https://orcid.org/0000-0003-1037-9731
Figures
Abstract
Granite quarrying, a cornerstone of the construction industry in South India, yields significant economic benefits but poses substantial environmental and social challenges, including land degradation, dust pollution, alternation of the water regime, and harsh working conditions. Rapid urban expansion has escalated granite demand in many countries, intensifying quarrying activities. This trend is particularly pronounced in Bengaluru, India, where rural-urban transformation causes concerns about environmental sustainability and social-ecological consequences of urban resource mining. This study proposes an innovative multi-modal framework to monitor granite quarrying in Bengaluru by combining deep learning with a 2024 dry-season multi-date Sentinel-2 composite for quarry segmentation and UAV SfM-MVS photogrammetry for volumetrics. We benchmark five CNN architectures—U-Net, PSPNet, DeepLabV3 + , FCN, and EMANet. In-area development results peaked with DeepLabV3+ (F1 ≈ 94.6%, IoU ≈ 89.7%), while an external, geographically independent audit established PSPNet as the most robust model (F1 = 93.4% [95% CI 90.8–95.9], IoU = 87.6%) with significantly fewer errors than alternatives (McNemar tests, FDR-adjusted p < 0.001). Applying the best model across the region yielded 252 candidates; 227 quarries were confirmed via field checks and sub-meter imagery, spanning 740 hectares. UAV photogrammetry at the Prasannacharipalya site (0.046 m grid; LoD95 masking), yielded a combined lowering volume of 9 280 051 m³ (acceptance area 97.2%; 95% CI ± 17 864 m³, 0.19%). The satellite-to-UAV pipeline enabled automated, scalable quarry footprint mapping with site-level volumetric quantification, offering actionable evidence for environmental management and oversight of quarrying in the quickly-urbanizing study region.
Citation: Himmy O, Nguyen TT, Vazhacharickal PJ, Buerkert A (2025) Monitoring of granite quarries using deep learning and UAV photogrammetry in Bengaluru, India. PLoS One 20(11): e0334493. https://doi.org/10.1371/journal.pone.0334493
Editor: Kristofer Lasko, USACE ERDC: US Army Engineer Research and Development Center, UNITED STATES OF AMERICA
Received: May 4, 2025; Accepted: September 29, 2025; Published: November 6, 2025
Copyright: © 2025 Himmy et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: “All relevant data have been uploaded into the permanent repository Zenodo as follows: - Quarry segmentation: https://doi.org/10.5281/zenodo.17321285 - Volume estimation: https://doi.org/10.5281/zenodo.17322409”.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
Granite quarrying is the process of extracting granite, an igneous rock, from the earth for construction, decoration, and sculpture [1,2]. It is prominent among natural stones globally, with a high demand in domestic and international markets. However, extraction comes with substantial environmental and social impacts [3–5].
Quarrying activities significantly alter landscapes. They involve the removal of soil and bedrock by drilling, cutting, and blasting techniques that break rocks into manageable pieces [6,7]. This entails the loss of topsoil, can result in a barren and unstable land surface [8–10], may cause soil erosion, and ultimately leads to the formation of pits and quarries that may pose safety risks and disrupt water drainage patterns [11]. In Odeda Local Government Area of Ogun State, Nigeria, quarrying activities have led to a decrease in areas covered by light forest, from 637.282 hectares in 1984 to 326.517 hectares by 2014, indicating significant land cover changes due to quarrying [12]. The extraction and processing of materials also generates dust, noise, and vibrations, which may adversely affect the environment and human health. Dust can cover vegetation thereby reducing photosynthesis, and affect air quality for nearby communities [12,13]. On the positive side, in South India many old, abandoned quarries traditionally serve as tanks, providing water for animals and irrigation during the dry season [14].
Remote sensing has emerged as a powerful tool for monitoring various mining activities, offering a cost-effective, efficient, and comprehensive approach to assessing the environmental impacts of mining operations [15,16]. The availability of publicly available datasets such as Landsat and Sentinel, along with the advancement of analysis-ready data archives [17], has facilitated widespread access to remotely sensed information. This, in turn, enables researchers and policymakers to quantify human-induced disturbances caused by mining activities [18].
Delineating mining areas is a critical task within the broader scope of Land Use/Land Cover (LULC) mapping, which is pivotal for environmental monitoring, planning, and management. Traditionally, pixel-based classification has dominated this process, yet it often lacks spatial context and yields “salt-and-pepper” noise and boundary fragmentation [19,20]. This is not only an aesthetic issue: fragmented rims inflate false positives/negatives near quarry edges, bias areal estimates, and propagate error to any downstream change/volume indicators built from those polygons. Object-based image analysis (GEOBIA) partially mitigates these errors via segmentation prior to classification, but the-crafted features and thresholds can generalize poorly across heterogeneous quarry landscapes triggering modern end-to-end, context-aware deep learning approaches [21,22].
