Forests are crucial for biodiversity, climate regulation, and hydrological cycles, requiring sustainable management due to threats like deforestation and climate change. Traditional forest monitoring methods are labor-intensive and limited, whereas UAV LiDAR offers detailed three-dimensional data on forest structure and extensive coverage. This study primarily assesses individual tree segmentation algorithms in two forest ecosystems with different levels of complexity using high-density LiDAR data captured by the Zenmuse L1 sensor on a DJI Matrice 300RTK platform. The processing methodology for LiDAR data includes preliminary preprocessing steps to create Digital Elevation Models, Digital Surface Models, and Canopy Height Models. A comprehensive evaluation of the most effective techniques for classifying ground points in the LiDAR point cloud and deriving accurate models was performed, concluding that the Triangular Irregular Network method is a suitable choice. Subsequently, the segmentation step is applied to enable the analysis of forests at the individual tree level. Segmentation is crucial for monitoring forest health, estimating biomass, and understanding species composition and diversity. However, the selection of the most appropriate segmentation technique remains a hot research topic with a lack of consensus on the optimal approach and metrics to be employed. Therefore, after the review of the state of the art, a comparative assessment of four common segmentation algorithms (Dalponte2016, Silva2016, Watershed, and Li2012) was conducted. Results demonstrated that the Li2012 algorithm, applied to the normalized 3D point cloud, achieved the best performance with an F1-score of 91% and an IoU of 83%.

Benthic communities, such as seagrass meadows, play a crucial environmental role in marine ecosystems and provide socio-economic benefits. Satellite remote sensing is currently used for their monitoring, and Deep Learning (DL) techniques offer improvements in mapping quality compared to traditional Machine Learning (ML). This study compares conventional ML and convolutional DL models for mapping Cymodocea nodosa meadows in El Río, Canary Islands, using WorldView-2 satellite imagery. An in-situ measurement campaign was conducted to generate an open dataset for segmentation. Evaluated models include Decision Trees, Gaussian Naïve Bayes, Support Vector Machines, K-Nearest Neighbors, Subspace KNN, Feedforward Neural Networks, U-Net, Attention U-Net, and Pix2Pix models. Results show that DL models significantly outperform conventional ML models in detecting Cymodocea nodosa. The best model (U-net) achieved an Intersection over Union (IoU) of 83% overall and 74% for Cymodocea nodosa, while the best ML model (FNN) only reached 62% and 23%, respectively. IoU was highlighted for its sensitivity to minor mapping changes. Additionally, a temporal analysis revealed a dramatic 96% reduction in Cymodocea nodosa coverage over 21 years, from 245.32 ha in 2001 to 9.31 ha in 2022. This study not only compares conventional ML and convolutional DL techniques for benthic habitat mapping but also provides a valuable methodology and dataset for future marine ecosystem monitoring research.

Accurate estimation of above-ground biomass (AGB) is essential to understanding carbon stocks and flows, monitoring forest health, assessing biodiversity, and tracking ecological disturbances, which together help to inform climate policies. Imminent global satellite biomass missions (such as ESA’s BIOMASS and NASA-ISRO’s NISAR satellites) will offer valuable environmental monitoring, but their low spatial resolution limits their application in detailed local assessments. In this study, we present BiomSHARP (Biomass Super-resolution for High Accuracy Prediction), a deep learning (DL) model that extends the Hierarchical Attention Transformer (HAT) architecture adapting it to enhance coarse-resolution biomass maps by fusing them with high-resolution multispectral data from sensors such as Sentinel-2 or Landsat. BiomSHARP achieves 25-meter biomass predictions—four times the spatial resolution of the input—bridging the gap between global-scale monitoring and local-scale applications. In a first set of experiments, conducted in a local area in Europe, we demonstrate that BiomSHARP outperforms both traditional interpolation methods and state-of-the-art DL interpolation and prediction approaches for high-resolution AGB estimation across all evaluated metrics (MAE, MSE, RMSE, PSNR and SSIM), while using a comparable/lower number of parameters. Moreover, the model exhibits strong global-scale generalization, as demonstrated by its ability to accurately estimate biomass across diverse climatic regions despite being trained on a limited subset of data. Furthermore, the model presents strong temporal generalization, achieving improved performance in estimating AGB from 2020 data even when trained solely on 2010 data. We also analyze the impact of different combinations of spectral bands on biomass estimation, identifying optimal subsets that reduce redundancy and improve computational efficiency. BiomSHARP represents a promising approach to advance global environmental assessments and support improved climate strategies. The code and models are publicly available at https://github.com/laiaalbors/biomsharp.