PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Samuel Dunthorne, Hashim A. Hashim

arXiv:2609.13013v1cs.CVcs.AI

TL;DR

Flat-roof defects threaten building performance, while autonomous UAS inspection faces onboard computing and data constraints. The paper uses tiled imagery and a lightweight CNN-SVM approach for roof-defect detection; its model achieves 94.4% mean test accuracy and outperforms GoogLeNet and AlexNet on the reported test set.

  • Problem

    Autonomous UAS roof inspection must combine accurate detection with low-latency onboard processing despite hardware constraints and limited defect-labeled data.

    Source: A. Related Work · p. 3

  • Method

    The framework combines image tiling with a lightweight CNN-SVM architecture to process roof imagery for defect detection on UAS hardware.

    Source: §VI · pp. 16–17

  • Results

    94.4% mean test accuracy (95% confidence interval ±0.4% over three seeds) outperformed GoogLeNet (89.2%) and AlexNet (79.8%).

    Source: D. CNN-SVM Time Complexities · p. 15

  • Takeaways & Limitations

    The tile-based architecture is designed to support rapid inference while preserving detail and computational capacity for critical UAS functions.

    Source: §VI · pp. 16–17

    The photo-level split can place images sharing pixels of the same roof across training, validation, and test sets; a roof- or flight-level split is planned to quantify and remove this effect.

    Source: E. Limitations and Future Work · p. 16

Abstract

from arXiv · show

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.

I. INTRODUCTION

Roof integrity matters for building energy use, emissions, moisture control, and occupant health. The introduction frames UAS-based detection as a response to inspection hazards and the challenge of achieving accurate, low-latency inference on constrained aircraft.

I. INTRODUCTION

  • Roof damage can reduce energy efficiency and insulation performance, increase energy demand, and allow moisture ingress that can lead to mold growth.The paper connects roof performance to household energy use, carbon emissions, air quality, and public health.Source: §I · pp. 1–2
  • Cool roofs, including vegetated roofs and reflective membranes, can reduce building cooling loads, with studies reporting reductions of up to 34%.The passage describes these approaches as reducing solar absorptance or heat gain.Source: §I · p. 2
  • Roofing defects range from major structural or moisture failures to minor maintenance issues, whose type distribution varies with regional climate and roof systems.The reported number of issues per 1,000 homes is relatively consistent across Northern Atlantic regions.Source: Figure 2 · p. 2, §I · p. 3

A. Related Work

  • Visual, thermal, physical, and destructive roof inspections rely on trained inspectors and can involve hazardous work, motivating UAS inspections.The cited account says UAS inspections reduce risk, save man-hours, and allow rapid, repeatable data collection.Source: A. Related Work · p. 3
  • Autonomous UAS inspection must balance accurate detection with low-latency onboard processing despite power, weight, resolution, and labeled-data constraints.Shadows, cloud cover, precipitation, and visual similarity between defects and healthy roofs can also reduce detection reliability.Source: A. Related Work · p. 3

B. Modern Machine Learning

The paper situates CNN-based defect detection among varied machine-learning approaches, noting constraints from limited datasets and UAS hardware. It then presents a tile-based CNN-SVM framework, a multi-altitude imaging rationale, and validation with real UAS imagery.

B. Modern Machine Learning

  • CNN-based approaches are consistently recommended for defect detection, but performance is constrained by limited datasets and onboard processing demands.The related-work discussion describes multiple alternatives while noting that heavier methods can exceed mobile hardware capabilities or interfere with the UAS mission.Source: B. Modern Machine Learning · pp. 3–4

C. Contribution and Structure

  • The proposed tile-based CNN-SVM architecture preserves high-resolution defect cues while reducing computational load for near-real-time inference on embedded UAS hardware.Its computation-aware design uses optimized hyperparameters and reduced feature-map dimensionality.Source: C. Contribution and Structure · p. 4
  • The computation-aware design is reported to achieve over 92% testing accuracy while remaining deployable on low-power processors.Source: C. Contribution and Structure · p. 4
  • The dual-altitude, multi-resolution imaging strategy is intended to target small, irregular, or partially occluded defects, but its benefit has not been isolated in a dedicated ablation.Source: C. Contribution and Structure · p. 4
  • The operational pipeline was validated with real imagery from onsite DJI Matrice 350 RTK visits and was designed to reduce human risk, lower inspection costs, and support sustainable maintenance.Source: C. Contribution and Structure · p. 4

II. PROBLEM FORMULATION

Autonomous roof inspection must balance accurate defect detection with the power, weight, and processing limits of UAS platforms. The paper addresses dataset and classification challenges including moisture-related ambiguity, limited labeled data, and variation in imagery.