In recent years, Convolutional Neural Networks (CNNs) have become standard for extracting mining footprints from multispectral imagery. Despite initial challenges -such as limited pre-labeled datasets and the difficulties in analyzing multispectral and hyperspectral imagery -representative studies report competitive performance across mining task and sensors. For artisanal/small-scale mining (ASM), a Sentinel-2 multispectral CNN detected sites reliably at a regional scale, outperforming traditional baselines and demonstrating policy relevance for unregistered sites [23]. For open-pit delineation on high-resolution imagery, an EMANet + CRF workflow achieved PA ≈ 98.1% and mIoU ≈ 89.5%, while another dense-encoder variant reported precision ≈ 97.7% and IoU ≈ 72.1%, illustrating both the potential and the variability of results across datasets and designs [24]. For tailings-related segmentation, optimized DeepLabv3 + variants have achieved mIoU ≈ 63% under challenging low-contrast conditions, underscoring the value of multi-scale context and boundary-aware decoders yet also the remaining generalization challenges [25]. These design choices address, in measurable terms, the shortcomings of pixel-based methods: multi-scale context reduces spurious speckle over bright soils and construction sites; skip connections and decoders stabilize quarry rims -the features that govern area/perimeter -and attention modules suppress noise in heterogeneous benches/haul roads. This is why recent mining-specific studies have shifted toward CNNs for footprint mapping.
While satellite-scale segmentation maps may show where quarries are and how they expand, regulators and planners also need to know how much material has been extracted. In this context, Unmanned Aerial Vehicles (UAVs) have emerged as effective means for mine mapping and feature quantification, producing centimeter-scale Digital Surface Model (DSM)/ orthophoto products suitable for cut-and-fill analysis. UAV SfM–MVS has matured into a low-cost, high-resolution topographic tool with well-documented accuracy characteristics; best-practice guidance highlights how careful flight geometry, inclusion of oblique imagery, and sensible ground control points (GCP) or RTK/PPK geotagging mitigate systematic errors and support robust change detection and volumetrics. Multiple studies in quarries and stockpiles have shown that UAV-derived volumes compare favorably with GNSS/terrestrial surveys, validating UAVs as a quantitative complement to satellite detection. In terms of precision, UAV SfM–MVS typically achieves absolute mapping accuracies of ~1–3 × GSD when supported by well-distributed GCPs or RTK/PPK image geotags (i.e., a few centimeters for 2–5 cm GSD), whereas no-GCP configurations can exhibit absolute biases at decimeter- to meter-scale [26–28]. Moreover, recent frameworks explicitly connect satellite triggers to targeted UAS acquisition for rapid updates of Digital Elevation Models (DEM), reinforcing the synergy between regional DL mapping and local 3D measurement [29–32]. This study addresses the above limitations with a multi-modal, end-to-end framework to monitoring granite quarrying activities around the South Indian megacity of Bengaluru. We integrated deep learning techniques with satellite imagery and UAV photogrammetry to: (i) develop a robust deep learning framework for automated quarry detection from multispectral satellite data withcontext-aware CNNs to stabilize quarry boundaries and reduce speckle, (ii) benchmark U-Net, PSPNet, FCN, DeepLabv3 + , and EMANet as strong complementary baselines for quarry segmentation. Finally, (iii) we integrate UAV photogrammetry to independently quantify excavation volume (cut-fill) and to cross-check segmentation geometry at centimeter GSD- explicitly reporting uncertainty and discussing practical constraints. This integrated design allows for actionable, scalable monitoring while retaining the precision required for on-site environmental management.
2. Materials and methods
2.1. Study area
According to official data, India, a globally significant granite producer, boasts 3.314 mining leases covering 38 major minerals on a total formal lease area of approximately 306.398 hectares [33]. Hereby, the states of Karnataka, Andhra Pradesh, Tamil Nadu, and Rajasthan contribute about 75% of the country’s total granite production. Bengaluru, the capital of Karnataka, has been chosen as a case study area due to the rapid expansion of granite quarrying and its major apparent impact on the local environment, vegetation, and people’s livelihoods (Fig 1). Known as the India’s ‘Silicon Valley’ and ‘City of Lakes’ Bengaluru’s burgeoning urban development and construction needs have accelerated the demand for granite. This increased demand has resulted in substantial land-use changes often at the expense of natural vegetation and habitats [4].
Map was created using ArcGIS (version: 3.5) from ESRI (http://www.arcgis.com/) under the institutional ESRI site license of the University of Kassel. Basemap satellite image accessed from World Imagery ESRTI Tile Layer. Credits: Esri, Earthstar Geographics, TomTom, Garmin, FAO, NOAA, USGS, © OpenStreetMap contributors, and the GIS User Community.
Bengaluru’s granite (Fig 2) results from the ancient Precambrian rocks foundation of the Peninsular Gneissic Complex (PGC). This complex includes granites, gneisses, and magmatites formed billions of years ago [34]. Over time, weathering has created many soil types ranging from red laterite to clayey deposits.
The environmental effects of granite quarrying in Bengaluru are pronounced. Excavation processes lead to extensive deforestation and the clearance of native vegetation. Soil erosion, water pollution from quarrying activities, and dust emissions challenge air and water quality in the region [35,36]. Further, granite quarrying practices disrupt the hydrological balance and jeopardize nearby water bodies while providing at the same time the possibility to establish new tanks for irrigation or leisure [37–39]. Often, however, the adverse impacts of these quarries extend beyond their immediate vicinity, jeopardizing both urban and rural areas [40].
This study’s focus area encompasses districts around Bengaluru with known active granite quarries or clear evidence of ongoing or historical quarrying. This approach ensured comprehensive coverage of relevant quarry sites and aligns with available baseline data that can support subsequent analysis (Fig 1).
