Since 2011, the proliferation of brown macro-algae of the genus Sargassum has considerably increased in the North Tropical Atlantic Sea, all the way from the Gulf of Guinea to the Caribbean Sea and the Gulf of Mexico. The large amount of Sargassum aggregations in that area cause major beaching events, which have a significant impact on the local economy and the environment and are starting to present a real threat to public health. In such a context, it is crucial to collect spatial and temporal data of Sargassum aggregations to understand their dynamics and predict stranding. Lately, indexes based on satellite imagery such as the Maximum Chlorophyll Index (MCI) or the Alternative Floating Algae Index (AFAI), have been developed and used to detect these Sargassum aggregations. However, their accuracy is questionable as they tend to detect various non-Sargassum features. To overcome false positive detection biases encountered by the index-thresholding methods, we developed two new deep learning models specific for Sargassum detection based on an encoder–decoder convolutional neural network (CNN). One was tuned to spectral bands from the multispectral instrument (MSI) onboard Sentinel-2 satellites and the other to the Ocean and Land Colour Instrument (OLCI) onboard Sentinel-3 satellites. This specific new approach outperformed previous generalist deep learning models, such as ErisNet, UNet, and SegNet, in the detection of Sargassum from satellite images with the same training, with an F1-score of 0.88 using MSI images, and 0.76 using OLCI images. Indeed, the proposed CNN considered neighbor pixels, unlike ErisNet, and had fewer reduction levels than UNet and SegNet, allowing filiform objects such as Sargassum aggregations to be detected. Using both spectral and spatial features, it also yielded a better detection performance compared to algal index-based techniques. The CNN method proposed here recognizes new small aggregations that were previously undetected, provides more complete structures, and has a lower false-positive detection rate.

Object detection is a crucial task in both computer vision and remote sensing. The performance of object detectors can vary across different modalities depending on lighting and weather conditions. To address these challenges, we propose a fusion module based on contrastive learning and Gaussian cross-channel attention, called Gaussian Contrastive Cross-Channel Fusion (G3CF). We integrate this module into a dual-YOLO architecture, forming YOLO-G3CF. The contrastive loss enforces similarity between the features sent to the detection head from both modality branches, as they should lead to the same detections. The Gaussian attention mechanism enables the model to fuse features in a higher-dimensional space, enhancing discriminative power. Extensive experiments on VEDAI, GeoImageNet, VTUAV-det, and FLIR demonstrate that G3CF improves detection performance, achieving a mAP increase of up to 6.64% over the best single-modality baselines and outperforming prior multimodal fusion methods. Regarding model complexity, our fusion method operates at a late stage, increasing the computational cost of single-modality YOLO by approximately 150% in terms of GFLOPs. For instance, YOLOv8 requires 52.84 GFLOPs, whereas YOLOv8-G3CF, due to its dual architecture and three G3CF modules, increases this to 131.22 GFLOPs. However, a single G3CF module requires only ~ 15 GFLOPs. Despite this overhead, our approach remains computationally less expensive than transformer-based models, e.g., ICAFusion requires 284.80 GFLOPs. Moreover, the proposed method still operates in real-time, achieving ~ 19 FPS on an NVIDIA RTX 2080. The code is available at https://github.com/abelmouhcine/YOLO-G3CF.

Object detection has achieved great success in recent years; however, small-object detection remains a major challenge, particularly in remote sensing scenarios. Due to the large distance between the sensor and the scene, many objects appear extremely small and become indistinguishable after downsampling, which degrades detection performance. While avoiding downsampling and/or using high-resolution imagery can help, these solutions significantly increase computational cost at inference time. Indeed, when the resolution is increased by a factor of $\sigma$, the computational cost (in GFLOPs) grows approximately by a factor of ${\sigma }^{2}$. Similarly, if downsampling is avoided at a stage corresponding to a factor of $\sigma$, the computational cost of the network increases by a factor of ${\sigma }^{2}$ from that point onward. Thus, we propose a dual-training process in which one network performs early upsampling and the other performs late upsampling. Using a collaborative training strategy based on knowledge distillation, the early-upsampling network serves as a guide to improve the feature representations learned by the late-upsampling detector. Finally, only the late-upsampling model is used during inference, preserving efficiency while benefiting from the enhanced learning process. Experiments on six small-object detection remote sensing datasets (VEDAI, VTUAV, AITODv2, LEVIR-Ship, Sentinel-2 Ship Detection, and UAVSwarm) and one large object detection dataset (GeoImageNet) demonstrate the effectiveness of our approach. Our method consistently improves performance over the YOLOv8 baseline on small-object detection tasks, and outperforms several recent state-of-the-art methods, all while maintaining a reasonable computational cost (211 GFLOPs for early upsampling, compared to 120 GFLOPs for late upsampling for an input resolution of $512 \times 512$). The code will be made publicly available upon acceptance at https://github.com/abelmouhcine/YOLODistUP.

