Modelo de classificação do coeficiente de atrito em pistas de pouso e decolagem com imagens aéreas e aprendizado profundo

Abstract

The reliability of landing and takeoff operations depends directly on runway skid resistance, which is primarily ensured by maintaining adequate friction levels on the pavement surface —a condition that can be compromised by rubber buildup resulting from aircraft braking. Traditional methods for measuring runway friction require specialized equipment and trained personnel, potentially causing partial or total disruption of airport operations. Considering the wide availability, low cost, and increasing resolution of satellite imagery provided by Google Earth, this study proposes the development of a classification model for runway friction coefficients using such imagery, supported by Convolutional Neural Networks (CNNs). A dataset of approximately 7,000 satellite images of Brazilian aerodrome runways was assembled and linked to actual friction coefficient measurements. Three image preprocessing techniques (CLAHE, Gaussian filter, and Wavelet) and five neural network architectures (simple CNN, ResNet50, DenseNet121, InceptionV3, and VGG16) were evaluated. The combination of CLAHE and Gaussian filtering with the DenseNet121 architecture yielded the best performance, achieving over 74% accuracy in cross-validation. This result was obtained using a binary classification model based on a 0.71 threshold – a technical reference commonly adopted to distinguish between adequate and critical friction levels. External validation with unseen images from three separate field campaigns revealed spatial variations in accuracy across runway thirds and lateral zones, with increased errors near the thresholds, where rubber accumulation is more prevalent. An additional test using images captured by an Remotely Piloted Aircraft (RPA) demonstrated limitations in the model's generalization to visual domains different from those used during training. The proposed approach shows potential as a complementary tool for direct friction assessment, enabling proactive identification of runway segments at risk of reduced skid resistance and compromised pavement performance.