2.2. Data acquisition and preprocessing
2.2.1. Elevation data.
Elevation data were obtained from the NASADEM Digital Elevation Model (DEM), a 30-meter resolution produced by a full reprocessing of Shuttle Radar Topography Mission (SRTM C-band InSAR; Feb-2000) data with improved phrase unwrapping and vertical control using ICESat. Remaining voids are primarily filled with refined ASTER GDEM v3. Distributed in WGS84 horizontal and EGM96 (vertical) datum and covers ~60° N to 56° S in June 2024. This product was accessed via Google Earth Engine (GEE) https://earthengine.google.com/. Published evaluations report typical absolute vertical errors of ~6–10 m Root Mean Square Error (RMSE) for NASADEM, with accuracy varying by terrain ruggedness, slope/aspect, vegetation, and region; consequently, precision is not uniform globally [41,42]. C-band penetration into vegetation can bias elevations in forested areas; users commonly observe reduced biases after reprocessing, but residual effects can persist locally [43,44].
The DEM was resampled using bilinear interpolation to match the spatial resolution of other datasets, ensuring consistent dimensionality across all inputs. This elevation layer helped in refining quarry detection and volume estimation processes by capturing surface morphology and improving the accuracy of subsequent analyses.
2.2.2. Sentinel-2 satellite data.
ESA’s Sentinel-2 MSI provides high-resolution multispectral imagery for accurately detecting granite quarries. Its spatial resolution (10–20 m) and broad spectral range (visible, near-infrared, and short-wave infrared bands: Table 1), enable clear discrimination between quarry surfaces and surrounding landscapes. Granite’s distinctive spectral response, particularly in the visible, Near-Infra Red (NIR), and Short-Wave Infrared (SWIR) range, aids in automated quarry detection using deep-learning models [45].
For this study, Sentinel-2 L2A images from the dry season of 2024 were accessed via Google Earth Engine (GEE). We deliberately limited data selection to the dry season (December–March) to maximize clear-sky observations in this monsoonal climate and to reduce vegetation-related spectral confusion. Dry winter months in India have substantially lower cloud fractions than monsoon months, and dry-season selections are widely adopted for LULC/mining mapping. To ensure data quality, images with cloud cover exceeding 20% were excluded, and the s2cloudless algorithm [46] was applied to remove remaining clouds and shadows. A median composite of all valid observations was generated, minimizing noise and highlighting stable quarry features. As aerosols (Band 1) and water vapor (Band 9) bands contributed minimally to land surface characterization, they were excluded. Six 20 m bands were resampled to 10 m resolution using bicubic interpolation [47]. To enhance detection accuracy, topographic information from the NASADEM was integrated with the Sentinel-2 composite, enabling the model to leverage both spectral and terrain-related features [48,49].
For our study, we selected an area of approximately 93 km by 52 km north of Bengaluru because of its abundance of granite quarries (Fig 1). Quarry polygons used for model training and validation were manually digitized in ArcGIS Pro (v3.3) by a single annotator with expertise in mining operation and practical experience in interpreting high resolution imagery of mining environment. Digitization relied primarily on high-resolution imagery from Google Earth Pro supplemented by Maxar satellite imagery [50]. Each delineated quarry underwent a secondary review using multi-temporal imagery (Sentinel 2) to confirm extraction activity and distinguish between active and abandoned sites. Ambiguous cases were cross-referenced by field surveys. The (Fig 3) composite was tiled into 256 x 256 pixel patches (10 m GSD; 2.56 × 2.56 km per patch) with a 128-pixel stride. To improve rotational invariance, we applied 90° rotations [51]. The dataset (1384 patches) was split into an 80−20 ratio for training and testing, respectively, ensuring proper model evaluation on unseen data while maintaining sufficient training samples for optimal learning [52,53]. This systematic approach to data preparation established a solid foundation for developing a generalizable model capable of accurate quarry detection across diverse landscapes.
2.2.3. Drone imagery.
High-resolution drone imagery was acquired on 8 March 2024 at the Prasannacharipalya granite-quarry site under clear skies (air temperature ≈ 25°C; relative humidity ≈ 24%; wind speed < 3 m s ⁻ ¹), conditions chosen to minimize motion blur and illumination change. A DJI Mavic 2 Pro drone equipped with a 20-megapixel Hasselblad L1D-20c camera flew a nadir grid flight generated in Dronelink [54], at an altitude of 200 meters above ground level (agl), with 75% forward/ 70% side overlap to ensure comprehensive site coverage and accurate photogrammetric processing [55–57]. In total 343 images (5472 × 3648 px) were acquired during the 35-min survey. Georeferencing relied on the drone’s onboard dual-frequency GNSS (absolute positional accuracy ≈ 1–2 m).
The high-quality imagery obtained from this mission served as the primary input for generating a detailed 3D model of the study site. Subsequent processing of these images using Pix4Dmapper software v.4.9 [58] yielded a high-resolution orthomosaic and a precise 3D model. To ensure compatibility between the pre-quarry and post-quarry data, we resampled the DEM using bilinear interpolation to match the resolution of the DSM generated from our 3D model. This technique is particularly effective in maintaining the integrity of the elevation data while allowing for a more refined representation of the terrain [59].
2.2.4. Deep learning-based quarry detection.
We leveraged state-of-the-art deep learning-based semantic segmentation architectures to identify mining sites from satellite imagery. Five primary models, U-Net,PSPNet, DeepLabV3 + , FCN, and EMANet were employed and benchmarked, following the workflow below (Fig 3).
The selection of U-Net, PSPNet, DeepLabV3 + , FCN, and EMANet was based on their complementary architectural strengths which have translated into state-of-the art quarry-segmentation results across multiple recent benchmarks, outperforming earlier CNNs and matching heavier transformer hybrids [60].
