TileNet: How a Lean CNN-SVM Hybrid Is Bringing Real-Time Defect Detection to Flat-Roof Drone Inspections

Main Story

Flat roofs are not passive building elements. They govern both structural performance and thermal efficiency, contributing directly to household energy consumption, carbon emissions, and long-term environmental sustainability. Yet their inspection has remained stubbornly manual, labour-intensive, and hazardous — inspectors must walk surfaces prone to hidden moisture damage and structural weakness. A new paper from the Department of Mechanical and Aerospace Engineering at Carleton University, authored by Samuel Dunthorne and Hashim A. Hashim, proposes a machine-learning-first alternative.

Posted to arXiv on 11 September 2026 under the cs.CV and cs.AI classifications, and submitted to the Journal of Safety Science and Resilience, the paper introduces TileNet: a real-time, UAS-based deep learning framework that autonomously detects defects using live imagery captured during dual-altitude aerial passes.

The core engineering problem the team had to solve is familiar to anyone working on embedded aerial intelligence: existing lightweight CNN detectors deployed on UAV platforms struggle with small object recognition, while standard, more capable architectures suffer from computational complexity that prohibits real-time inference on embedded hardware. TileNet's answer is a deliberate architectural trade-off — substitute model bulk for spatial decomposition.

The Tiling Strategy

Rather than feeding full-resolution roof images into a heavy classifier, TileNet partitions each frame into smaller tiles and classifies each tile independently. This keeps individual inference tasks computationally bounded, enabling the system to process 130 images per second on the onboard hardware. The approach is reinforced by the dual-altitude flight protocol: the multi-resolution flight strategy is designed to aid the identification of both small, fine-scale defects and larger structural issues, enabling more comprehensive assessments. Low-altitude passes surface granular membrane cracks and blistering; higher-altitude passes capture macro-level deformation and drainage failure zones.

The Architecture

The classifier itself is a custom network — not a fine-tuned transfer-learning derivative — comprising five convolutional layers and four dense layers, with a linear SVM head replacing the conventional softmax output. Substituting the final activation with a Support Vector Machine is a well-established technique for improving margin-based separation on the feature embeddings generated by the convolutional backbone; it has been explored in embedded FPGA contexts precisely because CNN-SVM hybrid algorithms are widely applied in image processing tasks and can be mapped efficiently to constrained compute resources.

The dataset on which TileNet was trained and evaluated is substantial: 43,383 training images, 3,869 validation images, and 2,540 test images — all tiled and augmented from real UAS footage collected during onsite visits. The model achieved a mean test accuracy of 94.4% (95% confidence interval ±0.4% across three seeds). Against established baselines, the gap is significant: GoogLeNet reached 89.2% and AlexNet 79.8% on the same split, with TileNet outperforming both while carrying a fraction of the parameter count those architectures require.

The Platform

Field data collection used the DJI Matrice 350 RTK, DJI's current flagship enterprise platform. The Matrice 350 RTK features an all-new video transmission system, a more efficient battery system, more comprehensive safety features, and robust payload and expansion capabilities. Key hardware characteristics relevant to inspection missions: the platform offers up to 55 minutes of flight time, an IP55 environmental protection rating, and an operating temperature range of -20°C to 50°C. On payload capacity, with a max payload capacity of 2.7 kg, the aircraft can carry up to three payloads simultaneously, meeting the needs of different operation scenarios like public safety, inspection, and mapping. Transmission is handled by DJI O3 Enterprise, which supports triple-channel 1080p HD live feeds and a maximum transmission distance of 20 km — a transmission margin far exceeding what a building-perimeter roof survey would demand, providing headroom for data-dense onboard inference streaming.

The Matrice 350 RTK also supports a broad sensor ecosystem: the four-sensor system of zoom camera, wide camera, thermal camera, and laser rangefinder works together, supplemented by advanced intelligent algorithms, to deliver sensing and imaging performance — a combination well suited to the multi-modal defect signatures (visual cracking, thermal moisture ingress, surface deformation) that flat-roof inspections need to resolve.

Why It Matters Beyond the Lab

Timely detection of roof defects is essential for reducing heating and cooling losses, preventing moisture-driven degradation such as mold growth, and supporting national climate-change mitigation goals. The paper situates TileNet within this building-performance framing, making the case that inspection frequency — currently constrained by cost and human risk — is a direct lever on a building's operational carbon footprint.

That framing aligns with a market actively transitioning from manual to autonomous methods. Industry studies confirm that UAV-based inspections reduce inspection time by up to 70% and lower costs by 40–60% compared to traditional manual inspections. The drone roof inspection segment is on a corresponding growth trajectory: the drone roof inspection market was valued at USD 1.2 billion in 2025 and is estimated to reach USD 3.5 billion by 2034, exhibiting a CAGR of 12.9%.

Technical Breakdown

Industry Impact

For inspection service operators and roofing contractors: TileNet's 130-image-per-second onboard throughput means defect maps can be generated and reviewed on the same site visit, compressing the feedback loop from days (off-site photogrammetric processing) to minutes. With repeatable flight paths and the right payloads, UAVs document membranes, penetrations, parapets, skylights, and rooftop equipment without disrupting building operations. A validated onboard AI layer removes the bottleneck of post-flight manual image review.

For building managers and asset owners: Building operators prioritise roof condition monitoring to reduce repair costs and extend asset life. A system that can be deployed frequently — rather than annually because of cost — shifts maintenance from reactive to genuinely predictive. The thermal and moisture-detection capability of the Matrice 350 RTK sensor suite, paired with TileNet's tile-level defect localisation, supports the granular condition records that insurance underwriting and ESG reporting increasingly demand.

For AI and edge-compute integrators: TileNet's custom five-layer architecture — purpose-built rather than transferred from ImageNet-scale models — illustrates that bespoke lightweight design can beat larger incumbent architectures on narrow, well-defined inspection tasks. The deployment of CNNs on UAV platforms has evolved from direct porting of standard architectures to the development of highly specialised, hardware-aware models — TileNet fits squarely in that trajectory. Integrators building inspection software stacks should note the SVM head as a practical mechanism for hardening classification boundaries without adding convolutional depth.

For the broader drone inspection market: The adoption of AI-powered drone inspection, autonomous UAV monitoring systems, and predictive maintenance analytics is enabling faster data collection, reduced operational risks, and improved inspection accuracy. Papers like TileNet provide validated, reproducible benchmarks that give enterprise buyers and insurers a concrete accuracy baseline to demand from commercial inspection vendors — raising the floor for the entire sector.

For regulators and standards bodies: The paper's submission to the Journal of Safety Science and Resilience signals intent to influence building maintenance standards, not just academic discourse. As autonomous UAS inspection moves toward regulatory normalisation under frameworks like FAA BVLOS waivers and EASA's specific category, peer-reviewed accuracy benchmarks tied to named production platforms become part of the compliance evidence stack that operators must assemble.