Jake Rizzo

Cadet, United States Military Academy at West Point
“A Satellite Image Segmentation Pipeline to Inform Autonomous Aerial Search Strategies”

Session Materials: Link

Session Recording: Link

Speaker Bio: 

I am a Second-Class Cadet at the United States Military Academy at West Point, where I am pursuing a degree in Operations Research within the Department of Mathematical Sciences. Beyond the classroom, I have academic experience working with special operations components to develop and integrate autonomous systems solutions with military applications. 

Abstract: 

Unmanned aerial systems (UAS) have rapidly evolved into critical assets for modern military and reconnaissance operations, offering the ability to operate at long ranges with reduced risk to personnel. However, the efficacy of autonomous search missions is often limited by a reliance on predefined, geometric flight paths such as spirals or lawnmower patterns that fail to account for the complexities of the underlying terrain. This paper presents the development of a novel pipeline that generates autonomous, probabilistic terrain maps to seed target belief management algorithms, used for target acquisition and search persistence in non-permissive
environments. The primary success of this research lies in the creation of a hybrid segmentation model that outperforms traditional deep learning architectures in terms of deployability and integration. While Convolutional Neural Networks (CNNs) and Transformers like SegFormer often require massive, labeled datasets and significant computational overhead, our approach integrates OpenStreetMap (OSM) data with the lightweight, probabilistic vegetation estimates of the Detectree algorithm. This combination allows for the generation of pixel-wise probability matrices that characterize terrain without the need for extensive retraining or specialized hardware. By avoiding rigid classification masks, the pipeline assigns probability vectors to every pixel, enabling the system to distinguish between high-belief areas relative to the expected behavior of a specific target class. This probabilistic output was successfully integrated into a target belief management algorithm, providing a nuanced input for belief initialization. A key feature of this work is the ability to shape initial belief based on mission-specific
parameters, weighting the search distribution according to the probability of different target types such as vehicles, hikers, or boats occupying specific terrain features. This ensures that the initial particle distribution reflects both environmental feasibility and tactical reality, providing broad flexibility to a range of different search missions. The evaluation of this pipeline demonstrates a superior balance between segmentation quality and computational feasibility. Ultimately, this work delivers a functional, rapidly deployable system that transforms raw satellite imagery into actionable, probabilistic data, providing a foundational capability for the next generation of context-aware autonomous search strategies.