U-Net’s encoder-decoder structure preserves fine spatial details by combining high-resolution features from early network layers with deeper contextual representations [61]. The model’s proven capability is to achieve high performance with limited training data, as evidenced by the mean Intersection over Union (mIoU) score of 83.5%. This design makes it particularly well-suited for delineating the detailed boundaries characteristic of quarry sites. PSPNet incorporates pyramid pooling to capture multi-scale global context [62], facilitating more robust detection of mining patterns spread over larger areas. DeepLabV3+ [63,64] leverages Atrous Spatial Pyramid Pooling and a lightweight decoder to capture global context without sacrificing edge sharpness, achieving the highest F1/IoU once larger multisensory datasets are available [65].
FCN [66], canonical fully convolutional baseline allowed for early end-to-end dense prediction (useful reference for gains due to modern context/attention blocks).
EMANet replaces full self-attention with an expectation-maximisation block cutting memory by an order of magnitude [67]; when paired with a CRF it delineated open-pit mines at 89.48% MioU 15 k × 15 k scenes [24]. These covered the main families (skip-decoder, multi-scale pooling, dilated (atrous) convolutions/ ASPP, attention, and baseline). Heavier transformer hybrids were not included to keep computing cost and inference latency compatible with the workflow.
All networks used a ResNet34 backbone (pre-trained on ImageNet) for a strong initialization for feature extraction, enhancing efficiency and mitigating gradient vanishing issues through residual connections.
Inverse frequency weighting was integrated into the cross-entropy loss function to address the class imbalance, emphasizing the accurate detection of less frequent quarry pixels [68]. This approach prioritized the precise prediction of minority mining sites. We used the AdamW optimizer [69] with weight decay to reduce overfitting. Hyperparameters were tuned through iterative experimentation, using an initial learning rate of and a cosine annealing schedule. Batch sizes of 4–8 were chosen to balance computational demands and model performance, and early stopping based on validation loss ensured resource-efficient training [70]. Training and inference were conducted on an NVIDIA RTX 4070 GPU supported by a 13th Gen Intel® Core™ i7-13620H CPU. Models were trained on the 80% training set and evaluated on the non-overlapping 20% test set. We computed per-pixel precision, recall, F1 score (Identical to Dice coefficient), and Intersection over Union (IoU) and aggregated them over test patches – reflecting performance from multiple perspectives [71, Table 2].
For proportions, we applied Wilson 95% confidence intervals; for F1 we computed bootstrap 95% Cis (2 000 resamples of TP/FP/FN) [72]. Pairwise model differences at the same points were tested using exact two-sided McNemar on discordant errors with Benjamini-Hochberg FDR across all pairs [73,74].
To quantify generalization beyond the training area, we defined an external AOI (green area of interest in Fig 1) in South-eastern Bengaluru. We formed two strata (quarry/non-quarry) and sampled 200 points per stratum (n = 400). Each point was visually interpreted on sub-meter imagery (quarry = 1; other = 0). All five models were sampled at the same points without interpolation, allowing for paired tests. This follows recommended practice for design-based accuracy assessment with probability samples [75].
For spatially explicit error analysis, we chose a micro-AOI (~1–3 km²). Quarries were polygonised and filtered to ≥ 0.5 ha, then rasterized (snap/cell size of the training grid) to a binary ground-truth mask. For each model, we computed TP/FP/FN rasters and pixel-wise confusion counts. To further improve segmentation results, morphological operations such as opening and closing were applied to remove noise and fill gaps, refining the quarry boundaries detected.
Validation was carried out through (1) direct field-based observations and measurements at selected sites and (2) comparisons with high-resolution Google Earth imagery to verify areas inaccessible for field validation.
2.2.5. Estimation of extracted granite volume.
The volume of extracted materials at the Prasannacharipalya mining site was determined using a photogrammetric workflow based on our high-resolution drone imagery. Following a preliminary quality check to discard blurred or poorly exposed images, Pix4D Mapper software was employed to process the data. The workflow began with a Structure from Motion (SfM) procedure [29], which identified and matched tie points across overlapping images and performed a bundle adjustment to refine camera orientations. Subsequently, a Multi-View Stereo (MVS) algorithm converted the sparse point cloud into a dense point cloud with an average density of 250 points per square meter [76,77]. From this dense point cloud, a high-resolution Digital Surface Model (DSM) was generated with a ground sampling distance of 5 cm per pixel.
For volume change detection, the UAV-derived DSM was differenced against a reference Digital Elevation Model (DEM) that represented the pre-extraction topography. This reference DEM was reconstructed from historical topographic data and undisturbed terrain near the mining area. Prior to differencing, we co-registered the UAV-DSM to the reference DEM over a mask of stable bedrock to remove residual horizontal/vertical biases and minor tilt using the Nuth & Kääb method [78], thereby minimizing systematic offsets and providing the residuals used to quantify the co-registration variance.
To examine scale effects on the volume estimate, we conducted the analysis at multiple grid sizes. The reference DEM was first resampled to match the working resolution of the DSM, and both the DSM and the reference DEM were then coarsened to a common 1 and 5-meter resolution, and a cut-and-fill analysis was executed in Python environment to compute volumetric differences.