Object detection is a crucial task in both computer vision and remote sensing. The performance of object detectors can vary across different modalities depending on lighting and weather conditions. To address these challenges, we propose a fusion module based on contrastive learning and Gaussian cross-channel attention, called Gaussian Contrastive Cross-Channel Fusion (G3CF). We integrate this module into a dual-YOLO architecture, forming YOLO-G3CF. The contrastive loss enforces similarity between the features sent to the detection head from both modality branches, as they should lead to the same detections. The Gaussian attention mechanism enables the model to fuse features in a higher-dimensional space, enhancing discriminative power. Extensive experiments on VEDAI, GeoImageNet, VTUAV-det, and FLIR demonstrate that G3CF improves detection performance, achieving a mAP increase of up to 6.64% over the best single-modality baselines and outperforming prior multimodal fusion methods. Regarding model complexity, our fusion method operates at a late stage, increasing the computational cost of single-modality YOLO by approximately 150% in terms of GFLOPs. For instance, YOLOv8 requires 52.84 GFLOPs, whereas YOLOv8-G3CF, due to its dual architecture and three G3CF modules, increases this to 131.22 GFLOPs. However, a single G3CF module requires only ∼ 15 GFLOPs. Despite this overhead, our approach remains computationally less expensive than transformer-based models, e.g., ICAFusion requires 284.80 GFLOPs. Moreover, the proposed method still operates in realtime, achieving ∼ 19 FPS on an NVIDIA RTX 2080. The code is available at https://github.com/abelmouhcine/YOLO-G3CF .

Web page classification has many applications and plays a vital role in web mining and semantic web. Web pages contain much irrelevant information that does not reflect their categories or topics, and operates as noise in the process of their classification, especially when using a text classifier. Thus, the use of information from related web pages can help to overcome the problem of noisy content and to get a better result after the classification. Web pages are linked either directly by hyperlinks or indirectly by user’s intuitive judgment. In this work, we suggest a post classification corrective method that uses the query-log to build an implicit neighborhood, and collectively propagate classes over web pages of that neighborhood. This collective propagation helps improving text classifier results by correcting wrongly assigned categories. Our technique operates in four steps. In the first step, it builds a weighted graph called initial graph, whose vertices are web pages and edges are implicit links. In the second step, it uses a text classifier to determine classes of all web pages represented by vertices in the initial graph. In the third step, it constructs clusters of web pages using Formal Concept Analysis. Then, it applies a first adjustment of classes called Internal Propagation of Categories (IPC). In the final step, it performs a second adjustment of classes called External Propagation of Categories (EPC). This adjustment leads to significant improvements of results provided by the text classifier. We conduct our experiments using five classifiers: SVM (Support Vector Machine), NB (Naïve Bayes), KNN (K Nearest Neighbors), ICA (Iterative classification algorithm) based on SVM and ICA based on NB, on four subsets of ODP (Open Directory Project). We also compare our approach to Classification using Linked Neighborhood (CLN) considered as the closest algorithm to EPC. Results show that: (1) when applied after SVM, NB, KNN or ICA classification, IPC followed by EPC help bringing improvements on results. (2) F1 scores provided by our approach with any of the five classifiers are significantly better than those obtained by CLN. (3) The performance provided by our proposed approach grows proportionally to the size of the query-log, and to the density of the weighted graph.