Abstract
The number of daily small Unmanned Aircraft Systems (sUAS) operations in uncontrolled low-altitude airspace is expected to reach into the millions in a few years. Consequently, UAS density prediction has become an emerging and challenging problem. This paper presents a deep learning-based UAS instantaneous density prediction model that utilizes two types of input: historical density data and future sUAS mission information. The architecture consists of four components: a Historical Density Formulation module, a UAS Mission Translation module, a Mission Feature Extraction module, and a Density Map Projection module. Experimental results demonstrate outstanding performance, with the model achieving a correlation score of 0.947 in simplified scenarios and 0.823 in realistic scenarios involving A* routing and no-fly zones. The model significantly outperforms baseline approaches, improving prediction accuracy by over 15.2%.
The Challenge of Dynamic Airspace
As sUAS become ubiquitous for delivery and surveillance, managing low-altitude airspace becomes a complex data challenge. To ensure safety and efficiency, Air Traffic Management (UTM) systems must predict "hotspots" of high drone density. Unlike road traffic, which follows fixed lanes, drone paths are highly dynamic, changing based on temporary no-fly zones, new landing pads, or mission priorities. Existing models often rely solely on historical patterns, failing when the environment shifts. This research bridges that gap by incorporating the future mission schedule directly into the prediction engine.
Proposed Architecture
The proposed end-to-end deep learning architecture consists of four primary modules designed to fuse historical context with future intent:
- Historical Density Formulation: A CNN-based module that extracts features from the average traffic density of the previous 10 simulation cycles.
- UAS Mission Translation: This converts future missions (start, end, and launch time) into a 2D image using a novel "Flow" representation. Waypoints near the destination are rendered brighter than those near the start to help the model distinguish travel direction.
- Mission Feature Extraction: The model uses channel segmentation to group missions into long-term, mid-term, and short-term categories. It employs 2D convolutions to extract features from these groups independently, learning the distinct impact of missions based on their launch proximity.
- Density Map Projection: A deconvolutional (transposed convolution) layer fuses historical and mission features to project a final 2D density map for a specific future time window.
Data Generation and Simulation
Due to the lack of large-scale real-world UAS traffic data, the researchers developed a Python-based simulator inspired by the Multi-Agent Traffic Resource Usage Simulator (MATRUS) framework. The simulation environment uses a 100x100 grid representing airspace. The dataset comprises 3,000 samples, covering both simplified traffic (straight-line paths) and realistic traffic (using A* routing to avoid no-fly zones and maintain safe distances).
Key Insights: Why 2D Convolutions?
A significant finding of this study is that 2D convolutions on segmented time channels outperformed 3D convolutions. While 3D filters are often used for spatio-temporal data, they tended to "blur" features across the entire time window in this application. In contrast, 2D filters on segmented channels (short, mid, and long-term) preserved the distinct impact of missions launching at different times, leading to higher accuracy.
Experimental Results
The model demonstrates robust performance across various metrics:
- Accuracy Boost: In simplified scenarios, the model achieved a correlation score of 0.947, a 15.2% improvement over baseline models like Vanilla CNNs or LSTMs.
- Hotspot Prediction: The model is exceptionally effective at identifying congestion, achieving an AUROC of 0.951 for hotspot prediction (identifying the top 1% densest areas).
- Realistic Scenarios: Even with complex A* routing and no-fly zones, the model maintained a high correlation score of 0.823, significantly outperforming Vanilla LSTMs which struggled with long-term dependencies and accumulated errors.
Conclusion and Future Work
This mission-aware approach proves that knowing future intent is more predictive than relying on history alone in dynamic environments. Future work will focus on generalizing the model to handle diverse trajectory planning algorithms and exploring performance in environments with "non-cooperative" drones that do not share mission lists.
Summary updated