For error propagation, we followed DEM-of-Difference (DoD) best practice if z1 and z2 are the elevations in the DSM and (resampled) DEM and σz1, σz2 their standard deviations (assumed to be independent), the per-cell elevation-difference uncertainty is:
We applied a 95% level-of-detection threshold () to exclude changes below detection. Volumetric uncertainties were then propagated by (i) summing independent per-cell variances and (ii) a correlated formulation that accounts for spatial error structure via an empirical correlation length, following geodetic/DoD practice [79,80]. To avoid over-confidence on steep facets, LoD95 was spatially varied across slope bins (0–5–10–15–25–35–90°) by increasing σz2 with slope; DoD cells were accepted only where |Δz| ≥ LoD95 and within the quarry masks.
As an additional robustness check, we performed a Monte-Carlo (MC) experiment on a 1 m grid in which the post-coregistration DSM was repeatedly perturbed by sub-pixel horizontal (dx, dy) and vertical (dz) jitters drawn from the measured coreg residuals. DoD and volume were recomputed per draw to produce sampling distributions and 95% MC confidence intervals [81].
3. Results and discussion
3.1. Detection and Segmentation Performance
Across the tested architectures, internal (in-distribution) performance was uniformly high (Table 3) in the training area of northern Bengaluru. DeepLabV3+ achieved the strongest balance (F1 ≈ 94.6%, IoU ≈ 89.7%), with PSPNet (F1 ≈ 94.2%, IoU ≈ 89.1%) and U-Net (F1 ≈ 93.6%, IoU ≈ 88.0%) close behind. EMANet and FCN trailed modestly but remained credible baselines. This ranking is consistent with the architectural biases of the models: multi-scale context aggregation (PSPNet’s pyramid pooling; DeepLab’s ASPP) tends to improve region coherence and suppress local texture confusion, while skip-decoder designs (U-Net) emphasize edge fidelity and small-object recovery. These observations mirror the original reports: PSPNet’s pyramid pooling explicitly aggregates global cues at multiple scales, DeepLabv3+ increases receptive field via atrous convolutions without sacrificing borders, and U-Net’s skip connections recover fine details from early layers [62,64]. Because the split is not geographically independent, we treat Table 3 as a development benchmark (descriptive). Still, the near-ceiling values confirm that modern CNNs can reliably map quarries at 10 m when trained and validated in-region, consistent with reports that context modules (pyramid/atrous) improve segmentation over early baselines [82].
On the geographically independent audit, models diverged (Table 4). PSPNet retained high performance with tight uncertainty: F1 = 93.4% (95% CI 90.8–95.9), IoU = 87.62% (TP = 177, FP = 23, FN = 2, TN = 198). The other models were more conservative—higher precision, lower recall—and therefore achieved lower F1/IoU (FCN F1 = 74.49%, IoU = 59.56%; DeepLabV3+ 61.97%/45.05%; EMANet 54.23%/37.36%; U-Net 49.40%/32.97%). Paired McNemar tests on discordant errors confirmed that PSPNet yielded significantly fewer errors than every alternative (FDR-adjusted p < 0.001), establishing a statistically supported rank ordering. This behavior matches broader evidence that explicit context modules (pyramid/atrous) are more robust when background heterogeneity increases (urban fringe, laterite, bare soil) and aligns with recent mining-RS studies using Sentinel-2 CNNs for open-pit/ASM detection [23,83,84]. Operationally, employing a recall-prioritizing model like PSPNet is preferable when the cost of missing active/illegal sites is high; the modest FP surplus is manageable in analyst triage. Where false alarms are costly, a two-stage workflow (contextual veto with DEM/indices) can keep precision high [85,86].
Fig 4 summarizes the “generalization gap” visually: per model, bars show in-area IoU (Table 3) versus external IoU (Table 4). The drop is moderate for PSPNet, but more pronounced for DeepLabV3 + , FCN, EMANet, and U-Net – consistent with the notion that explicit context modules confer robustness beyond the training footprint.
Consistent with the audit statistics in Table 4, PSPNet retains high recall and delineates irregular pits, whereas U-Net, DeepLabV3+, EMANet, and FCN are more conservative and miss small or occluded quarries (Fig 5).
Map was created using ArcGIS (version: 3.5) under the institutional ESRI site license of the University of Kassel, Germany (http://www.arcgis.com/). Basemap satellite image accessed from World Imagery ESRTI Tile Layer. Credits: Esri, Maxar, Earthstar Geographics, TomTom, Garmin, FAO, NOAA, USGS, © OpenStreetMap contributors, and the GIS User Community.
Pixel-wise confusion maps over a hand-edited micro-AOI reveal three interpretable, repeatable error modes (Fig 6): false positives over bright construction/storage yards and lateritic exposures (spectral/texture look-alikes to benches and spoil piles in VNIR/SWIR), false negatives for small (< 0.5 ha), vegetated, or water-filled pits (occlusion and size effects). Boundary raggedness along terraced faces (sub-pixel mixing at 10 m).
Map was created using ArcGIS (version: 3.5) under the institutional ESRI site license of the University of Kassel, Germany (http://www.arcgis.com/). Basemap satellite image accessed from World Imagery ESRTI Tile Layer. Credits: Esri, Maxar, Earthstar Geographics, TomTom, Garmin, FAO, NOAA, USGS, © OpenStreetMap contributors, and the GIS User Community.
In the micro-AOI, PSPNet produced the most compact TP cores with fewer FN halos; FCN was thinner and more fragmentary; U-Net occasionally over-segmented rock outcrops at quarry margins. The three-panel failure in Fig 6 illustrates typical cases we observed.
