TileNet: Tile-Based CNN-SVM Architecture for Autonomous Unmanned Aerial Systems Inspection of Flat Roofs

AI summary

Overview

• Research area: Computer vision and embedded machine learning for unmanned aerial systems (UAS), applied to automated building-envelope inspection — specifically flat-roof defect detection.
• Technical level: Intermediate. The paper assumes familiarity with convolutional neural networks, support vector machines (SVMs), and embedded computing constraints, but the concepts are explained accessibly enough for an applied engineering audience.
• Scope: The paper proposes and validates TileNet, a tile-based CNN-SVM classifier that performs binary defect/no-defect classification on aerial flat-roof imagery using low-power onboard hardware.

What This Paper Is About

Flat roofs govern building energy performance and are prone to defects that cause moisture ingress, mold growth, and heat loss, but inspecting them manually is dangerous and expensive. The authors build a deep learning system that runs directly on a drone to autonomously flag roof defects from live imagery, engineered specifically to fit the power, weight, and processing limits of embedded UAS hardware. The core tension the paper addresses is that accurate CNNs usually demand more computation than a drone can carry, so the authors split high-resolution images into tiles and pair a small CNN with an SVM classifier head.

Key Contributions

1. A tile-based CNN-SVM architecture that splits high-resolution aerial images into 640×360-pixel tiles, preserving fine defect detail while cutting feature-map size and computational load enough for near-real-time inference on embedded UAS hardware.
2. A computation-aware model design with tuned hyperparameters and reduced feature-map dimensionality, achieving over 92% test accuracy while remaining deployable on low-power processors.
3. A multi-resolution, dual-altitude UAS imaging strategy — a high-altitude pass to capture large structural problems and a low-altitude pass for fine-grained defects — adopted as a design rationale, though the authors acknowledge further work is needed to isolate its individual benefit.
4. An end-to-end operational pipeline, validated with real imagery collected using a DJI Matrice 350 RTK drone (Zenmuse H30T camera) during onsite roofing visits, including a public demonstration video.

Main Findings

• Test accuracy of 94.4%: The final model achieved a mean test accuracy of 94.4% with a 95% confidence interval of ±0.4% across three random seeds, evaluated on a photo-level split of 43,383 training, 3,869 validation, and 2,540 test tiled and augmented images.
• Outperforms established architectures: TileNet exceeded GoogLeNet (89.2%) and AlexNet (79.8%) on the same classification task.
• Optimizer choice was decisive: Early training plateaued below 70% accuracy under the Adam optimizer. Switching to Stochastic Gradient Descent with a fixed learning rate of 0.005 and momentum of 0.2 pushed accuracy above 90% and stopped early-stopping callbacks from triggering prematurely.
• Squared-hinge SVM head replaces softmax: The final dense layer uses a linear squared-hinge L2-SVM head rather than a conventional sigmoid cross-entropy output, assigning the defective class when the SVM decision value is non-negative.
• Tiling preserves information while reducing cost: A 3,840×2,160-pixel source image splits evenly into 36 tiles, keeping total pixel information intact while shrinking feature maps and training times.
• Capture rate can be modest: With 80% image overlap, capture rates of 0.5–1.4 images per second proved sufficient, since roof surfaces are static and overlapping frames guarantee defect visibility across neighboring images.
• Environmental ambiguity is a real problem: Damp membrane from recent rainfall looks nearly identical in RGB imagery to prolonged standing water caused by structural failure, which the authors flag as a persistent source of false positives.
• Dataset scope is deliberately narrow: All imagery covers modified-bitumen (mod-bit) membranes; generalization to TPO, EPDM, or PVC systems is explicitly not claimed.

Methodology in Plain English

The researchers assembled a custom dataset because no labeled rooftop defect dataset existed publicly. They combined sparse online imagery with high-resolution photos captured by a drone flown over real roofs, then cropped everything to a standard 3,840×2,160-pixel format and sliced each image into 640×360-pixel tiles. A single annotator with five years of roofing-company experience labeled every tile as either containing a defect or not, consulting in-house roofing experts on ambiguous cases. Defect categories included surface cracking, alligatoring, granular loss, blistering, exposed roof deck, moisture staining, failed seams, and membrane seam separation.

