Application of UAVs to Support Blast Design for Flyrock Mitigation: A Case Study from a Basalt Quarry

Abstract

Blasting operations in surface mining pose a risk of flyrock, which is a critical safety concern for both personnel and infrastructure. This study presents the use of unmanned aerial vehicles (UAVs) and photogrammetric techniques to improve the accuracy of blast design, particularly in relation to controlling burden values and reducing flyrock. The research was conducted in a basalt quarry in Lower Silesia, where high rock fracturing complicated conventional blast planning. A DJI Mavic 3 Enterprise UAV was used to capture high-resolution aerial imagery, and 3D models were created using Strayos software. These models enabled precise analysis of bench face geometry and burden distribution with centimeter-level accuracy. The results showed a significant improvement in identifying zones with improper burden values and allowed for real-time corrections in blasthole design. Despite a ten-fold reduction in the number of images used, no loss in model quality was observed. UAV-based surveys followed software-recommended flight paths, and the application of this methodology reduced the flyrock range by an average of 42% near sensitive areas. This approach demonstrates the operational benefits and enhanced safety potential of integrating UAV-based photogrammetry into blasting design workflows.

1. Introduction

The use of explosives for rock fragmentation remains the primary method of extraction in surface mining operations for rock raw materials. The application of blasting agents allows for effective resource exploitation; however, it also poses significant environmental and safety risks. One such hazard is flyrock, which presents a danger to mine personnel, equipment, and surrounding infrastructure. Although this risk is critical, it remains difficult to accurately assess.

During blasting operations, flyrock can be ejected over considerable distances. As a result of the detonation of explosives within blastholes, rock fragments are displaced, typically causing horizontal movement of the entire blasted rock mass. When the blasting sequence is carefully designed and properly executed, the process should not pose substantial risks. Nevertheless, in addition to the expected displacement, flyrock may be expelled unpredictably, creating safety hazards for personnel, equipment, and structures within its range. Flyrock refers to any rock fragment projected into an unintended area [1].

The extent of the flyrock zone depends on a number of factors that can be quantitatively determined under controlled conditions through measurements. Identifying and monitoring these parameters enables the implementation of measures to mitigate flyrock hazards. The key contributing factors include [2,3,4,5,6,7,8,9,10,11,12,13,14]:

• Burden and spacing of blastholes,
• Specific charge (powder factor),
• Inadequately characterized geological structure,
• Blasthole deviation,
• Length and quality of stemming,
• Presence of back fractures,
• Initiation sequence,
• Human error.

In most cases, these factors are interrelated; for instance, blasthole deviation affects burden and spacing, which in turn can locally increase the powder factor. For this reason, the literature offers a range of empirical methods and formulas for estimating the maximum flyrock range, such as the Swedish method [15,16], the American method (Roth) [17], Pokrowski’s method [18,19,20,21], and those based on explosive energy indices. Traditional empirical formulas and statistical regression models often proved inadequate for accurate flyrock prediction, showing low predictive capacity. In the past decade, a wide range of artificial intelligence (AI) and machine learning (ML) techniques have been applied to this problem, generally achieving much higher accuracy than empirical approaches. This section reviews these techniques, grouped by model type, and discusses their predictive performance, practical field applications, and noted limitations (e.g., data scope and input sensitivity).

1.1. Artificial Neural Network (ANN) Approaches

Artificial neural networks have been extensively used to model blast-induced flyrock. Early studies demonstrated that ANN models can significantly outperform conventional statistical methods. For example, Trivedi et al. [22] applied back-propagation neural networks (BPNNs) to predict flyrock distance in Indian limestone quarries and found the ANN’s predictions to be far more accurate than multivariate regression analysis (MVRA) on the same data. In a follow-up study, Trivedi et al. [22] simultaneously predicted flyrock distance and rock fragmentation using a BPNN. The network yielded substantially lower root-mean-square error (RMSE) and mean absolute error (MAE) than MVRA, confirming the ANN’s superior accuracy. Monjezi et al. [23] trained a three-layer ANN on 192 blast records from Sangan iron mine (Iran) and likewise observed reliable predictions, identifying key influential inputs, such as powder factor, stemming length, and charge-per-delay. Notably, by using the ANN’s insights to adjust blast design (e.g., reducing the charge per delay), the mine was able to reduce observed flyrock distances from 165 m to 25 m, illustrating a direct practical benefit of ANN modeling in the field. Overall, these studies established that well-trained ANNs can capture the highly nonlinear relationships governing flyrock and serve as a better predictive tool than regression models [22,23], with typical reported coefficients of determination (R²) above 0.90 in training and validation phases.

Researchers have also explored optimizations and variants of ANN to further improve accuracy. Monjezi et al. [24] introduced a neuro-genetic model, using a genetic algorithm (GA) to optimize an ANN’s architecture and parameters. This hybrid ANN-GA model (sometimes called a genetic algorithm optimized neural network) achieved higher predictive efficiency than a standard neural net on the same dataset. Similarly, other works integrated evolutionary algorithms with ANN training (see the Hybrid Models Section below). In general, ANNs have proven highly effective for flyrock prediction, consistently outperforming empirical equations and providing actionable insights (e.g., sensitivity analysis from ANN weights often highlights the most critical factors controlling flyrock throw). However, one limitation noted is that an ANN’s performance and derived insights are only as good as the training data—if the data come from a single site or do not span a wide range of conditions, the network may not generalize well to other mines (a point revisited later). To mitigate this, some studies trained ANNs on multi-mine datasets: for instance, Trivedi et al. [25] pooled data from four different limestone mines (125 blast events) to train an ANN and still achieved a high prediction accuracy (the ANN’s outputs correlated much better with observed flyrock distances than any regression fit). This indicates that with sufficient and diverse data, ANN models can be robust tools for flyrock forecasting in practice.