Applying the best model (PSPNet) across the rural and urban administrative boundaries of Bengaluru yielded 252 granite quarries 227 of which were confirmed via field checks and high-resolution imagery, covering an estimated 740 hectares (Fig 7). This validation resulted in a confirmation rate of 90%, reinforcing the model’s reliability in identifying quarries in previously unobserved regions. Among confirmed sites, 124 quarries were classified as abandoned based on indicators such as water accumulation, vegetation regrowth, and lack of recent excavation – an approach similar to post-classification refinement strategies used in mining land-use studies [23,87]. Conversely, 103 quarries displayed ongoing mining activities, evidenced by exposed rock, visible machinery, and detectable site alterations in sequential imagery, paralleling multi-temporal validation practices employed in mining footprint detection [24].
(B) Heatmap of granite quarry density. It was created using ArcGIS (version: 3.5) under the institutional ESRI site license of the University of Kassel, Germany.
3.2. 3D model and volumetric analysis
The 3D model of the granite quarry site (Fig 8) effectively captured the intricate terrain over an area of approximately 1.43 km2 with a Ground Sampling Distance (GSD) of 4.6 cm. The project successfully calibrated and geolocated all 343 images, achieving a mean reprojection error of just 0.166 pixels, underscoring the model’s precision. The resulting dense point cloud consisted of over 55 million points, with an average density of 30.4 points per cubic meter, facilitating the creation of a detailed digital surface model (DSM) and orthomosaic. Recent advances in geometry-aware point-cloud learning demonstrate centimeter-scale localization accuracy in similarly cluttered scenes [88]. This underscores the suitability of high-density point clouds for downstream quantitative tasks such as our cut-and-fill volume estimation.
To this end authors’ drone images taken in July 2023 were mosaicked.
Despite the absence of GCPs, georeferencing was accurate reflecting optimized internal camera parameters [89].
The volumetric analysis of two extraction areas (Fig 9) within the granite quarry yielded quantitative insights into material removal and displacement after coregistration and parametric, slope-dependent LoD95, the DoD acceptance mask covered 96.8–98.1% of the mapped area across sites and resolutions. At the native UAV grid (0.046 m) the lowering components and their correlated 95% confidence intervals (CI) are presented in Table 4.
Repeating the full workflow at 1 m and 5 m yields lowering totals within +0.23–0.38% of the 0.046 m baseline (Table R1). Specifically, the combined lowering is 9 315 638 m³ at 1 m (+35 587 m³, + 0.38%) and 9 301 761 m³ at 5 m (+21 710 m³, + 0.23%). Correlated 95% CIs remain small (0.20–0.38% if lower) indicating that resampling contributes negligibly to the overall uncertainty compared with vertical accuracy and residual coregistration variance.
A MC jitter test at 1 m (sub-pixel dx/dy and dz drawn from the measured coreg residuals) yielded following a 95% CI: Site 1: ± 50 422 m³; Site 2: ± 27 722 m³; Combined. ± 77 970 m³ (≈ 0.84% of the MC mean lowering; Table 5). This independently corroborates the LoD-propagated CIs and shows that residual coreg errors have a minor impact; Table 6.
Based on an assumed average rock density of 2.7 g cm-3 the mass of the extracted granite amounts to 25.1 million metric tons (Mt); the 95% CI on mass is ± 0.048 Mt (scales directly with the 0.19% lowering CI). Laboratory and regional gravity data show that granites of the Peninsular Gneissic Complex span 2.60–2.75 g cm--3 (mean 2.65 g cm-3) depending on feldspar–quartz ratio and degree of alteration [90]. Applying these end members to the extracted volume yields a mass range of 24.13–25.52 Mt, i.e., - 3.7% to +1.8% around the 2.70 g cm-3 estimate. This density-driven uncertainty has been included in Table 5 and the overall error budget. Future surveys combining drill-core density measurement and high-resolution gravity data could further constrain this parameter.
Our UAV-derived volume estimate represents a significant improvement in precision compared to data derived from traditional survey methods. Interestingly, the volume distribution was not uniform across the extraction area, with notably higher concentrations in the northeastern quadrant. This non-uniformity likely stems from the site’s unique geological features and the strategic extraction patterns employed by the quarry operators. The 3D model also revealed subtle variations in the quarry floor topography, contributing to a more accurate volume calculation by accounting for undulations that might have been overlooked in conventional 2D surveys.
3.3. Limitations and future research
PSPNet, despite its sophisticated pyramid pooling architecture, exhibits notable limitations in precisely delineating quarry boundaries, particularly in complex scenarios [91] involving abandoned mining sites. While the network excels at capturing multi-scale contextual information through its pyramid structure, it apparently struggles with detailed boundary information and smooth edges in challenging landscapes. This limitation became particularly evident at abandoned quarry sites where water accumulation interrupted the typical spectral signature of granite exposures.
For future work, it may be beneficial to expand the dataset used for training and validation hereby including a broader range of geographical areas and mining site characteristics. This could involve incorporating data from different regions or countries to improve the model’s adaptability. Also, transfer learning techniques could help leverage existing models trained on similar tasks to enhance performance in new contexts [92].
Another avenue for improvement is the integration of temporal analysis using time-series satellite imagery to monitor changes in quarry sites over time. This approach could provide valuable insights into mining activities and their environmental impacts. However, it might face significant practical challenges. The historical limitations of satellite data, with Landsat imagery only available from the 1990s onwards, coupled with a substantial gap in Google Earth coverage for Bengaluru between 1985 and 2011, create significant obstacles for comprehensive temporal analysis. This is particularly problematic as many regional quarries predate such temporal datasets, and the absence of consistent historical imagery makes it difficult to verify quarry evolution patterns.
