This study presents an innovative approach to high-resolution land cover classification using deep learning, tackling the challenge of working with an exceptionally small dataset. Manual annotation of land cover data is both time-consuming and labor-intensive, making data augmentation crucial for enhancing model performance. While data augmentation is a well-established technique, there has not been a comprehensive and comparative evaluation of a wide range of data augmentation methods specifically applied to land cover classification until now. Our work fills this gap by systematically testing eight different data augmentation techniques across four neural networks (U-Net, DeepLabv3 + , FCN, PSPNet) using 25 cm resolution images from Cantabria, Spain. In total, we generated 19 distinct training sets and trained and validated 72 models. The results show that data augmentation can boost model performance by up to 30%. The best model (DeepLabV3 + with flip, contrast, and brightness adjustments) achieved an accuracy of 0.89 and an IoU of 0.78. Additionally, we utilized this optimized model to generate land cover maps for the years 2014, 2017, and 2019, validated at 580 samples selected based on a stratified sampling approach using CORINE Land Cover data, achieving an accuracy of 87.2%. This study not only provides a systematic ranking of data augmentation techniques for land cover classification but also offers a practical framework to help future researchers save time by identifying the most effective augmentation strategies for this specific task.

Gorgonians play a fundamental role in the deep sea (below 200 m depth), composing three-dimensional habitats that are characterized by a high associated biodiversity and playing an important part in biogeochemical cycles. Here we describe the use of a benthic lander to monitoring polyps activity, used as a proxy of gorgonian feeding activity of three colonies of Placogorgia sp. Images cover a period of 22 days with a temporal resolution of 30 min. In addition, this seafloor observatory is instrumented with oceanographic sensors that allows continuous monitoring of the hydrographic conditions in the site. Deep-learning is used for automatic detection of the state of the polyps registered in the images. More than 1000 images of 3 large specimens of gorgonians are analyzed, annotating polyps as extended or retracted, using the semantic segmentation algorithm ConvNeXt. The segmentation results are used to describe the feeding patterns of this species. Placogorgia sp. shows a daily pattern of feeding conduct, depending on the hours of day and night. Using a Singular Spectrum Analysis approach, feeding activity is related to currents dynamics and Acoustic Doppler Current Profile (ADCP) return signal intensity, as proxy of suspended matter, achieving a linear correlation of 0.35 and 0.11 respectively. This is the first time that the behavior of the Placogorgia polyps, directly related to their feeding process, is described.

Despite the well-known strong influence of spatial resolution on the quality of burned area mapping and the need for timely environmental information, global wildfire monitoring services are commonly based on coarse spatial resolution (300–500 m) reflectance imagery and deliver products months or years after the present date. The paper presents, for the first time, an algorithm that provides highly accurate near-real-time medium spatial resolution burned area, from 20 m Sentinel-2 imagery. The paper exploits a pioneering sensor-independent potential of a mapping method, based on land surface reflectance modelling and machine learning, originally optimised for Sentinel-3 imagery. The mapping method uses predictions of time series of burned area from a neural network, which are combined with the spatio-temporal density of active fire detections. The mapping method was calibrated and validated using reference datasets for the years 2020 and 2019, respectively. The novelty of this method lies in its high accuracy and multi-latency flexibility: it achieves a Dice coefficient (DC) of 82.7% with zero-day latency, already surpassing the 81.8% accuracy of current state-of-the-art non-time critical methods. As reflectance data availability increases, accuracy scales to DC 84.7% and 85.4% with 5 and 10 days of latency, respectively, and to DC 87.2% for monthly composites with 45 days of latency.