1.2. Support Vector Machines and SVR Models

Support vector machine (SVM) algorithms have also been successfully applied to flyrock prediction, especially in cases with complex, nonlinear parameter interactions. SVM-based models generally outperform linear regression and even perform favorably against ANNs in certain cases. Amini et al. [26] tested SVM for predicting flyrock at the Sungun copper mine and found it “faster and more precise” than an ANN model on the same data. Khandelwal and Monjezi [27] developed an SVM regression model for flyrock prediction and compared it to MVRA. The SVM achieved a coefficient of determination R² ≈ 0.95, dramatically higher than the 0.44 achieved by multiple regression, and reduced prediction error (MAE) to about 3.1 m versus 7.7 m for MVRA [27]. These results demonstrate SVM’s ability to capture nonlinear patterns that simple regressions miss, yielding much improved accuracy [28]. The success of SVM is further reflected in its variants like support vector regression (SVR). Guo et al. [29] proposed an ensemble of SVR models (with different kernels) combined via a regularized linear meta-learner. This stacked SVR model achieved excellent performance (test R² = 0.993 and RMSE ~ 3.74 m), significantly better than any single SVR (whose R² ranged from 0.92 to 0.97). Likewise, Li et al. [30] utilized a hybrid SVR optimized by the Harris Hawks algorithm, which attained test R² ≈ 0.97 and RMSE ~9.7 m, outperforming not only a standard SVR but also other machine learners like ANN and extreme learning machine (ELM) on the same dataset. These studies confirm that SVM-based approaches are highly capable for flyrock modeling, often matching or exceeding ANN accuracy. A practical consideration with SVM/SVR models is the need to tune kernel and regularization parameters. Advanced optimization techniques (genetic algorithms, particle swarm, etc.) have been applied to automate this tuning with great success in flyrock problems (e.g., yielding the R² ~ 0.97 noted above for hybrid SVR [31]). As with ANNs, SVM models benefit from adequate training data—in fact, Khandelwal and Monjezi’s SVM was trained on the same small dataset (from one mine) as their regression yet generalized far better [27]. This robustness suggests SVMs can be a powerful tool even when data are limited or noisy, provided the model parameters are properly optimized.

1.3. Fuzzy Logic and Neuro-Fuzzy Systems

Fuzzy inference systems provide an alternative AI approach that has been applied to handle the uncertainty and nonlinearity in flyrock phenomena. In a pure fuzzy logic model, expert knowledge or data-driven rules map input parameters to flyrock distance using linguistic fuzzy rules. Rezaei et al. [32] developed a Mamdani-type fuzzy model for an iron mine in Iran and reported its performance was “much better” than a conventional statistical regression on the same 490-blast dataset. The fuzzy model yielded more accurate flyrock predictions and was able to incorporate vague relationships between inputs and outputs that crisp equations could not capture. Similarly, Ghasemi et al. [33] compared an ANN and a fuzzy inference system for flyrock prediction in the Sungun mine—both models proved useful, but the fuzzy model slightly outperformed the ANN in predictive performance. These outcomes suggest that fuzzy systems, which do not require an explicit functional form, can effectively model the complex interplay of blasting factors. A noted advantage is that fuzzy models provide interpretable if–then rules, which can be meaningful for engineers (e.g., a rule might relate “high powder factor and short stemming” to “long flyrock distance” in linguistic terms).

To leverage the strengths of both neural networks and fuzzy logic, researchers have employed the Adaptive Neuro-Fuzzy Inference System (ANFIS) for flyrock prediction. ANFIS is essentially a neural network that implements fuzzy inference rules and can learn rule parameters from data. Hudaverdi and Agan [34] applied ANFIS (alongside a standard ANN and Gaussian process regression) to data from a Turkish quarry, using important input features selected via statistical analysis. The ANFIS model proved most accurate, with a mean absolute error (MAE) of only ~5.36 m—outperforming the back-propagation ANN in that study. In fact, the intelligent models (ANN, ANFIS, and GPR) all predicted flyrock with <6 m error (mean percentage errors < 10%), but ANFIS slightly edged out the others [35]. In a related study, Hudaverdi [36] demonstrated that simplifying the input set can improve neuro-fuzzy models: using variable selection procedures, he developed two ANFIS models with only 3–4 inputs each, which achieved median errors < 5 m and MAPE < 8%. These ANFIS models were not only accurate but also less complex, underlining the benefit of focusing on the most predictive parameters [34]. Another advantage of neuro-fuzzy approaches is their ability to highlight influential inputs. For instance, Hudaverdi’s studies identified the burden-to-hole diameter ratio and specific charge (powder factor) as consistently significant features for flyrock prediction.

Recent research has pushed the performance of ANFIS even further by integrating metaheuristic optimizers (essentially creating hybrid neuro-fuzzy models). Nguyen et al. [37] introduced an ANFIS enhanced by a Lévy-flight Jaya algorithm (a modern swarm intelligence optimizer) to predict flyrock. This ANFIS–LJ model achieved R² ≈ 0.981 on the test set (with MAE ~ 1.42 m), and when validated on 13 new production blasts it reached R² = 0.988 with only ~1.3 m MAE. Such high accuracy (within ~1–2 m of observed values) is exemplary for flyrock studies and was obtained using 5-fold cross-validation to ensure the model generalizes well. In a complementary study, another team led by Nguyen [38] tested ANFIS models optimized by different algorithms (adaptive differential evolution, genetic algorithm, fireworks, and bee colony). They found the JADE-ANFIS (ANFIS optimized by JADE) to be the best, with an overall prediction accuracy of ~97.8% and very low error (MAPE = 1.1%). Competing models in that comparison achieved 89–96% accuracy (MAPE up to 3%), so the JADE-ANFIS was clearly superior [37,38]. Interestingly, the sensitivity analyses from these neuro-fuzzy models often concur with earlier findings. For example, the JADE-ANFIS study noted that stemming length was the most critical factor influencing flyrock distance, which aligns with other works that frequently identify stemming and charge-related variables as key drivers (these factors essentially control the confinement of explosive energy). In summary, fuzzy and neuro-fuzzy approaches (especially when boosted with learning algorithms) have attained excellent predictive performance in flyrock modeling. They combine accuracy with interpretability, making them attractive for practical implementation—though like all data-driven models, their rules are valid within the range of training data and should be applied cautiously beyond that range [37,38].

