Jose Manuel Fonseca

Remote sensing image super-resolution aims to reconstruct high-quality spatial details from low-resolution observations, which is essential for land-cover interpretation, object detection, and downstream geospatial analysis. However, existing convolutional neural network (CNN)- and Transformer-based methods may struggle to preserve sharp boundaries, high-frequency textures, and structural consistency in complex remote sensing scenes. To address these issues, this paper proposes an Edge-Guided Kolmogorov–Arnold Network (EGKAN) for remote sensing image super-resolution. The proposed method integrates the adaptive nonlinear representation capability of Kolmogorov–Arnold Networks (KANs), the global context modeling ability of Transformers, and explicit edge-aware structural modeling to enhance structural reconstruction in complex scenes. Specifically, EGKAN consists of three key components: KAN-based nonlinear feature modeling, edge-aware structure modeling, and multi-scale structural consistency supervision. First, a B-spline-based KAN activation module and Feature KAN Attention are introduced to enhance feature representation through adaptive nonlinear mappings and efficient feature-wise interaction. Second, an edge-aware structure modeling framework is designed, which progressively enhances edge representation through an Edge Attention Block, Multi-direction Edge KAN Modeling, and Edge-aware KAN Refinement, supporting the reconstruction of structural boundaries and fine-grained textures. Finally, a multi-scale edge consistency loss is proposed by jointly incorporating Sobel and Laplacian gradient constraints, which enforces structural consistency and improves boundary fidelity across different spatial scales. Extensive experiments on a self-constructed Sentinel-2 dataset demonstrate that EGKAN achieves favorable quantitative performance compared with representative CNN-, GAN-, and Transformer-based methods in terms of PSNR, SSIM, and perceptual quality. The results indicate that KAN-based nonlinear modeling can contribute to high-frequency structure reconstruction when integrated with Transformer-based global modeling and explicit edge-aware components, demonstrating the potential of KAN–Transformer hybrid architectures for remote sensing image super-resolution.

1. Earth observation remote sensing plays a pivotal role in monitoring forest ecosystems (Xiao et al. 2019). An example of this is the Global Ecosystem Dynamics Investigation (GEDI), which was spec...

Age Related Macular Degeneration (ARMD) is a frequent disease appearing mainly in elderly people, commonly diagnosed by ophthalmologists during patient examination. ARMD is characterized by the accumulation of extra cellular materials under retina seen in retina images as yellowish spots and called Drusen spots. Until now, ophthalmologists analysis has done manually, based only on qualitative aspects. However this process is highly dependent on the ophthalmologist and is very difficult to reproduce. In this paper we propose a methodology for an automatic Drusen analysis, which models each Drusen spot calculating important information like Drusen area, volume and location. The proposed methodology is a reproducible process independent of the clinician allowing the comparison of, images taken in different instants to evaluate treatment effectiveness and disease evolution. To achieve our objectives four different algorithms were used: the first divides the original image in smaller images grouping Drusen neighbours, the second algorithm is based in a Gaussian Blur filter and corrects the non-uniform illumination. The third algorithm based on labelling gradient paths and is responsible for giving Drusen spots location and intensity. The last algorithm is based in Levenberg-Marquardt methodology and is responsible for this modelling of each Drusen spot.

The undeniable importance of forests in the sustainability of the planet Earth and its impact on the climate change increases interest in monitoring these ecosystems. Within the parameters that support the management of forests there is the Forest Height (FH), which not only provides information by itself, but also can be used as a proxy of other parameters, as the biomass. Multispectral observations jointly with machine learning regressors have been used to produce FH maps, however most of them were fitted using large calibration sets covering the area being mapped. In this article, a FH mapping model optimised to be calibrated with sparse and limited calibration sets is proposed. To achieve this objective, Sentinel-2 and ancillary data is used, with robust feature generation and processing steps, together with a stacking regressor. For the calibration and posterior testing, Light Detection and Ranging (LiDAR) data acquired during Airborne Laser Scanning campaigns was used. The approach was tested in 16 study areas covered by Mediterranean forests, achieving a RMSE/rRMSE in the range 1.50-6.06m/13.58-27.15%, respectively, although it only resorted to calibration sets that only covered 25% of the study areas (decreasing the amount of LiDAR data needed for calibrating FH regressors).