The volume estimation of the granite quarry using advanced 3D modeling techniques faces several limitations that merit consideration. The absence of ground control points (GCPs) introduces potential uncertainties in absolute georeferencing accuracy, particularly for large-scale projects. Although a dense network of GCPs remains the gold standard for absolute georeferencing, in unregulated or artisanal quarry settings the installation, and repeated surveying of targets can be impractical or even unsafe due to unstable slopes, active blasting, and access restrictions. To reduce the absolute georeferencing error inherent in our GCP-free workflow, future campaigns will employ RTK-PPK-enabled UAVs, which adding it with a minimal GCP network improves vertical accuracy from decimeter to just ±2–4 mm and overall 3-D RMSE to <5 mm [93].
Moreover, the use of DEM data, with its relatively coarse resolution (30 m), potentially misses subtle topographic features crucial for accurate volume calculations in complex quarry terrains. Quarry change detection will also move from single-epoch snapshots to quarterly UAV campaigns, enabling sub-decimeter SfM point-cloud differencing and early warning of slope instabilities.
The lack of historical extraction records and reference data hinders the validation of the model’s accuracy. This limitation reflects a common challenge in studies involving industrial sites with complex histories or changes in operational management [94].
4. Conclusions
Our study demonstrates the effectiveness of integrating deep learning techniques with multi-temporal satellite imagery analysis and UAV photogrammetry to monitor granite quarrying activities. The PSPNet-based deep learning model achieved high accuracy in detecting and segmenting granite quarries, providing a reliable tool for large-scale monitoring of mining activities. Identifying 227 granite quarries covering 740 hectares highlights the extensive nature of quarrying in the surroundings of the southern Indian megacity of Bengaluru region and its potential environmental impacts.
The high-precision volume estimation using UAV photogrammetry offers valuable insights into the scale of granite extraction, with implications for resource management and environmental assessment. This integrated approach provides a robust framework for monitoring quarrying activities, offering decision-makers and environmental managers a powerful tool for sustainable land use planning and policy formulation.
Future research should focus on expanding the geographical scope of such studies, incorporating temporal analysis to monitor changes over time, and exploring the integration of additional data sources to enhance the accuracy and applicability of the model. This comprehensive approach to quarry monitoring has the potential to significantly contribute to sustainable mining practices and environmental conservation efforts in Bengaluru and beyond.
Acknowledgments
This study was inspired by and associated to the interdisciplinary German-Indo-Moroccan-Ghanaian Research Unit “Sustainable Rurbanity - Resources, Society, and Regulatory Systems” (FOR5903) supported by the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG).
References
- 1. Rochette P, Sciuto C, Raneri S, Sutter A, Rovazzani A, Folco L, et al. Non-destructive magnetic and chemical characterization of granite column shafts traded in the Mediterranean area: the case of Piazza dei Miracoli in Pisa (Italy) and Basilica of Saint-Martin d’Ainay in Lyon (France). J Phys: Conf Ser. 2022;2204(1):012037.
- 2. Freire-Lista DM, Campos BB, Moreira P, Ramil A, López AJ. Building Granite Characterisation, Construction Phases, Mason’s Marks and Glyptography of Nossa Senhora de Guadalupe Church, Mouçós e Lamares, Galicia-North Portugal Euroregion. Geoheritage. 2023;15(1).
- 3. Ningappa C, Sannappa J, Chandrashekara MS, Paramesh L. Studies on radon/thoron and their decay products in granite quarries around Bangalore city, India. Indian J Phys. 2009;83(8):1201–7.
- 4. Nithya* P, Arulselvi DrG. Pattern Classification Technique to Assess Land Use/Cover Changes in Granite Quarry Area of Dharmapuri and Krishnagiri Districts of Tamilnadu. IJITEE. 2019;8(12):2947–56.
- 5. Lawal JI. Natural Radionuclides Content in Granites from Operational Quarry Sites. Detection. 2019;07(01):1–15.
- 6. Omar M, Abdelrhman Y, M. Hassab I, Khierldeen WM. Experimental Study On Compressive Strength And Flexural Rigidity Of Epoxy Granite Composite Material. JES Journal of Engineering Sciences. 2021;49(2):198–214.
- 7. Kareem Salati L, Shola Adeyemo M. Examining the Effect of Powder Factor Variability on Granite Productivity. Recent Advances in Mineralogy. IntechOpen. 2023.
- 8. Darwish T, Khater C, Jomaa I, Stehouwer R, Shaban A, Hamzé M. Environmental impact of quarries on natural resources in Lebanon. Land Degrad Dev. 2011;22(3):345–58.
- 9. Quintela-Sabarís C, Masfaraud J-F, Séré G, Sumail S, van der Ent A, Repin R, et al. Effects of reclamation effort on the recovery of ecosystem functions of a tropical degraded serpentinite dump site. Journal of Geochemical Exploration. 2019;200:139–51.
- 10. Mendoza-García M, Godínez-Alvarez H. Spontaneous vegetation succession and recovery of ecosystem structure and function in a 40-year abandoned stone quarry in a Mexican tropical dryland. Bot Sci. 2021;100(1):86–92.