II. PROBLEM FORMULATION

  • UAS roof inspection requires accurate detection while operating within strict power, weight, and processing constraints.The paper identifies real-time onboard inference as challenging for high-resolution imagery.Source: §II · p. 4

A. Dataset and Preparation

  • Recent-rain dampness can resemble defective prolonged ponding in RGB imagery, complicating supervised classification.The paper identifies expanded datasets, stronger annotation protocols, or complementary sensing as possible differentiation strategies.Source: A. Dataset and Preparation · pp. 4–5, Figure 3 · p. 5

B. Image Tiling

  • With no suitable public labeled dataset identified, the authors assembled and annotated a custom mod-bit roof dataset from online and UAS imagery.Images were tiled into 640×360-pixel regions to train on localized areas and improve computational efficiency.Source: B. Image Tiling · p. 5
  • Tiles were labeled defect if they contained a recognized mod-bit failure mode, otherwise no defect.The binary labels aggregate cracking, alligatoring, granular loss, blistering or wrinkling, exposed deck, moisture staining or failed seams, and seam separation.Source: B. Image Tiling · p. 5
  • The dataset and model are specific to modified-bitumen membranes, and generalization to TPO, EPDM, or PVC is not claimed.A single expert annotator labeled the data, so interannotator agreement could not be computed.Source: B. Image Tiling · p. 5

C. Data augmentation

  • Random rotation, flipping, brightness variation, and noise injection expanded the dataset while retaining each tile’s class label.The authors report that augmentation mitigated classification gaps under previously unseen illumination, membrane colours, and surface conditions.Source: C. Data augmentation · p. 5

D. Image Information Density and Distance

The section motivates using imagery from multiple distances to capture both broad anomalies and fine defects, while addressing training-image clutter and the resolution–computation trade-off. It also describes the CNN feature-extraction rationale and the defect categories represented in the tiles.

D. Image Information Density and Distance

  • Fig. 5 represents cracking, granular loss, alligatoring, wrinkling or blistering, moisture staining, seam failures, and exposed deck as defect tiles, with a healthy tile for contrast.These modified-bitumen failure modes are aggregated into binary defect/no-defect labels.Source: Figure 5 · p. 6
  • A dual-pass strategy is suggested: survey at high altitude for large anomalies, then fly low for fine-scale defect detection.Close-range, high-resolution imagery helps reveal cracking or alligatoring, while higher-altitude imagery supplies context for missing membrane or pooling water.Source: D. Image Information Density and… · pp. 6–7, Figure 4 · p. 6

E. Dataset Compilation and Exclusions

  • Images containing irrelevant rooftop objects were excluded early to reduce classification ambiguity and overfitting.The long-term objective is to reintroduce such images with appropriate annotations as the dataset expands and the model matures.Source: E. Dataset Compilation and Exclusions · p. 7

III. METHODOLOGY

  • CNN convolutions extract hierarchical image features, while pooling reduces feature-map dimensionality to balance computation and feature preservation.The paper notes that higher-resolution inputs support fine-defect detection, whereas standard 256 × 256-pixel downscaling can compromise feature-map detail.Source: §III · p. 7

A. tiling

Tiling reduces image-processing load while retaining classification-relevant detail, and varying flight altitude helps address defects that lose form across tiles. The proposed CNN-SVM processes tiled RGB images through convolutional and dense layers to make binary defect decisions.

A. tiling

  • 640 × 360-pixel tiles reduce feature-map sizes and processing time while retaining detail; varying flight altitudes helps interpret larger objects that may lose form across tiles.Images at 1,280 × 720, 1,920 × 1,080, and 3,840 × 2,160 pixels are divided into four, nine, and 36 tiles, respectively.Source: A. tiling · p. 7
  • With 80% image overlap, photo capture at 0.5–1.4 images per second supported successful results while reducing computational demands relative to continuous video capture.Photos are used because overlap preserves visibility of defects in neighboring images and roof surfaces are static.Source: A. tiling · p. 8

B. CNN-SVM Model architecture

  • The architecture processes tiled RGB inputs through augmentation, convolution and pooling, then four dense layers ending in a linear squared-hinge SVM head for binary decisions.The final SVM decision value assigns a tile to the defective class at or above zero, and non-defective otherwise.Source: Figure 8 · p. 9, B. CNN-SVM Model architecture · p. 9
  • The adopted squared-hinge loss is differentiable, easing gradient-based optimization and penalizing margin violations more strongly than hinge loss.It replaces the conventional sigmoid cross-entropy head used in standard CNN classifiers.Source: B. CNN-SVM Model architecture · p. 9

IV. WORKFLOW, CHALLENGES, AND MITIGATION

The preliminary model plateaued below 70% training accuracy, a problem attributed to overfitting. Switching from Adam to SGD substantially improved performance, while the final architecture used five convolutional and four dense layers.