The importance of mapping the Forest Height (FH) is increasing due to the more frequent impacts of the climate change in the society (wildfires, droughts, and extreme weather events). Remote Sensing is often used for mapping this variable; however, it usually relies in costly and extensive field or airborne campaigns. Also, when using Synthetic Aperture Radar most approaches do not use freely available data. Considering this, in this work a model is proposed that resorts to ALOS-2, Sentinel-1 and ancillary data. Airborne Laser Scanning (ALS) data is used for local calibration but, with the aim of developing a more scalable model, the latter is optimised to work with small calibration datasets (representative of just 25% of the study area to be mapped). With this purpose, the model combines a featuring generation and a features processing stage with a Stacking Regressor to produce estimates at the pixel level. Their impact was assessed, and an improvement of 8.11 percentual points (pp) and 2.01pp in the relative Root Mean Square Error (rRMSE) was achieved by including the features generation and features processing stages, respectively. Additionally, when the multifrequency dataset was used, the model achieved a rRMSE better than when using only a C-band dataset (Sentinel-1) or only a L-band dataset (ALOS-2), respectively by 4.21pp and 3.05pp. Finally, the model achieved an average R²/rRMSE of 0.6240/24.30% and 0.5901/22.64% for the validation and test study areas, respectively. The proposed approach revealed to be effective on mapping the FH resorting to multifrequency synthetic aperture radar and small calibration datasets acquired by ALS.

The impacts of the climate change in the society make forest monitoring increasingly important. Consequently, there is a growing interest in mapping variables as the Forest Height (FH). The direct measurement of the FH through field campaigns is expensive and difficult to scale. Alternatively, Airborne Laser Scanning (ALS) campaigns can be used to map it, however they share the same disadvantages of the previous approach. Therefore, Remote Sensing (RS) data have been used for local and large-scale mapping of the FH. In this paper a Regression Methodology (RM) that combines GEDI, ALOS-2, Sentinel-1/2, and ancillary data is proposed for mapping the FH in Mediterranean forests. The proposed RM, tested for the 15 regions of interest, achieves a RMSE/rRMSE of 4.95m/33.93%, when evaluated with GEDI data, and 5.11m/41.70%, when evaluated with ALS data.

In recent years, the application of neural networks in the field of satellite image super-resolution has become increasingly widespread. This paper aims to explore how to utilize images from two completely different data sources (Sentinel-2 and Google Earth Pro) as training data to surpass the spatial resolution limitation of 10 meters provided by the Sentinel-2 satellite and obtain higher resolution image. The Sentinel-2 satellite offers open data with four bands with spatial resolution of 10 meters and a resampling cycle every 5 days. This provides abundant data that can be used as Low-Resolution (LR) images for research. Adopting a multi-image super-resolution processing approach allows for the full utilization of geographical information contained within different images of the same location captured at shorter time intervals, while also eliminating concerns regarding information loss due to weather-affected low-quality images. Google Earth Pro provides images with a resolution of up to 0.15 meters, allowing the High-Resolution (HR) image resolution to be adjusted according to the needs. In our experiments, we proposed Sentinel-2 Google Earth Pro Network (SGNET). We used HR images with a resolution of 2 meters, and through SGNET implementation, a five-fold increase in spatial resolution for Sentinel-2 images was achieved, yielding very satisfactory results.

The knowledge of the Forest Height (FH) is important for monitoring the forests, and it can be used as a proxy variable of other forest parameters as the aboveground biomass. It is also important for understanding the climate change and prepare the wildfire seasons. The most effective way to map the FH is through field campaigns or airborne laser scanning, but both are expensive and not scalable. Alternatively, spaceborne Synthetic Aperture Radar (SAR) data may be used. However, it often relies on the acquisition of large ground truth datasets. In this paper, a new Regression Methodology (RM) that makes use of SAR data and a Stacking Regressor that minimises the amount of data needed to map the FH of a region is presented. Tested on a total of 16 regions between Portugal and Spain, plus one in California, the RM achieved a R ² between 42.12%-62.62%, and a RMSE between 0.96m-4.49m.