1.4. Hybrid Metaheuristic and Advanced Machine Learning Models

Given the inherent complexity of blasting processes, many researchers have turned to hybrid models that integrate machine learning with evolutionary or swarm optimization techniques. The motivation is two-fold: (1) to optimize model parameters/architecture for better accuracy, and (2) to possibly find optimal blasting designs that minimize flyrock. A common strategy is coupling an ANN or other learner with an optimizer to fine-tune its weights or hyperparameters. For example, Armaghani et al. [39] combined particle swarm optimization with an ANN (PSO-ANN) to improve convergence and avoid local minima in training. Applied to 44 blast cases in Malaysian quarries, the PSO-ANN was able to predict flyrock distance with “a high degree of accuracy”, better than a standard back-propagation ANN, and the authors reported that powder factor and charge per delay were the most influential inputs on the model. Marto et al. [40] used a different evolutionary algorithm (imperialist competitive algorithm, ICA) to optimize an ANN and showed that this ICA-ANN hybrid outperformed a conventional ANN as well as two newly developed empirical equations on a dataset of 113 blasts. Likewise, Monjezi et al. [24] had earlier demonstrated a GA-optimized ANN (neuro-genetic model) with superior performance to an unoptimized network, underscoring the value of such hybridization.

Beyond ANN enhancements, researchers have also explored hybrid models with alternative learning algorithms. Extreme learning machine (ELM) is a fast neural network variant, and Murlidhar et al. [41] optimized it using biogeography-based optimization (BBO). Their BBO-ELM achieved a testing R² ≈ 0.94, compared to R² ≈ 0.79 for a plain ELM on the same data (262 blast samples). In fact, the hybrid BBO-ELM slightly outperformed a PSO-tuned ELM as well (which had R² ≈ 0.93), making BBO-ELM the top model in that study [40]. A recent study by Bhatawdekar et al. [42] introduced an equilibrium optimizer coupled ELM (EO-ELM) and benchmarked it against PSO-ELM and PSO-ANN models. Using 114 blast records for training/testing, the EO-ELM attained an impressive R² ≈ 0.97 (RMSE ~ 32 m) on the test set, whereas the PSO-ANN yielded R² ~ 0.87 (RMSE ~ 64 m) and PSO-ELM ~ 0.88 (RMSE ~ 49 m). This huge accuracy gain (EO-ELM reducing error by ~50% relative to the ANN model) demonstrates how modern metaheuristic optimizers can dramatically improve model performance [41]. It is worth noting that such hybrids also often include a sensitivity analysis component; for instance, Bhatawdekar et al. [42] identified powder factor and blastability index as the most sensitive inputs (each with a normalized sensitivity of 0.98) in their EO-ELM model.

Another category of hybrid models aims to combine predictive modeling with optimization of blasting parameters. Zhou et al. [43] took a two-stage approach, first training an ANN to predict flyrock, then using a PSO algorithm to search for optimal blast design parameters that minimize the predicted flyrock. In a case study, they showed that by adjusting the pattern (within practical limits), the anticipated flyrock distance could be reduced to ~34 m under ideal conditions (or ~109 m under typical field conditions), whereas the baseline scenario had much greater flyrock throw. This highlights the potential of AI not just to predict outcomes, but to actively improve blasting practices. Similarly, Saghatforoush et al. [44] combined an ANN with an ant colony optimization (ACO) procedure. Their trained network predicted flyrock and back-break with high accuracy (for the ANN alone, Ea ≈ 0.014 and RMSE ≈ 0.063 in scaled units), and then the ACO algorithm was used to optimize the blast layout. By implementing the ACO-derived design changes, the mine could potentially achieve a 61% reduction in flyrock distance (and 58% reduction in back-break) according to the model’s estimations. These studies demonstrate a practical extension of predictive models into prescriptive tools for blast optimization.

1.5. Tree-Based Ensembles and Other Methods

In addition to ANN and SVM, various tree-based machine learning methods have been applied to flyrock prediction, often yielding performance on par with the best ANN/SVM models. Decision tree ensembles like random forests and gradient boosting machines can naturally handle nonlinear interactions and have the advantage of built-in variable importance analysis. Yari et al. [45] evaluated four tree-based models—decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), and adaptive boosting (AdaBoost)—for a mine dataset. All four techniques provided accurate predictions (R² > 0.95), with AdaBoost being the most precise (R² ≈ 0.99 for both training and testing). Such near-perfect fits indicate that the models captured almost all variance in the data, though the authors cautioned that burden and spacing had the least impact on the model output while powder factor had the highest influence. Hasanipanah et al. [46] similarly found that a regression tree model outperformed a multiple linear regression for flyrock distance prediction in a Malaysian quarry, based on metrics like median absolute error and R². The regression tree was more precise, and a sensitivity check again pointed to powder factor as the dominant factor affecting flyrock throw. These results reinforce that ensemble or rule-based tree models can be highly effective. One advantage is that they are relatively straightforward to apply and interpret; for example, a regression tree yields if–then rules similar to expert reasoning (e.g., “if stemming < X and charge > Y, expect flyrock > Z”). Indeed, Hudaverdi and Akyildiz [35] proposed a novel classification approach using multiple discriminant analysis (MDA) to categorize flyrock throw severity in a quarry. Instead of predicting an exact distance, their model classifies blasts into risk categories (no flyrock, moderate flyrock, and severe flyrock) based on blast parameters, aligned with regulatory safety criteria. The MDA-based model successfully classified flyrock outcomes for test blasts and has the benefit of simplicity—it does not require specialized software or extensive training to use. The authors even produced a territorial map for the quarry so that engineers can easily determine the expected flyrock severity class from blast design parameters [35]. This approach, while not yielding a numeric prediction, is very practical for field use as a quick risk assessment tool.

