The original release of YOLO-MOD already offered something genuinely useful: a general-purpose framework for running YOLO object detection models inside standard QGIS workflows. Users could perform multi-class object detection using models in both PyTorch and ONNX formats, visualize the results as georeferenced vector layers overlaid on their maps, and export annotated datasets in the YOLO format for training future models. Basic post-processing utilities, such as layer merging, canvas tiling, and previewing exported annotations, rounded out an end-to-end workflow. But there was a significant constraint. The first version could only detect and classify five predefined object categories: ships, aircraft, helicopters, airports, and storage tanks. Anyone wanting to detect something else, from solar panels to wind turbines, was simply out of luck.

The updated version 2.1.1 removes that ceiling entirely. Users can now load and apply their own trained Ultralytics YOLO models, compatible with YOLOv8 and later architectures, with an arbitrary number and type of object classes. The technical mechanism behind this is elegantly simple. When a model stored in the PyTorch .pt format is loaded through the plugin’s interface, YOLO-MOD automatically extracts the object class names embedded in the model and dynamically generates the appropriate configuration controls in the user interface for each class. This means researchers can configure visualization and processing options for their custom classes exactly as they would for the built-in standard models, without ever touching a line of plugin code. The same workflow now applies to virtually any detection task based on horizontal bounding boxes.

The implications for applied remote sensing are considerable. A coastal monitoring team could train a model on harbor infrastructure and vessel types specific to their region, load it into QGIS, and immediately begin scanning aerial imagery for those objects. An airport safety analyst could build a detector for ground support equipment. An environmental agency could train a network to flag illegal dumping sites or deforestation edges. In each case, the results appear directly as color-coded bounding boxes on georeferenced layers, superimposed on basemaps such as Bing Aerial, meaning detections land in the same coordinate system as every other dataset in the project. That integration eliminates the clumsy export-import cycles that have long characterized deep learning workflows in the geospatial domain.

The update also tackles one of the most practical frustrations of working with automated detection: false positives. A new Edit Detections tab extends the plugin’s post-processing capabilities with interactive tools for removing individual detections from an existing vector layer. Users can select one or more detected objects, highlight them visually on the map, and permanently delete them without rerunning the entire detection process. This matters more than it might sound. In operational settings, analysts routinely spend hours cleaning up automated results before they can be used, and every detection run that must be repeated from scratch wastes computational resources and time. Being able to surgically remove spurious detections within QGIS itself keeps the analytical workflow inside a single environment.

The editing workflow is designed to feel native to QGIS users. Analysts select features using the familiar QGIS Selection Toolbar, for instance with the Select Features by Area or Single Click tool, and the plugin’s Selection Editing panel reports how many features are ready to be removed with a single button press. The layer merging functionality from the previous version has been retained and enhanced alongside these new tools, so refining a large multi-part detection result remains straightforward. The developers describe this as a way to let detection results be refined before further analysis or export, which is precisely the point in the pipeline where most real-world projects stall.

The third major enhancement concerns training data preparation, often the most labor-intensive phase of any machine learning project in remote sensing. The Canvas Tiling tool, which automatically splits the current QGIS map extent into image tiles suitable for training object detection models, has been upgraded with an interactive preview of the tile grid before any files are generated. Users configure tile width and height, the overlap percentage between adjacent tiles, the padding color for regions outside the available map content, and the output directory. Rather than committing immediately, a Preview and Start Tiling button first displays the proposed grid, letting the user verify the layout and then confirm or cancel the operation. Areas beyond the map content are automatically padded in the user-defined color, white by default, ensuring that every generated tile has identical dimensions regardless of the shape or size of the selected extent.

Under the hood, the plugin is built entirely in Python and designed for QGIS 3.x, with testing reported on versions 3.40 and 3.42 under Windows 11 using the embedded Python environment. Dependencies include PyTorch, with CUDA-enabled builds recommended for GPU acceleration, ONNX Runtime for flexible CPU and GPU inference, and NumPy. The software is distributed under the GNU General Public License version 3.0, versioned with Git, and publicly available through its GitHub repository, keeping it fully open source. That licensing choice matters for public agencies and research groups in developing regions, for whom commercial geospatial AI platforms can be prohibitively expensive and whose data policies sometimes forbid cloud-based processing altogether.

The broader context is a field moving fast. Object detection in remote sensing has been propelled by the growing availability of high-resolution imagery and by advances in detection algorithms since the original You Only Look Once architecture was introduced in 2016. Yet although several QGIS plugins now support selected deep learning tasks, the developers note that flexible frameworks for multi-category object detection remain relatively limited. YOLO-MOD’s approach of dynamically adapting its interface to whatever model the user supplies is a notable contribution to closing that gap, particularly for teams that lack the software engineering resources to build custom integrations.

What makes this release interesting beyond its immediate feature list is what it signals about the direction of geospatial software. The division between the GIS, where spatial data lives, and the machine learning environment, where models run, has been eroding for years, and tools like YOLO-MOD represent the pragmatic middle path: rather than rebuilding GIS capabilities inside a deep learning framework, they embed the models directly into the analyst’s existing environment. For the growing community of researchers tracking ships in shipping lanes, aircraft at airfields, and infrastructure across vast landscapes, the ability to load any YOLOv8 or later model, clean up its outputs interactively, and tile fresh training imagery from the same map canvas constitutes a genuinely streamlined loop, one that could shorten the cycle from data collection to working detector from weeks to days.

