Conference Agenda

.

Synthetic Aperture Radar (SAR) super resolution (SR) is a tough challenge for various reasons. The application of machine learning (ML) algorithms for SAR-SR has been an active field of research in recent times. ML models must learn the mapping between low- and high-resolution acquisitions of the same scene. Due to physical properties of SAR images, namely speckle statistics, resolution-coherence coupling, and cross-sensor radiometric differences, synthetic dataset generation cannot fully capture the real sensor behavior. This stresses the need for empirical heterogeneous pairs so that the models generalize correctly to operational sensor conditions. Coregistration errors are inevitable because methods like tie-point matching and geometric warp estimation are imperfect, yet even minor misregistration causes pixel-wise loss functions to fail during training.

Beyond resolution enhancement, the ultimate goal of SAR-SR is to improve performance in downstream applications. To achieve this, we propose shifting the loss computation away from the pixel domain. We are developing a set of SAR SR algorithms that uses application-specific error metrics to train the ML architectures. These methods, which we term as Application-Specific Super-Resolution for SAR (ASSURE-SAR), would also alleviate the general challenges related to SAR-SR (such as misregistration and speckle noise).

The following are the two flavors of ASSURE-SAR that we are currently investigating. However, ASSURE-SAR is a generic framework, and we plan to investigate a family of algorithms to achieve application-specific SR algorithms for SAR images.

(1) The first formulation computes the loss in the compressed domain. Both the output of the SR model and the high-resolution reference are passed through a shared compression operator C(·), such as the Discrete Fourier Transform (DFT), before the error is calculated. The compression operators used in our work are shift-invariant, meaning that misregistrations between training pairs no longer cause a problem.

(2) The second formulation computes the loss in the detection domain. Both the output of the SR model and the high-resolution reference are passed through a pipeline containing ROI segmentation stage and a pretrained object detector, i.e. SAR YOLO. The resulting detection outputs are then compared to compute the error that the training objective minimizes. Since the detector operates on segmented object patches, the loss is robust to small geometric offsets.

Both formulations allow the model to be optimized for perceptual and semantic quality, even in the presence of the coregistration uncertainty inherent to real-world paired acquisitions.

The increasing availability of high-resolution spaceborne Synthetic Aperture Radar (SAR) data is enabling new capabilities for monitoring dynamic activities across maritime and terrestrial environments. Unlike optical sensors, SAR provides consistent observation capabilities independent of daylight and weather conditions, making it particularly valuable for security, defense, and geospatial intelligence applications. However, the effective operational exploitation of SAR imagery requires advanced processing techniques capable of robust object detection, classification, and monitoring under complex imaging conditions and across large volumes of data.

SATIM specialises in advanced geospatial intelligence (GEOINT) solutions, leveraging proprietary Artificial Intelligence (AI) and Synthetic Aperture Radar (SAR) analysis to deliver actionable insights across defense, security, and commercial sectors. This presentation introduces SATIM’s approach to automated detection and classification of objects in SAR imagery through SATIM’s OREC (Object Recognition and Classification) software, a purpose-built solution for automated SAR imagery analysis.

OREC is an AI-based software solution specifically engineered for object detection and classification. It possesses a vast library covering approximately 300 classes of military-relevant objects, including naval vessels (ships), grounded aircraft, and various ground vehicles. This comprehensive classification ability allows for rapid and precise identification of key assets in diverse environments, regardless of weather or time of day, a critical advantage offered by SAR technology.

The robust performance of the OREC software is achieved through a meticulous development process that combines synthetic SAR signatures with extensive real-world data. This hybrid training methodology ensures a resilient and accurate solution that is both data and platform agnostic, meaning it can ingest and process imagery from virtually any SAR satellite or aerial sensor. Quantitatively, OREC demonstrates exceptional reliability, boasting F1 scores for both detection and classification consistently above 90%.

Beyond automated object recognition, SATIM's capabilities extend to sophisticated change detection, persistent monitoring, and large-area surveillance. By transforming complex SAR data into immediate, understandable intelligence, SATIM empowers operators and analysts to achieve superior situational awareness, streamline monitoring workflows, and make critical decisions with greater speed and confidence.

We present advanced AI-based feature detection systems utilizing Deep Neural Networks for marine vessel and aircraft detection on high-resolution Synthetic Aperture Radar (SAR) imagery, developed through the SEA-SPARK-ADVANCED AI/ML VESSEL DETECTION AND CLASSIFICATION ON HIGH-RES SAR IMAGERY (4000137472/22/I-DT-lr) and ADVANCED AI BASED FEATURE DETECTION FOR SECURITY APPLICATIONS (4000139425/22/I-DT) projects. These results stand out as all-weather, day-and-night complementary automated detection systems to electro-optical sensors, which frequently suffer from low visual contrast in conditions such as snow cover.

For maritime domain awareness, we developed SAR segmentation models capable of detecting vessels ranging from 10 meters to the largest operating ships in diverse open sea and port environments. Through rigorous training and the cross-validation of 24 models against a Constant False Alarm Rate (CFAR) baseline, we demonstrated the superior performance stability of our DNN models and identified specific edge cases for both methods. To ensure segmentation robustness, we constructed a highly balanced ICEYE dataset distributed across diverse satellite types and incidence angles. Additionally, we identified the detection of "dark ships" with disabled AIS as a highly viable extension into multi-modal analytics.

For the aviation domain, we benchmarked Convolutional and Vision Transformer architectures on manually curated datasets from both ICEYE and Capella Space providers. We successfully trained a "provider-agnostic" model by combining these datasets, which further improved final detection performance. During aircraft model evaluations, we identified persistent difficulties in segmenting closely grouped objects (parked in blocks) and handling unusual formations. Furthermore, high-backscatter ground support equipment and shelters occasionally triggered false positive identifications. During our research, we also found that addressing data scarcity through synthetic SAR data generation remains a highly complex task, as none of the existing market solutions offered a sufficient technical readiness level.

Ultimately, we integrated these robust algorithms into the existing SpaceKnow Guardian analytics platform for persistent monitoring scenarios, where the functionality was successfully tested and validated by pilot users.

Artificial Intelligence–based Automatic Target Recognition (ATR) from Synthetic Aperture Radar (SAR) imagery is increasingly adopted in security and situational-awareness applications. However, the performance and generalization capability of machine learning models remain strongly dependent on the availability, diversity, and quality of labeled training data. In many operational scenarios, collecting sufficiently large annotated datasets is difficult, particularly for rare targets or specific acquisition geometries.

This work presents an integrated framework for SAR ATR that combines simulation, data management, annotation, model training, and inference within a unified processing environment. The framework includes a SAR image simulation module capable of generating physically consistent synthetic data based on electromagnetic scattering models, a labeling environment for efficient annotation of SAR imagery, a training module for machine learning model development, a centralized database for managing imagery and metadata, and an inference pipeline for operational deployment.

Particular emphasis is placed on the use of convolutional neural networks supporting oriented bounding boxes, which are well suited to represent the arbitrary orientation of ships and other objects in SAR imagery. The integration of simulated and real SAR data enables systematic experimentation and improves model robustness in situations where labeled data are limited. The simulation component further allows controlled variation of acquisition geometry and target conditions, including the modelling of ship motion and associated defocusing effects that frequently occur in maritime SAR acquisitions.

The framework is demonstrated on a case study involving naval targets observed in very high resolution SAR imagery. Results highlight how the combination of simulation-assisted data generation and AI-based detection approaches can support more scalable ATR development workflows. The proposed environment provides a flexible platform for advancing SAR exploitation capabilities and accelerating the adoption of AI-enabled SAR applications in the security domain.