A. Overfitting and Optimization Issues

  • Preliminary training accuracy and loss plateaued below 70%, which the authors attributed to overfitting; data augmentation alone did not resolve the issue.The early-stopping callback also triggered immediately and repeatedly before the optimizer was changed.Source: A. Overfitting and Optimization Issues · pp. 9–10
  • The described architecture has five convolutional layers, with filter counts of 32, 64, 128, 256, and 512 across the first through fifth layers, and two max-pooling layers.Source: A. Overfitting and Optimization Issues · p. 10
  • After switching from Adam to SGD with a 0.005 learning rate and 0.2 momentum, model accuracy exceeded 90% and early stopping no longer triggered prematurely.The earlier callback repeatedly triggered around the third epoch; the final configuration monitored loss with a 0.001 minimum-improvement threshold.Source: A. Overfitting and Optimization Issues · p. 10

B. Workflow

The proposed workflow uses overlapping images from two altitude passes, tiles each image, and classifies the tiles for operator review. Only tile-level classification and its training, evaluation, and inference-time results were implemented and benchmarked; the remaining end-to-end functions are future work.

B. Workflow

  • Two passes at different altitudes target complementary scales: higher imagery captures large structural issues, while lower imagery resolves fine-grained defects.Images are captured at approximately 1–2 per second with 50%–90% overlap, then tiled and processed sequentially at 640 × 360 pixels.Source: B. Workflow · p. 10
  • The implemented and benchmarked work covers tile-level classification, training and evaluation, and inference-time results; autonomous flight, onboard execution, georeferenced mapping, and automated revisits remain future work.The described end-to-end workflow is therefore a concept beyond the realized per-tile classification stage.Source: B. Workflow · p. 10

V. EXPERIMENTAL RESULTS

The supplied results passages describe an architecture sweep and two training-choice comparisons for the tiled-image model. They report qualitative effects of optimizer and augmentation choices but do not provide the corresponding cell values.

V. EXPERIMENTAL RESULTS

  • The architecture sweep covers 2–6 convolutional and 2–4 dense layers, with held-out test accuracy, training accuracy, training time, and epochs reported.The caption identifies the adopted five-convolution/four-dense configuration but supplies no cell values here.Source: Table II · p. 11
  • The image and classification breakdown is presented in Table III, but its caption provides no counts or class proportions.No quantitative breakdown can be reported from the supplied caption alone.Source: Table III · p. 11
  • Replacing SGD with Adam collapses training to the majority class, while removing augmentation nearly perfectly fits training data and widens the train-test gap.The table caption cautions that these single-run values vary by run and points to seed-averaged accuracy elsewhere.Source: Table IV · p. 11

A. Dataset Download

The study archives its dataset, scripts, and trained weights for reproducibility, and evaluates model configurations using a photograph-level split. The selected CNN-SVM balances test accuracy and training cost.

A. Dataset Download

  • The dataset, split manifest, processing and evaluation scripts, and final trained weights are openly archived under CC BY 4.0.A fixed random seed of 42 was used across experiments.Source: A. Dataset Download · pp. 11–12

B. Proposed CNN-SVM Performance and Test Accuracy

  • The five-convolution/four-dense model reached 94.1% test accuracy, within roughly half a point of the best configuration, while training about an hour faster.Across tested configurations, test accuracy ranged from 87.8% to 94.5%; increasing convolutional depth improved test accuracy modestly, while dense depth had little effect.Source: B. Proposed CNN-SVM Performance and… · p. 12
  • Photographs were split 75%/15%/10% into training, validation, and test sets, keeping each photograph’s tiles and augmentations within one split.The resulting sets contained 43,383 training, 3,869 validation, and 2,540 test images; validation and test used original tiles only.Source: B. Proposed CNN-SVM Performance and… · p. 12
  • 94.4% mean test accuracy (95% CI ±0.4% over three seeds) was achieved after iterative revisions with the proposed CNN-SVM.The reported model also reached approximately 97.6% training accuracy and 94.0% validation accuracy.Source: B. Proposed CNN-SVM Performance and… · p. 12

C. Comparison to Cutting-edge Techniques

On the shared roof-defect test set, CNN-SVM achieved higher accuracy than AlexNet and GoogLeNet while requiring less inference time in the reported comparisons. The section also reports class-wise errors, error patterns, and computational costs.