Ensemble learning has also been used in creative ways. Guo et al. [29] (as mentioned earlier) combined multiple SVRs and a GLM-net aggregator to exploit diverse models’ strengths, achieving R² = 0.99-level accuracy. Nguyen et al. [47] took an ensemble of ANNs approach: they trained five different ANN models and then ensembled their predictions to train a meta-ANN (referred to as EANNs). This stacked model attained R² ≈ 0.986 (with very low error, e.g., RMSE ~4.3 m) on the flyrock dataset, outperforming any single ANN and even a benchmark ANN of identical structure. Such ensemble methods benefit from averaging out individual models’ errors, often yielding more stable and accurate predictions. However, a potential downside is the complexity of implementation and the need for more data to train multiple models. In practice, these approaches show that combining models can push the predictive accuracy to extremely high levels—albeit with the caveat that models trained and tested on the same site’s data (especially if limited in size) might overfit to that context. For example, an AdaBoost or EANN model achieving ~99% accuracy on a given quarry’s blasts [45,47] may not maintain that level on a different mine without retraining. Recognizing this, some recent studies explicitly tested generalization: Zhang et al. [48] optimized a random forest using two novel algorithms (light spectrum optimizer and puma optimizer) and evaluated it on two separate quarry datasets (one from Malaysia and one from Istanbul). Both optimized RF models performed very well (test R² ≈ 0.96–0.97 and RMSE on the order of 6–16 m) and significantly outperformed an unoptimized RF. More importantly, when the model trained on one site was validated on a completely different site (an “engineering case” not seen in training), it still achieved a respectable R² ~0.88. This indicates strong generalization capacity, likely due to the robustness conferred by optimization and the use of diverse training data. Moreover, by using SHAP (Shapley Additive Explanations) analysis, Zhang et al. [48] identified the critical factors for each site’s model—for instance, hole diameter and maximum charge per delay were most important in the Malaysian quarry, while specific charge and the spacing-to-burden ratio mattered most in the Istanbul site. Such insights are valuable for practitioners to understand regional differences and to focus on controlling the right parameters in each operation.

1.6. Practical Applications and Common Findings

The breadth of studies above shows a consensus that AI/ML techniques can predict blast-induced flyrock with far greater accuracy than older empirical methods. Models such as ANN, SVM, tree ensembles, and hybrids routinely achieve R² values in the 0.90–0.99 range, with RMSE on the order of only a few meters to a few tens of meters (depending on dataset scale)—a substantial improvement over classical formulas that often had R² values well below 0.5 in comparative tests [27]. Notably, many different model types (from simple fuzzy systems [32] to deep ensembles [29]) have proven successful, suggesting that as long as the major influencing factors are captured, a variety of AI models can map them to flyrock outcomes effectively. Across the literature, there is strong agreement on which input parameters are most influential in flyrock distance. Nearly all studies highlight some measure of explosive charge and rock confinement as key drivers, e.g., powder factor (explosive per rock volume) emerges as a top factor in numerous sensitivity analyses, while maximum charge per delay and stemming length are also repeatedly identified as critical inputs. In contrast, parameters like burden, spacing, or rock density often rank lower in influence, although their effects are not negligible and can be site-specific. The general finding is that flyrock risk is most governed by how much explosive energy is released and how well that energy is confined—which aligns with practical blasting experience. This gives confidence that AI models are picking up meaningful physical relationships, not just statistical coincidences.

From a practical standpoint, several research works have moved beyond pure prediction into field implementation and validation. For instance, Monjezi et al. [23] used their ANN model’s recommendations to adjust blasting in an active mine, achieving drastic flyrock reduction (as noted, from 165 m down to 25 m). Armaghani et al. [28] built a regression-based model and then performed a Monte Carlo simulation to establish a probabilistic safe distance for flyrock; interestingly, the simulated mean flyrock (236.3 m) closely matched the observed mean (238.6 m), giving credence to using such models for defining blast clearance zones. However, they cautioned that models should be used only under conditions similar to those they were developed for and directly applying them elsewhere is not recommended. This caution about generalizability is echoed by many authors. Flyrock is a highly site-specific phenomenon, influenced by local geology and blast practices, so an AI model trained on one mine’s data may need retraining or recalibration for another site. The risk of overfitting looms, especially when models boast extremely high R² on limited datasets (e.g., an AdaBoost or ensemble ANN that fits one mine’s blasts almost perfectly [41,45]). Without rigorous cross-validation or external testing, such performance might be optimistic. To address this, recent studies increasingly adopt techniques like k-fold cross-validation (as in [37]) and multi-site data fusion (as in 48]) to ensure models generalize. Some works also explicitly exclude less relevant inputs to reduce noise—e.g., using random forest or statistical selection to prune variables—which can improve model robustness when moving to new conditions [36,49,50].

