models were designed to forecast the chlorophyll-α levels with a 2-week prediction horizon. To enhance the models’ accuracy, a sliding window method combined with a blocked cross-validation procedure for time series was also applied to these techniques. Two input strategies were also tested in this approach: using only chlorophyll-α time series and incorporating exogenous variables. The proposed approach significantly improved the accuracy of the predictive models, no matter the forecasting technique employed. Results were remarkable, with |$\overline{\sigma}$| values reaching approximately 0.90 for the 0.5-m depth level and 0.80 for deeper levels. The proposed forecasting models and methodologies have great potential for predicting eutrophication episodes and acting as decision-making tools for environmental agencies. Accurate prediction of eutrophication episodes through these models could allow for proactive measures to be implemented, resulting in improved environmental management and the preservation of the ecosystem.