Accurate estimation of forest Canopy Height Model (CHM) is essential for understanding ecosystem structure, biomass, and carbon dynamics. Machine learning (ML) and deep learning (DL) models have shown strong potential when trained using LiDAR-derived (Light Detection and Ranging) canopy heights as reference data. However, the influence of LiDAR point density on model performance remains poorly understood. This study systematically evaluates three modeling frameworks, Extreme Gradient Boosting (XGBoost), Convolutional Neural Networks (CNN) and hybrid CNN–Transformer architectures, for predicting CHM across multiple regions in Spain, using LiDAR data with varying point densities (1–5 points/m²) from the Plan Nacional de Ortofotografía Aérea (PNOA). Models were trained using multiple sources of remote sensing data: Sentinel-2 spectral bands (visible and near-infrared bands at 10 m) and Sentinel-1 VH backscatter and topographic variables (slope and elevation). Model performance was assessed in terms of predictive accuracy, generalization capacity and sensitivity to LiDAR point density data. Results show that CNN-based approaches outperform classical ML across all datasets, and that integrating Transformer modules further improves performance particularly in areas with higher LiDAR density. The best results were achieved on the dataset with higher point density (5 points/m²), where the CNN + Transformer model reached a Mean Absolute Error (MAE) of 3.37 ± 0.1 m and an R² of 0.66 ± 0.02. Lower-density LiDAR datasets degraded the accuracy for all methods. These findings highlight that both model architecture and LiDAR point density critically influence CHM estimation. The results provide practical guidance for designing large-scale structural mapping workflows in regions with heterogeneous LiDAR coverage, helping to support forest management, carbon accounting and environmental decision-making.

Land cover mapping is essential for territorial management due to its links with ecological, hydrological, climatic, and socioeconomic processes. Traditional methods use discrete classes per pixel, but this study proposes estimating cover fractions with Sentinel-2 imagery (20 m) and AI. We employed the French Land cover from Aerospace ImageRy (FLAIR) dataset (810 km² in France, 19 classes), with labels co-registered with Sentinel-2 to derive precise fractional proportions per pixel. From these references, we generated training sets combining spectral bands, derived indices, and auxiliary data (climatic and temporal variables). Various machine learning models—including XGBoost three deep neural network (DNN) architectures with different depths, and convolutional neural networks (CNNs)—were trained and evaluated to identify the optimal configuration for fractional cover estimation. Model validation on the test set employed RMSE, MAE, and R² metrics at both pixel level (20 m Sentinel-2) and scene level (100 m FLAIR). The training set integrating spectral bands, vegetation indices, and auxiliary variables yielded the best MAE and RMSE results. Among all models, DNN2 achieved the highest performance, with a pixel-level RMSE of 13.83 and MAE of 5.42, and a scene-level RMSE of 4.94 and MAE of 2.36. This fractional approach paves the way for advanced remote sensing applications, including continuous cover-change monitoring, carbon footprint estimation, and sustainability-oriented territorial planning.

El seguimiento preciso del ciclo del carbono en ecosistemas agroforestales es un reto crítico frente al cambio climático. Este trabajo presenta una metodología innovadora que integra modelos mecanísticos de fotosíntesis con teledetección multiespectral y térmica del Programa Copernicus (Sentinel-2/3 y Landsat). A diferencia de los modelos tradicionales de eficiencia en el uso de la luz (LUE) basados en ajustes estadísticos, esta aproximación incorpora estimaciones de conductancia estomática derivada de modelos de balance de energía como variable crítica, permitiendo capturar con precisión el impacto del estrés hídrico en la fijación de carbono. La metodología se aplicó en una parcela piloto silvopastoral en Cantabria. Se empleó el modelo de balance energético de dos fuentes (TSEB) con el fin de separar la transpiración de la evaporación y derivar así la conductancia estomática como indicador del estado fisiológico. Este parámetro alimentó un modelo de fotosíntesis basado en Farquhar para calcular flujos diarios y anuales de carbono, obteniendo indicadores como la Productividad Primaria Bruta (GPP) y el Intercambio Neto del Ecosistema (NEE). Los resultados de 2024 revelaron variaciones significativas en la fijación de carbono según la densidad vegetal y la humedad del suelo, permitiendo clasificar las parcelas como sumideros o fuentes de CO2.