Another important practical aspect is data acquisition for model development and use. Traditional methods of measuring flyrock (manual spotters and video analysis from a single camera) can be error-prone (e.g., due to parallax) and dangerous. The adoption of modern monitoring technologies like unmanned aerial vehicles (UAVs) and high-speed videography is making it easier to collect accurate flyrock data for both model training and operational prediction. Lawal et al. [51] performed a comprehensive bibliometric review and noted that conventional means of locating flyrock fragments suffer from biases and errors, and they demonstrated a field trial where UAV-mounted cameras captured flyrock trajectories more reliably. The UAV-based method showed clear advantages in capturing the full flight of rock fragments, which were then analyzed with motion analysis software, and a soft computing model was used to predict the recorded flyrock distances [51]. Similarly, Mishra et al. [52] used high-speed cameras onsite (including UAV-mounted) to record flyrock from blasts at five mines, generating a rich dataset of launch velocities and distances. They then trained several ML models, deploying the best one (an extremely randomized trees regressor) in a prototype software tool to simulate flyrock trajectories for new blasts. This integration of AI models into user-friendly interfaces allows engineers to input planned blast parameters and visualize the predicted flyrock hazard zone, greatly aiding in blast design and risk mitigation. More generally, UAV photogrammetry and remote sensing techniques are increasingly utilized in mining, as they can provide rapid, high-resolution mapping of blast sites and measure outcomes like fragmentation and throw with centimeter-level precision. Studies have shown UAV-based monitoring to be efficient, removing the need for personnel in hazardous areas and improving data quality. For example, Bamford et al. [53,54] used UAVs to continuously monitor open-pit blasts, capturing detailed images for fragmentation analysis and flyrock observation, which offered insights for process improvement that traditional manual monitoring could miss. The proliferation of such technologies, as reviewed by Colomina and Molina [55] and Minh and Dung [56], underscores that UAVs have become an excellent tool in mining for tasks ranging from surveying to blast monitoring. In the context of flyrock prediction, better data acquisition means more reliable models and the possibility of near-real-time model updating and validation during blasting operations.

In conclusion, the application of AI and machine learning to flyrock prediction in open-pit blasting has yielded marked improvements in predictive accuracy and practical risk management. Diverse models—ANNs, SVMs, fuzzy systems, tree ensembles, and numerous hybrids—have each demonstrated high accuracy (with R² often in the high 0.9 range and errors reduced by 50–80% compared to older methods) in various case studies [27,29,33,45]. Hybrid models that combine techniques (ANN + GA, ANFIS + PSO, optimized SVR, etc.) tend to achieve the best performance, reflecting a common finding that no single method is universally best, but combining algorithms can harness their complementary strengths for superior results [37,38,41,42,57]. A recurring theme across studies is the identification of crucial parameters: in almost all models, powder factor (or specific charge) and stemming length emerge as primary controls on flyrock distance, whereas parameters like burden, spacing, or rock quality may have secondary effects. This consistency builds confidence in the physical relevance of the models. On the other hand, researchers consistently warn about limitations: models are usually data-driven and thus valid only within the domain of the training data. For instance, Armaghani et al. [28] explicitly state that their site-specific model should not be directly applied to other conditions without recalibration. To improve generalizability, recent works have embraced cross-validation, larger multi-site datasets, and input reduction to avoid overfitting [37,38]. Another concern is data representativeness—since flyrock events are stochastic and sometimes rare (extreme flyrock incidents might not occur in every dataset), ensuring the model has seen enough “bad cases” is important for reliable hazard prediction. Despite these challenges, the trend in the literature is toward increasingly sophisticated and validated models, often coupled with modern monitoring (e.g., UAV and high-speed cameras) for both model input collection and output verification. Many studies have reported successful field implementations—from adjusting blast designs based on model insights to deploying software that can simulate flyrock before a blast is carried out [23,26,35]. These advancements suggest that AI-driven flyrock prediction is maturing from academic research into practical technology [57].

Although this study does not directly apply artificial intelligence methods for flyrock prediction, the comprehensive literature review included in the introduction aims to highlight the dynamic development of this research area and underscore the need for integrating advanced computational tools with geospatial data acquired through UAV-based surveying. The presented case study represents a step toward such integration, demonstrating the potential of precise geometric analysis as a foundation for the future development of predictive models grounded in real-world spatial data.

2. Materials and Methods

2.1. Geodetic Technologies Supporting Blast Design

The design of a blasting hole pattern that meets safety requirements for ongoing operations must be based on precise geometric measurements of the burden zone to be excavated from the active deposit. The selection of appropriate geometric parameters for the blasthole grid—in correlation with the type of explosive used and the charge configuration—is a key condition for ensuring operational safety while optimizing the utilization of detonation energy from the explosive charges placed in the holes.

In the blast design process, geodetic tools play a critical role, particularly in the open-pit extraction of hard rock raw materials. Among the most commonly used devices are 3D scanners (e.g., laser total stations), which enable accurate mapping of surface topography and bench face geometry. Global navigation satellite systems (GNSSs) are also of significant importance, providing high-precision positioning of blastholes and other critical points within the mining area. Increasingly widespread is the use of UAVs, which, through rapid photogrammetric data acquisition, support the geometric inventory of the excavation site and enable efficient planning of blasting operations.

Geospatial data obtained from these tools are increasingly being integrated directly into specialized blast design software, such as BlastPlan, JKSimBlast, QuarryX, Strayos, or O-Pitblast. These platforms allow for detailed modeling of the blast layout, forecasting of blasting outcomes (e.g., fragmentation, muckpile displacement, and vibration levels), and verification of compliance with applicable safety standards and environmental constraints.

2.2. Laser Total Stations for 3D Terrain Modeling

Laser total stations are currently among the most widely used instruments for conducting the types of measurements described above, offering high precision in determining horizontal and vertical angles as well as distances. Classical tacheometry involves polar coordinate-based surveying to determine the planimetric and altimetric position of detail points, using trigonometric leveling. Based on the collected point cloud data, a 3D model of the bench face or its section can be generated in appropriate processing software. Figure 1 presents a sample 3D model visualized as a point cloud, showing the roof, bench face, and floor of the excavation wall along with the planned layout of the blasthole pattern.

In the case of highly irregular bench faces, there is a high probability that an optimal measurement position for the instrument cannot be found—one that would enable complete geometric reconstruction of the entire face. As seen in Figure 1, the perspective from which the scan was conducted limited the ability to capture the full dimensions of individual surfaces—roof, floor, and face. In such cases, supplementary scans must be carried out from additional vantage points to dimension surfaces that were not visible during the initial measurement.

Another issue involves the coordinate systems required by the instrument operator. While mapping the model in a standard Cartesian coordinate system (x, y, z) is generally unproblematic for commercially available devices, mapping into custom or local coordinate systems requires the establishment of well-defined reference points. These must precisely define geographic length, width, and elevation. Moreover, these reference points must be visible from all scanning perspectives used during the survey of a given area of the excavation site.

