This study addresses the challenge of adapting pretrained models, originally designed for three-band (RGB) imagery, to multispectral data with more than three bands, a mismatch that often leads to suboptimal performance in remote sensing tasks. To overcome this limitation, we propose the development of pretrained multispectral deep learning models tailored to the spectral characteristics of the PeruSat-1 sensor for remote sensing applications. We evaluated several architectures from the ResNet family (ResNet34, ResNet52, ResNet101, ResNet152) and VGG16 using a curated dataset that preserves spatial-spectral priors. Each model was trained and tested on classification tasks involving 2, 3, and 4 land cover classes, using both RGB and multispectral inputs. The results show that the proposed multispectral models generalize well, particularly in low- and medium-complexity scenarios, supporting their suitability for transfer learning. These findings underscore the importance and feasibility of creating pretrained models specifically designed for multispectral imagery, enabling more accurate and efficient environmental monitoring, resource management, and other remote sensing applications, without relying on suboptimal adaptations of RGB-based models.

The performance of hyperspectral image classification (HIC) models strongly depends on the informativeness and representativeness of the training data, which directly impacts classification accuracy. Active learning (AL) has been introduced as a strategy to enhance classification performance by selecting informative and representative samples from unlabeled data and incorporating them into the training process. Although AL has shown promising results in various applications, it requires an oracle to label new data. In this work, we eliminate the need for an oracle and adapt the principles of AL to the supervised learning paradigm. We integrate key concepts from AL into supervised learning by iteratively updating a supervised classifier with subsets of labeled and (potentially) informative data extracted from a fully labeled dataset. Experiments conducted on real hyperspectral data demonstrate that our method outperforms conventional supervised learning when implemented with a standard neural network architecture.

Nowadays, the ever-increasing amount of information provided by hyperspectral sensors requires efficient solutions for facilitating subsequent data analysis. Dimensionality reduction plays a central role in this context, as it allows the extraction of meaningful and compact representations from high-dimensional hyperspectral data. Existing methodologies address data representation problems through dimensionality reduction techniques, predominantly employing Principal Component Analysis (PCA), Autoencoders (AE), and more recently, Hyperspectral Orthogonal Autoencoders (HOAE). However, these approaches commonly rely on the entire image to build projection models, which may result in high computational costs. A pragmatic attempt to mitigate such computational challenge is to use a subset of the image data to construct accurate data representation models. In this work, we investigate the extent to which using a reduced number of training samples affects the quality of the latent space generated by AE and HOAE models, and how this impacts classification performance. Experiments conducted on the Pavia University hyperspectral dataset demonstrate that the representation efficacy of the AE and HOAE models significantly exceeds that of traditional hyperspectral dimensionality reduction algorithms, such as PCA. We also show that competitive classification results can be obtained even when the representation models are trained with a small portion of the image, which opens the door to more computationally efficient pipelines.

Despite advances in Deep Learning (DL) for road extraction, this task remains challenging. First, domain shifts in data distribution hinder the inference of pre-trained models to new areas, leading to a drop in classification accuracy. Second, DL-based models require a large amount of labeled training data to achieve robust performance. To address these challenges, this work proposes an Unsupervised Domain Adaptation (UDA) approach leveraging the Domain Adversarial Neural Network (DANN) strategy applied to Unmanned Aerial Vehicle (UAV) imagery. While most existing approaches rely on satellite imagery, they may not generalize well to UAV data, as very high-resolution images with fine-grained road details introduce additional domain adaptation challenges. Furthermore, since DANN operates at the feature level, the design of the feature extractor plays a key role to achieve the domain alignment. To investigate this, we evaluate our approach with three segmentation models: DeepLabv3+, ResU-Net, and Att-ResU-Net, the latter incorporating attention-enhanced skip connections. Experimental results demonstrate that UDA effectively deals with domain shift, improving road extraction performance by 1.79-6.07% on F1-Score and 2.12-7.85% on IoU when tested on the target domain without labeled training data. Among the evaluated architectures, Att-ResU-Net achieves the highest UDA performance. The qualitative analysis through further illustrates how architectural differences impact UDA for UAV-based road extraction.

Offshore oil platforms are critical infrastructures that require regular inspection to detect corrosion and maintain structural integrity. While traditional manual inspections are prone to human bias and high operational costs, recent advancements in automated inspection using 360-degree imagery have shown promise. This study presents a comprehensive evaluation of state-of-the-art metric monocular depth estimation methods—Depth Anything V2, ZoeDepth, Metric3Dv2, and Patchfusion—applied to 360-degree images of offshore oil platforms, a novel application in this domain. Metric depth estimation may also benefit downstream tasks such as corrosion and object detection by providing additional spatial context. Our comparative analysis assesses the performance and suitability of these methods in the context of the unique visual characteristics of offshore industrial environments and panoramic imagery. The findings offer valuable insights into the limitations and strengths of current approaches and serve as a basis for future work aimed at improving depth estimates, including domain-specific fine-tuning. This work contributes to ongoing efforts to enhance the efficiency, accuracy, and safety of structural health monitoring in challenging industrial settings. Code will be made publicly available upon acceptance.

Accurately measuring vegetation height is a key challenge in environmental monitoring, biomass estimation, and sustainable land

management. Recent advances in monocular depth estimation from RGB imagery have introduced models capable of predicting

canopy height without relying on costly LiDAR data. However, these models are typically trained on geographically constrained

datasets, raising concerns about their ability to generalize to different ecosystems and landscapes. In this study, we assess the

performance of a pre-trained monocular depth estimation model when applied to a dataset composed of natural forest areas and

a semi-urban environment characterized by a mixture of trees and built structures. Our evaluation reveals a significant degrada-

tion in accuracy when the model is used directly without any adaptation, highlighting the limitations of cross-domain transfer in

depth-based canopy height estimation. To address this issue, we perform a targeted fine-tuning using our dataset, which results in

considerable improvements across key metrics, including Mean Absolute Error (MAE), and Intersection over Union (IoU). These

findings demonstrate that even lightweight adaptation strategies are effective for tailoring monocular depth estimation models to

distinct environmental contexts, reinforcing the importance of local calibration for accurate and reliable canopy height mapping.The

source code used in this study is publicly available at: https://github.com/Geo99pro/depth-any-canopy.