To expand the dataset, they applied standard augmentations — random rotation, flipping, brightness variation, and noise injection — pairing each original image with an augmented twin. They deliberately excluded images dominated by irrelevant objects like brick walls, parked vehicles, wooden decks, and mechanical units to avoid confusing the model during early training.

The classifier itself is a compact CNN with five convolutional layers and four dense layers. The first dense layers use ReLU activations with L2 regularization, and the final layer is a linear SVM head trained with squared-hinge loss. Images enter at 640×360, get resized to 180×320, augmented, then passed through convolution and max-pooling stages before flattening. Training initially stalled below 70% accuracy until the authors swapped Adam for SGD, after which performance climbed past 90%. The deployed workflow has the drone fly two passes at different altitudes, capture overlapping images at 1–2 frames per second, buffer them, tile them, and classify each tile sequentially.

Why This Matters

Impact on research: The paper demonstrates that pairing image tiling with an SVM output head can deliver competitive accuracy on a highly constrained embedded platform, offering a template for other resource-limited visual inspection tasks. It also documents the Adam-to-SGD optimizer failure mode explicitly, which is useful practical guidance often omitted from published work.

Real-world applications:

• Residential and commercial roof inspection: Automated drone surveys that flag defects before they cause leaks, mold, or insulation failure, reducing the need for inspectors to physically climb onto roofs.
• Insurance claim assessment: Faster, repeatable, and more standardized documentation of roof condition for underwriting and claims processing.
• Building energy audits: Identifying moisture-damaged or degraded roof sections that degrade thermal performance, supporting retrofit prioritization.
• Municipal and climate policy support: Enabling large-scale roof condition surveys that feed into urban heat island mitigation and cool-roof adoption programs.

Industry relevance: Roofing accounts for roughly one-third of construction-related deaths in the United States between 1992 and 2009, and heating and cooling consume about 65% of residential energy use in Canada. A system that reduces human risk while lowering inspection costs and supporting energy-efficiency goals addresses both a safety and an economic problem. The work was supported by Mitacs and Roofmaster Ottawa Inc., indicating direct commercial interest.

Future Directions

1. Extending beyond mod-bit membranes: The model is trained exclusively on modified-bitumen roofs; adapting it to TPO, EPDM, and PVC systems would require new labeled data and validation.
2. Isolating the dual-altitude benefit: The authors explicitly note that further work is needed to quantify how much the multi-resolution imaging strategy contributes to detection performance on its own.
3. Resolving environmental ambiguity: Distinguishing transient moisture from true standing-water defects may require complementary sensing such as thermal infrared imaging or expanded annotation protocols.
4. Reintroducing excluded imagery: Images containing decks, mechanical units, and roof edges were removed early on; the long-term goal is to reincorporate them with proper labels to improve real-world robustness and reduce overfitting to idealized conditions.

Target Audience

This paper benefits most from applied machine learning engineers, robotics and UAS researchers, and computer vision practitioners working on embedded or edge deployment. It is also relevant to roofing and building-envelope professionals, insurance technologists, and building energy analysts evaluating automated inspection tools. Readers seeking theoretical advances in deep learning will find less value here; the contribution is squarely in applied system design under hardware constraints.

Authors’ abstract

Flat roofs are among the most influential components of the building envelope, governing both structural performance and thermal efficiency, and thereby contributing directly to household energy consumption, carbon emissions, and long-term environmental sustainability. 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. This paper presents a real-time, Unmanned Aerial System (UAS)-based deep learning framework that autonomously detects defects using live imagery captured during dual-altitude aerial passes. 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. To meet the strict computational and power constraints of embedded UAS hardware, the proposed framework integrates a tile-based architecture with a lightweight Convolution Neural Network-Support Vector Machine (CNN-SVM) classifier designed for low-latency onboard inference. The final model-comprising five convolutional layers and four dense layers, the last a linear SVM head, achieved a mean test accuracy of $94.4\%$ ($95\%$ confidence interval $\pm0.4\%$ over three seeds) on a photo-level split ($43,383$ training, $3,869$ validation, and $2,540$ test tiled and augmented images), outperforming GoogLeNet ($89.2\%$) and AlexNet ($79.8\%$). Experimental evaluations using real UAS imagery collected by onsite visits with DJI Matrice 350 RTK drone demonstrate that the system supports rapid, repeatable, and safe roof inspections while reducing human risk, lowering operational costs, and enabling more sustainable building maintenance.