These considerations highlight the fact that measurements performed using stationary laser-based instruments may, under certain conditions, prove insufficiently functional. This can directly affect the quality and completeness of the collected data, which are subsequently used to determine the input parameters for blast design.

2.3. UAV-Based Photogrammetry: Capabilities and Limitations

In recent years, a significant increase has been observed across various industries in the use of measurement techniques based on photogrammetry—a scientific and technical discipline that reconstructs real-world dimensions, shapes, and spatial relationships of objects through the analysis of image sequences captured from multiple viewpoints. Particularly notable is the rapid development of close-range photogrammetry using UAV, equipped with specialized cameras that minimize optical distortions, such as chromatic aberration (Figure 2).

This trend is clearly reflected in the mining sector as well. A comprehensive literature review by Minh Dang Tuyet and Nguyen Ba Dung [56] analyzed 113 scientific publications over the past 12 years, documenting the systematic increase in UAV applications across key mining activities. These include mine surveying and inventory (application in surveying mines), design and monitoring of drilling and blasting operations, mine safety and risk management, ground subsidence monitoring, rock slope stability analysis and monitoring, detection of underground coal fires, and erosion detection.

An optimal—though not exclusive—solution for blasting design is the use of aerial photogrammetry. This technique, by virtue of its mobile measurement platform, inherently eliminates issues related to limited perspective or access typical of ground-based scanning. Figure 3 illustrates an example flight path planned for UAV scanning of a selected production bench.

Key rules to keep in mind before flying a bench [58] are as follows:

• At least two-thirds of the photos should be taken with the camera pointing straight down (nadir images).
• Capture images along flight paths with 70–80% overlap.
• Effective results have been reported at flight altitudes of 75 ft (23 m), 150 ft (46 m), and 250 ft (76 m).
• Always overshoot the mapping area by at least 20 ft (6 m) to avoid stitching or reconstruction issues at the edges.
• Fly the UAV forward with the camera tilted 5° forward; when beginning the next pass, rotate the UAV instead of flying backward.

As previously mentioned, the use of stationary terrestrial laser scanners for single or series-based surveys may present operational limitations due to structural and functional constraints of the device. These limitations often result in extended measurement durations, increasing the risk of incomplete geometric data acquisition, which in turn can compromise the accuracy of blast design.

UAV deployment significantly mitigates problems related to instrument positioning and scanning geometry, thereby improving overall operational efficiency. Photogrammetric techniques based on UAV imagery are typically more functional and offer potentially higher spatial resolution. However, UAV-based measurements also come with certain technical and regulatory limitations. These include high sensitivity to weather conditions (e.g., wind, rain, and fog), limited operational time due to battery capacity, the requirement to maintain stable data transmission with the ground control station, and—under most regulations—the necessity of maintaining visual line of sight (VLOS).

Additionally, photogrammetric data quality can be affected by lighting conditions and surface characteristics—low contrast, homogeneous textures, or reflective surfaces may degrade the final 3D model, requiring supplementary ground-based measurements for validation.

It should be noted that stationary terrestrial laser scanning does not require formal operator licensing. In contrast, UAV operations in public airspace are legally regulated. Operators must obtain UAV pilot certification, which includes theoretical and practical training followed by an official examination. The required qualification level depends on UAV mass and the operational risk category.

Furthermore, effective 13 November 2025, national aviation regulations [59] will mandate liability insurance for all UAV operators. Operating UAVs without valid insurance or in violation of legal regulations may result in substantial financial penalties [59].

3. Results and Case Study: UAV-Assisted Blast Design in a Basalt Quarry

3.1. Site Description and Context

The methodology for designing a blasthole pattern using photogrammetric techniques was implemented in one of the basalt quarries located in the Lower Silesian region of Poland. The exploited deposit is characterized by a high degree of fracturing, which significantly complicates rock fragmentation using blasting methods. Accurate determination of the actual and optimal burden values on the working face is critical to ensuring the safety of blasting operations. This must account for the potential risk of excessive flyrock and the necessity to control seismic impacts on the surroundings of the mining site.

An increase in the geometric parameters of the burden directly correlates with a rise in the resistance of the rock mass, which in turn affects both the intensity of seismic vibrations and the degree of rock fragmentation. Precise determination of key parameters, such as blasthole length, angle and direction of inclination, and subdrill depth, is an essential step in blast design. Errors introduced at this stage are difficult if not impossible to correct in subsequent phases of the process, such as selecting the type and mass of explosives, determining the initiation sequence, and setting millisecond delays.

Therefore, it is essential to use measurement and analytical tools that enable continuous verification of the design assumptions throughout all stages of blasting operations from planning to execution.

The presented case study aims to illustrate the scope of challenges involved in designing blasting operations with a focus on controlling the extent of flyrock. Work on implementing and optimizing the use of UAV-based photogrammetry for blasting applications began in late 2023. Since then, each blast series has been designed using actual 3D data obtained from the excavation site. The 3D models obtained from UAV flights were validated using reference measurements conducted at fixed control points within the quarry area, utilizing precise geodetic instruments, such as a total station. This verification confirmed the high geometric accuracy and reliability of the generated models.

3.2. UAV Equipment and Photogrammetric Data Acquisition

To define the design assumptions and monitor their implementation, photogrammetric measurement data were collected using a DJI Mavic 3 Enterprise (DJI Sky City, No.55 Xianyuan Road, Nanshan District, Shenzhen, China) unmanned aerial vehicle (Figure 1). This UAV is equipped with a wide-angle camera featuring a 4/3 CMOS sensor. The 20-megapixel sensor includes a mechanical shutter capable of speeds up to 1/2000 s and an image capture interval of just 0.7 s. These features prevent motion blur and enable the operator to conduct photogrammetric missions at flight speeds of up to 15 m/s.