C. Comparison to Cutting-edge Techniques

D. CNN-SVM Time Complexities

  • The model identified 1,065 of 1,157 defective test tiles, giving 92.0% recall and an 8.0% false-negative rate.It correctly classified 1,330 of 1,383 non-defective tiles, for a 96.2% true-negative rate and 3.8% false-positive rate; the paper notes that the threshold can be lowered to trade precision for additional recall.Source: D. CNN-SVM Time Complexities · p. 15
  • Missed defects were mainly thin cracks near a tile’s resolution limit, while false alarms were triggered by small surface debris.Membrane seams were generally classified correctly as non-defective.Source: Figure 13 · p. 14, D. CNN-SVM Time Complexities · p. 15
  • The approximately 117-million-parameter model requires about 446 MB of FP32 weights and 243 million operations per tile.A full 36-tile image requires about 8.7 billion operations; inference was profiled on a workstation GPU, with embedded-accelerator profiling identified as future work.Source: Table VII · p. 15, D. CNN-SVM Time Complexities · p. 16

E. Limitations and Future Work

The reported evaluation has boundaries in data splitting, hardware testing, dataset coverage, and flight-strategy validation.

E. Limitations and Future Work

  • Photograph-level splitting prevents tile-level leakage but still allows overlapping views of the same roof across sets, so reported accuracy warrants modest optimism.A roof- or flight-level split is planned to quantify and remove this potential overlap effect.Source: E. Limitations and Future Work · p. 16
  • Training and inference were measured on a workstation GPU; onboard deployment and profiling of power, latency, and thermal behavior remain future work.The model requires approximately 243 million operations per tile and under 1 GB of parameter memory.Source: E. Limitations and Future Work · p. 16
  • The dataset excluded images containing walls, vehicles, or mechanical units, and performance across environmental conditions and roof membranes was not evaluated.The benefit of dual-altitude data collection has not been isolated through a dedicated ablation.Source: E. Limitations and Future Work · p. 16
  • Reported metrics come from three training runs, and additional repeated runs are planned to better characterize variance.The runs used seeds 42, 43, and 44.Source: E. Limitations and Future Work · p. 16

VI. CONCLUSION

The conclusion presents a lightweight, tile-based UAS roof-inspection model with strong test accuracy and rapid inference, alongside an end-to-end inspection workflow. It also reports the dataset, reproducibility resources, and author contributions.

VI. CONCLUSION

  • 94.4% mean test accuracy (95% CI ±0.4% over three seeds); the final CNN-SVM outperformed AlexNet and GoogLeNet while requiring substantially less processing time.It processed a full 36-tile image in roughly 7.8 ms, about 130 full images per second.Source: §VI · p. 16
  • The inspection workflow covers image acquisition, tiling, classification, and operator feedback while preserving capacity for navigation, communication, and collision avoidance.The bespoke dataset was built mainly from onsite DJI Matrice 350 RTK UAS surveys and supplemented with online images.Source: §VI · pp. 16–17

CREDIT AUTHORSHIP CONTRIBUTION STATEMENT

  • Dunthorne contributed original-draft writing, visualization, validation, software, methodology, investigation, formal analysis, and conceptualization; Hashim contributed review and editing, supervision, funding acquisition, and other work.Both authors are credited with visualization, validation, methodology, investigation, and conceptualization.Source: CREDIT AUTHORSHIP CONTRIBUTION STATEMENT · p. 17

DATA AVAILABILITY

  • The tiled and augmented imagery dataset, split manifest, scripts, and trained model weights are openly archived on Zenodo.The passage specifies CC BY 4.0 for the dataset and states that experiments use fixed random seed 42.Source: DATA AVAILABILITY · p. 17

APPENDIX A CALCULATION OF PERFORMANCE AND TIME

The appendix describes how classification outcomes are counted and how trainable parameters and operation counts are estimated for convolutional and dense layers.

APPENDIX A CALCULATION OF PERFORMANCE AND TIME

  • True positives and true negatives are correctly classified defective and non-defective images; false positives and false negatives are the corresponding misclassifications.The passage points to Table VIII and Fig. 14 for metric calculations and illustrations.Source: Appendix A · p. 17
  • Convolutional-layer parameters are counted as (w × h × c + 1) × k, with filter size, input channels, and filter count defining the terms.Here w × h is filter size, c is input channels, and k is the number of filters.Source: Appendix A · p. 17
  • Dense-layer parameters are computed as n × m + m, where n and m are the input- and output-unit counts.The formula is the dense-layer parameter count described in Eq. (6).Source: Appendix A · p. 17
  • Convolutional-layer operations use the input feature-map dimensions, filter dimensions, channel count, and stride; dense-layer operations use n and m input and output units.The convolutional count is 2(c × w × h)(M − w + s)(N − h + s); the dense-layer count is m(2n + 1).Source: Appendix A · p. 17
Loading 2609.13013v1…