Application of artificial intelligence on drone-collected data for cut-and-fill volume estimation of Chatoloma, Kasungu-Jenda road.

Abstract

Like many other engineering projects in developing countries, the Kasungu-Jenda Road has faced numerous challenges in data collection and analysis, including slowed surveying works, underestimations, and overestimations of cut and fill volumes. These challenges often lead to increased financial and time-related costs. This study examined the application of Artificial Intelligence processing techniques on drone-collected data for road surveying on the Kasungu-Jenda Road in Malawi. Specifically, it assessed the effectiveness of machine learning (ML) algorithms in object classification, evaluated their potential in predictive analytics for cut -and-fill volume estimation, and compared their volumetric accuracy with traditional computation methods. A quantitative experimental design was employed, utilising drone imagery, traditionally collected as-built data, and road design data. This data was processed and analysed through Convolutional Neural Networks (CNNs), Random Forest, XGBoost, and Support Vector Machine (SVM) algorithms. CNN models demonstrated high effectiveness in object classification, achieving 90.91% accuracy across multiple land cover types, including vegetation, bare land, and earthworks. For predictive analytics, XGBoost achieved the highest accuracy (MAE = 0.0052 m; R² = 0.9847), followed by Random Forest (MAE = 0.0106 m; R² = 0.9451) and SVM (MAE = 0.0150 m; R² = 0.8908), all indicating low error margins. On the other hand, a comparison of XGBoost with the Prismoidal Method in cut and fill volumet ric estimations using 2-meter chainage drone elevations, and 20-meter traditional chainage elevations revealed statistically significant differences by a paired samples t -test (p < 0.05). The difference is attributed to XGBoost's ability to detect finer terrain variations from high-resolution data, which are not captured by traditional methods. The study concludes that while conventional approaches remain useful, ML-based models in combination with drone technology offer improved precision and spatial detail in engineering surveying. It therefore recommends incorporating these technologies to improve both the accuracy and reliability of survey work