Subject of Research: A YOLO-based multi-category object detection plugin for optical remote sensing imagery integrated into the QGIS geographic information system.

Article Title: Version [ 2.1.1 ] – [ YOLO-MOD: YOLO-based multi-category object detection plugin for maritime, aviation, and infrastructure objects in optical remote sensing images in QGIS ]

Article References: Ciecholewski, M., & Strzelecki, M. (2026). Version [2.1.1]- [YOLO-MOD: YOLO-based multi-category object detection plugin for maritime, aviation, and infrastructure objects in optical remote sensing images in QGIS]. SoftwareX, 35, Article 103002. https://doi.org/10.1016/j.softx.2026.103002

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103002

Keywords: YOLO-MOD, object detection, QGIS, remote sensing, deep learning, YOLOv8, geographic information systems, Ultralytics, satellite imagery, maritime detection, infrastructure detection, open-source software

At its core, MEBS represents a leap forward by incorporating adaptive capabilities that enable it to recognize and segment images where traditional methods fall short. Most existing segmentation tools apply fixed rules or assumptions about data characteristics, often tailored for ideal or Gaussian noise environments. However, many real-world images, such as medical scans or satellite captures, contain non-Gaussian noise and irregular patterns that stymie conventional approaches. The novelty of MEBS lies in its ability to dynamically adjust to these atypical features, improving detection fidelity without manual recalibration.

The underlying mathematical principles of MEBS are rooted in energy-based models combined with multiple change point detection techniques. This synergy allows the system to autonomously pinpoint shifts in image properties, such as intensity or texture, that signify boundaries or regions of interest. By modeling these shifts as change points, MEBS segments images more accurately, especially when dealing with subtle or ambiguous structures often masked by noise. This approach is particularly important for medical imaging, where precise delineation of tumours or fluid accumulations can critically affect diagnostic outcomes.

In practical applications, MEBS’s adaptive segmentation capability enables healthcare providers to detect abnormalities in X-rays and mammograms with enhanced clarity. The model’s sensitivity to nuanced changes translates to earlier and more reliable identification of tumours and pathological fluid buildups. This advancement stands to significantly augment diagnostic workflows by reducing false negatives and enabling more targeted treatment planning, ultimately improving patient outcomes.

Environmental monitoring similarly benefits from the precision of MEBS. Wildfire management, a pressing concern exacerbated by climate change, demands rapid detection of nascent hotspots to mobilize containment efforts effectively. The adaptive model’s facility to parse satellite images laden with atmospheric noise allows it to detect small yet critical ignition points with unprecedented speed. Such capability promises to revolutionize how wildfire data is processed and applied in real-time crisis management.

Beyond health and environmental science, MEBS also offers substantial utility in biological research, particularly in plant biology and agricultural domains. Accurately counting and tracking cellular growth patterns is essential for understanding developmental processes and optimizing crop yields. Traditional imaging tools frequently struggle with cell segmentation when confronted with variable lighting or heterogeneous tissue samples. MEBS’s energy-based adaptive segmentation provides robust solutions to these challenges, enabling researchers to gather precise data that informs genetic and agronomic advancements.

This innovative technology’s development was driven by a dedicated team of UBCO students — including lead author Jiatao Zhong, along with Shiyin Du, Canruo Shen, Yiting Chen, Medha Naidu, and Min Gao — who collaboratively undertook the tasks of coding, experimentation, and validation. The students’ contributions showcase the synergy between academic mentorship and student initiative, providing a practical learning environment that bridges theoretical mathematics and applied data science.

The research effort was also bolstered by collaboration with Dr. Yuejiao Fu, further enriching the multidisciplinary nature of the project. Together, the team rigorously tested MEBS across various datasets representing real-world complexities to validate its performance gains over existing segmentation techniques. This comprehensive evaluation underscores the model’s versatility and adaptability in different domains.

The significance of MEBS lies not only in its academic novelty but also in its practical implications. Automatic adaptation to the inherent irregularities of images eliminates the need for extensive manual tuning, which is often time-consuming and prone to human error. This feature facilitates scalable application across industries where data volume, diversity, and quality vary widely, from hospitals to space agencies.

Funded by the Natural Sciences and Engineering Research Council of Canada and UBC Okanagan’s Vice-Principal, Research and Innovation office, the project exemplifies the vital role of institutional support in driving frontier scientific research. The outcomes pave the way for future explorations into energy-based methods and adaptive algorithms that can further elevate the capabilities of image processing technologies.

Published in the esteemed journal Scientific Reports in July 2025, the MEBS study not only pushes forward the boundaries of image segmentation but also resonates with a broader scientific community eager for solutions to complex pattern recognition problems. It reflects an exciting intersection of applied mathematics, computer science, and environmental and health sciences that is set to inspire subsequent innovations.

MEBS stands as a testament to how interdisciplinary collaboration and advanced mathematical modeling can produce tools with profound real-world impact, providing a new lens through which scientists and practitioners can extract meaningful insights from the most challenging visual data.

Subject of Research: Not applicable

Article Title: Energy-based segmentation methods for images with non-Gaussian noise

News Publication Date: 16-Jul-2025

Web References:
https://www.nature.com/articles/s41598-025-09211-8

References:
DOI: 10.1038/s41598-025-09211-8

Keywords:
Complex analysis, Computer science, Applied physics, Applied mathematics, Energy resources, Industrial science, Information science, Network science, Technology