Neural Networks for Detection of Last-Mile Warehouse Properties from Satallite Data
Client: Private Equity Logistics Fund
Neural Networks for Detection of Last-Mile Warehouse Properties from Satallite Data
Using YOLO and SAM2 to automate detection and segmentation of last-mile logistics warehouses from satellite imagery.
Business Case
Private equity investors increasingly target last-mile warehouses due to their high yield potential, yet identifying unlisted properties within strategic locations is highly resource-intensive. To gain a competitive edge, the client required a systematic approach to scout and analyze logistics hubs across all of Germany. Specifically, they needed to pinpoint facilities located within a critical 30-minute drive time of major population centers to assess their viability for high-return investments.
Outcome
We built an end-to-end geospatial intelligence pipeline that scanned and analyzed satellite imagery across the entirety of Germany. By deploying YOLO and SAM2 models, the system successfully detected, categorized, and segmented logistics properties at scale. This automated pipeline generated a comprehensive, data-driven investment report, transforming raw imagery into a high-value pipeline of actionable, off-market real estate opportunities for the client.
Detailed Report
What is Last-Mile Logistics?
The last mile represents the final, most critical leg of the supply chain the journey of a product from a local fulfillment hub or sorting facility to its final destination (usually a consumer’s doorstep or a retail store).
Though it spans the shortest physical distance, it is the most complex, costly, and operationally intensive part of the entire logistics cycle. In fact, studies show that last-mile delivery routinely accounts for up to 41% to 53% of total supply chain costs (see Capgemini Research Institute titled “The Last-Mile Delivery Challenge.”).
To combat this friction, the modern supply chain relies heavily on last-mile warehouses (or urban micro-fulfillment centers). These are strategically positioned distribution nodes located within a tight, high-density radiustypically defined by institutional investors as a 5-to-15 km radius or a 30-minute drive time of major population centers.
Toolset & Techstack
OpenStreetMap (OSM)
OpenStreetMap provides the foundational and open source geospatial data layer. It is used to map road networks and calculate the precise isochrones for the 30-minute drive-time zones around population centers, isolating the exact geographical boundaries where satellite imagery needs to be scanned. You can visit their sites here.
CVAT (Computer Vision Annotation Tool)
CVAT is the open source training data engine. Annotators use it to draw precise bounding boxes and masks over raw satellite tiles, labeling key logistics features like loading docks and large parking lots to create the high-quality ground truth needed to train the AI models. You can visit their github here.
YOLO (You Only Look Once)
YOLO acts as the rapid object detection engine. It scans massive, high-resolution satellite swaths across Germany in a single pass, instantly identifying and drawing bounding boxes around potential commercial buildings while ignoring irrelevant terrain like fields or forests. You can visit their github here.
SAM2 (Segment Anything Model 2)
SAM2 handles pixel-perfect image segmentation. Taking the initial locations flagged by YOLO, this foundation model precisely traces the exact boundaries of a warehouse footprint, allowing the system to accurately calculate building square footage for the investment report. You can visit the github pae here.