In addition to the wide-angle camera, the DJI Mavic 3 Enterprise is equipped with a 12 MP telephoto lens that enables up to 56× hybrid zoom. This makes it suitable not only for standard photogrammetric missions but also for detailed inspections from considerable distances.

The UAV can remain airborne for up to 45 min, allowing it to cover areas of up to 2 km² in a single mission. It uses the next-generation DJI O3 Enterprise (DJI Sky City, No.55 Xianyuan Road, Nanshan District, Shenzhen, China) transmission system, ensuring a stable video feed (1080 p/30 fps) even under challenging environmental conditions.

For enhanced safety, the Mavic 3 Enterprise is equipped with the upgraded DJI APS 5.0 system, which uses six wide-angle “fish-eye” optical sensors to provide omnidirectional obstacle avoidance without blind spots. An advanced return-to-home (RTH) system automatically calculates the most efficient return route to the launch point, conserving time, energy, and battery life.

Precise measurements are further supported by an integrated real-time kinematic (RTK) module, which enables centimeter-level positioning accuracy [60].

The acquired data were processed using Strayos software [58], which is dedicated to blast design and analysis.

3.3. Blast Design Using 3D Geospatial Data

A total of 30 blastholes were included in the designed blasthole pattern, with individual hole lengths ranging from 15.4 to 15.8 m. Figure 4 presents the parameters of two example blastholes, developed based on a 3D model of the selected quarry section.

Thanks to the integration of an RTK module, each point in the 3D model is precisely georeferenced. This enables accurate determination of blasthole lengths while accounting for terrain irregularities, hole inclination, and orientation—with centimeter-level precision in reaching the target bench elevation (Figure 5). The designed blasthole pattern and the actual positions of the drilled holes are verified by importing drilling data from the HNS (Hole Navigation System) into the Strayos software, where the final burden verification is performed based on the true geometry of the executed pattern.

Analysis of the burden geometry based on the 3D model enables a reliable calculation of actual burden values for each segment of the bench face. These values are obtained through photogrammetric evaluation using a fully three-dimensional visualization of the geometric dataset. As shown in Figure 6, the burden is defined as the shortest distance from the blasthole axis to the nearest point on the bench face, measured continuously along the entire bench height.

In contrast to the conventional 2D burden measured in a single vertical plane perpendicular to the hole axis, the 3D burden provides a more accurate representation of local resistance to detonation by incorporating variations in face geometry along all spatial axes. This enhances the quality of blast assessments, particularly for charge placement optimization and flyrock risk reduction.

Figure 6 presents the results of the spatial burden analysis, indicating that the minimum burden value required under the assumed safety criteria is 2.5 m. The vertical lines represent the positions of the blastholes, while the thin black horizontal lines correspond to burden lines at different elevation levels. Colored lines—red, green, and blue—highlight the locations on the bench face where the minimum burden occurs. In this case study, the 3D model revealed several areas where the burden fell below the safety threshold. Red zones represent areas with insufficient burden (<2.5 m), which significantly increase the risk of flyrock. Green zones denote adequately designed burden values (2.5–4.0 m), considered optimal for the given geological and operational conditions. Blue zones indicate excessive burden (>4.5 m), which may result in poor rock fragmentation and elevated seismic effects, especially near the toe of the bench. Access to detailed geospatial data allows blast designers to proactively revise layouts before drilling begins. In the execution phase, these data support the selection of appropriate explosives, charge masses, and configurations, matched to the actual burden in each blasthole.

Figure 7 compares burden estimates derived from 3D and 2D analyses. In the 3D model, purple indicates areas with undersized burden, while blue denotes properly dimensioned regions. The 2D burden is illustrated as a black line through yellow reference points. Notably, hole No. 8 shows a partially detached rock fragment in the profile, an anomaly unlikely to be detected using traditional 2D laser scanning. Furthermore, the toe zone exhibits a significant discrepancy: the 2D burden value is more than double the actual 3D burden, emphasizing the risk of underestimating flyrock potential in conventional assessments.

The most comprehensive assessment is achieved using a 3D burden heatmap, which visualizes actual burden values not just along individual hole profiles but across the entire blast area (Figure 8). In this representation, green indicates burden values considered optimal based on geological and operational context, red highlights zones with insufficient burden, and blue identifies areas with excessive burden.

Overall, the integration of UAV-based 3D data with specialized blast design software allowed for dynamic adjustment of the drilling layout, minimizing risk zones and improving overall blast performance.

4. Discussion

The conducted study confirms that the use of photogrammetry supported by unmanned aerial vehicles (UAVs) significantly improves the accuracy of acquiring geometric data for blast design and contributes to enhanced safety by reducing the risk of excessive flyrock. These findings are consistent with the general trend in surface mining, where remote sensing technologies and UAV-based measurement systems are increasingly applied to support operational planning. Minh and Dung [56], in their comprehensive literature review, highlighted the growing importance of UAVs across key mining activities, including drill and blast design.

In the presented case study carried out at a basalt quarry in Lower Silesia (Poland), UAV-supported methods enabled the creation of a highly detailed 3D model of the bench face and the generation of centimeter-precision burden heatmaps. This allowed the identification of zones with inadequate burden (<2.5 m) and excessive burden (>4.5 m), which is critical both in terms of flyrock risk and blasting efficiency. Compared to conventional 2D measurements, the 3D analysis revealed burden discrepancies exceeding 100%, particularly in toe zones, which underscores the insufficiency of flat projections for safe blast pattern design.

The reduction in the number of required images for 3D modeling by nearly a factor of ten, along with an 80% decrease in processing time, confirmed the operational efficiency of this approach. Moreover, the average reduction of flyrock range by an average of 42% (the average value was estimated based on photographic and video documentation) in areas near protected structures demonstrated the practical effectiveness of blast optimization using 3D spatial analysis.

However, certain limitations of the proposed methodology should be acknowledged. UAV-based photogrammetry is sensitive to environmental factors, such as wind, precipitation, fog, lighting conditions, surface reflectivity, and visibility of ground control points, all of which can affect data quality. Additionally, UAV operations require operator certification and are subject to legal regulations, including the obligation to hold valid liability insurance in Poland starting November 2025 [60].

It is important to note that this study did not incorporate predictive modeling methods, such as AdaBoost, random forest, or artificial neural networks, which have been shown to provide highly accurate forecasts of flyrock extent. These techniques allow probabilistic estimation of hazardous areas based on multiple input variables, such as burden, spacing, powder factor, and hole deviation. Combining such models with 3D measurements could further improve prediction accuracy. However, the performance of machine learning models is heavily dependent on the quality and representativeness of the training data, which may be limited in field practice.

A review of the current scientific literature clearly shows a growing trend in the use of advanced mathematical and computational models for predicting flyrock throw in blasting operations. In many studies, the reported coefficients of determination (R²) exceed 0.95, indicating an almost perfect fit between predicted and observed values. Such high accuracy, however, raises valid concerns from a practical standpoint, considering the significant variability and unpredictability of geological conditions in real-world mining operations. Rock formations differ not only in lithology, but also in bedding, weathering degree, presence of joints, and voids, all of which critically affect the propagation of detonation waves and rock fragmentation.

Another methodological concern relates to how the influence of different design parameters is interpreted. Some studies suggest that the powder factor is the most influential parameter in predicting flyrock distance, while burden has a negligible impact. Such conclusions are questionable because these two parameters are directly interrelated—burden represents the resistance of the rock mass to explosive energy, which in turn dictates the amount of energy (and hence explosive material) required for effective fragmentation and displacement. Ignoring this relationship leads to oversimplified and potentially misleading predictive models.

Incorporating accurate burden values into predictive models remains a particularly challenging task. Figure 7 clearly illustrates the substantial variation in actual burden along the profile of a single blasthole. Traditional approaches that assume a constant burden along the entire hole length introduce significant errors in estimating the true rock resistance. Similar issues arise with the powder factor, which is typically calculated as an average across the entire blast pattern and does not account for localized variations in charge distribution, influenced by bench geometry or the presence of detached rock blocks.

For blasting engineers responsible for the planning and execution of blast operations, it is essential to have access to tools that are both simple and effective—tools that allow for the rapid, high-resolution mapping of bench face geometry and the subsequent adaptation of explosive charge mass and design to actual field conditions. Only with such tailored design can flyrock throw be minimized effectively.

It is also important to recognize that even the most sophisticated predictive models cannot replace the experience and local geological knowledge of a skilled blasting engineer. Such an expert, familiar with the specific characteristics of the deposit and its geological variability, can design blasting operations that are both efficient and safe. Therefore, the future of effective flyrock mitigation lies not only in the continued development of predictive tools but also in their intelligent and responsible application by trained professionals.

At the same time, it should be emphasized that no tool currently offers complete assurance against the occurrence of excessive flyrock. One of the main challenges remains the lack of fast and reliable methods for detecting voids between the blasthole pattern and the face geometry. In this context, the precision of geodetic measurements and their integration with specialized blast design software become critical components in ensuring safety in mining operations.

When analyzing only equipment-related costs, the use of UAV technology proves economically favorable compared to traditional surveying methods, such as laser total stations (based on average pricing in the Polish market). The purchase of a high-quality laser total station typically ranges between USD 40,000 and 55,000, with an additional software cost of approximately USD 4000–6000. In contrast, the cost of acquiring a UAV system with RTK functionality ranges from USD 5000 to 8000. Additional expenses include UAV operator training (approximately USD 500, valid for five years), annual insurance (~USD 200), and software licensing fees ranging from USD 200 to 1000 per month, depending on the software’s capabilities. Considering average values, the total investment in UAV-based surveying becomes comparable to that of a total station after approximately 5–6 years of use, assuming no major equipment failures occur. Furthermore, operating a total station usually requires two personnel, while UAV data acquisition can be conducted by a single trained operator. The time required to survey the same section of a bench face is also significantly shorter with UAVs, resulting in greater operational efficiency and reduced labor costs.

In summary, the integration of photogrammetric measurements, 3D modeling, and advanced blast design tools significantly enhances the reliability and safety of blasting operations. Future research should focus on hybrid approaches that combine real-time 3D measurements with machine learning algorithms, especially in geologically complex environments or areas close to protected infrastructure. Nevertheless, increasing technical awareness and accountability among blast designers and supervisors remains essential, as human expertise continues to be the most vital link in the decision-making and execution chain.

5. Conclusions

The presented work focused on the application of UAV-based photogrammetry in the design of blasthole patterns in a basalt quarry. The use of high-resolution aerial imagery and RTK-supported 3D modeling enabled precise reconstruction of the geometry of the bench face, including critical parameters, such as hole length, inclination, and actual burden distribution. Compared to traditional 2D approaches, the 3D analysis revealed significant spatial variations, especially in toe zones, directly impacting blast efficiency and flyrock risk.

The method led to improved safety conditions, including a documented reduction in flyrock range by an average of 42% near protected structures. In addition, the photogrammetric approach allowed for a ten-fold reduction in the number of images and an 80% decrease in model processing time without compromising mapping accuracy. It should be noted that results may vary depending on specific mining conditions, and it is recommended that similar UAV-supported blast design approaches be further tested in other quarry environments. All UAV flights were conducted in accordance with Strayos software recommendations regarding image overlap, flight path, and camera orientation.

Preliminary cost comparisons indicate that, over time, UAV-based methods offer a cost-effective and efficient alternative to traditional surveying techniques, particularly in operations requiring frequent and rapid terrain documentation.

Despite operational constraints, such as weather sensitivity, licensing requirements, and surface reflectivity issues, UAV photogrammetry proved to be a robust and efficient solution for operational blast planning. Further integration with predictive models and machine learning may offer even greater control over risk assessment and blast outcome optimization.