- 11. Capitano C, Peri G, Rizzo G, Ferrante P. Toward a holistic environmental impact assessment of marble quarrying and processing: proposal of a novel easy-to-use IPAT-based method. Environ Monit Assess. 2017;189(3):108. pmid:28210890
- 12. Adedeji O, Olayinka O, Tope-Ajayi OO, Azeez O. Land use/land cover changes due to quarrying in Odeda local government area of Ogun State, Nigeria: An assessment and implication for rural livelihood. https://www.semanticscholar.org/paper/land-use-land-cover-changes-due-to-quarrying-in-of-Adedeji-Olayinka/6857a44451e8e266dbb0d2b9f8f2aa3c97c720da. 2020.
- 13. Jill Wells JW. Environmental concerns and responses in small-scale stone quarries in Nairobi. Enterprise Development & Microfinance. 2000;11(2):28–38.
- 14. Buerkert A, Ganeshaiah KN, Siebert S. Water Use In Human Civilizations: An Interdisciplinary Analysis Of A Perpetual Social-ecological Challenge. Front Agr Sci Eng. 2021;8(4):512.
- 15. Götze C, Beyer F, Gläßer C. Pioneer vegetation as an indicator of the geochemical parameters in abandoned mine sites using hyperspectral airborne data. Environ Earth Sci. 2016;75(7).
- 16. Saedpanah S, Amanollahi J. Environmental pollution and geo-ecological risk assessment of the Qhorveh mining area in western Iran. Environ Pollut. 2019;253:811–20. pmid:31344542
- 17. Dwyer JL, Roy DP, Sauer B, Jenkerson CB, Zhang HK, Lymburner L. Analysis Ready Data: Enabling Analysis of the Landsat Archive. Remote Sensing. 2018;10(9):1363.
- 18. Werner TT, Bebbington A, Gregory G. Assessing impacts of mining: Recent contributions from GIS and remote sensing. The Extractive Industries and Society. 2019;6(3):993–1012.
- 19. Blaschke T, Hay GJ, Kelly M, Lang S, Hofmann P, Addink E, et al. Geographic Object-Based Image Analysis - Towards a new paradigm. ISPRS J Photogramm Remote Sens. 2014;87(100):180–91. pmid:24623958
- 20. Huang L, You J, Zhu X, Shuai T, Liao Y. Object-oriented polarimetric SAR image classification via the combination of a pixel-based classifier and a region growing technique. European Journal of Remote Sensing. 2023;56(1).
- 21. Blaschke T. Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing. 2010;65(1):2–16.
- 22. Kucharczyk M, Hay GJ, Ghaffarian S, Hugenholtz CH. Geographic Object-Based Image Analysis: A Primer and Future Directions. Remote Sensing. 2020;12(12):2012.
- 23. Gallwey J, Robiati C, Coggan J, Vogt D, Eyre M. A Sentinel-2 based multispectral convolutional neural network for detecting artisanal small-scale mining in Ghana: Applying deep learning to shallow mining. Remote Sensing of Environment. 2020;248:111970.
- 24. Ren Z, Wang L, He Z. Open-Pit Mining Area Extraction from High-Resolution Remote Sensing Images Based on EMANet and FC-CRF. Remote Sensing. 2023;15(15):3829.
- 25. Wang K, Zhang Z, Yang X, Wang D, Zhu L, Yuan S. Enhanced Tailings Dam Beach Line Indicator Observation and Stability Numerical Analysis: An Approach Integrating UAV Photogrammetry and CNNs. Remote Sensing. 2024;16(17):3264.
- 26. Zhang H, Aldana-Jague E, Clapuyt F, Wilken F, Vanacker V, Van Oost K. Evaluating the potential of post-processing kinematic (PPK) georeferencing for UAV-based structure- from-motion (SfM) photogrammetry and surface change detection. Earth Surf Dynam. 2019;7(3):807–27.
- 27. Ferrer-González E, Agüera-Vega F, Carvajal-Ramírez F, Martínez-Carricondo P. UAV Photogrammetry Accuracy Assessment for Corridor Mapping Based on the Number and Distribution of Ground Control Points. Remote Sensing. 2020;12(15):2447.
- 28. Szypuła B. Accuracy of UAV-based DEMs without ground control points. Geoinformatica. 2023;28(1):1–28.
- 29. Westoby MJ, Brasington J, Glasser NF, Hambrey MJ, Reynolds JM. ‘Structure-from-Motion’ photogrammetry: A low-cost, effective tool for geoscience applications. Geomorphology. 2012;179:300–14.
- 30. Raeva PL, Filipova SL, Filipov DG. Volume Computation Of A Stockpile – A Study Case Comparing Gps And Uav Measurements In An Open Pit Quarry. Int Arch Photogramm Remote Sens Spatial Inf Sci. 2016;XLI-B1:999–1004.
- 31. James MR, Antoniazza G, Robson S, Lane SN. Mitigating systematic error in topographic models for geomorphic change detection: accuracy, precision and considerations beyond off‐nadir imagery. Earth Surf Processes Landf. 2020;45(10):2251–71.
- 32. White CT, Reckling W, Petrasova A, Meentemeyer RK, Mitasova H. Rapid-DEM: Rapid Topographic Updates through Satellite Change Detection and UAS Data Fusion. Remote Sensing. 2022;14(7):1718.
- 33.
Indian Bureau of Mines IB. Indian mineral industry at a glance 2020-21. 40th edition. Controller General Indian Bureau of Mines. 2021. https://ibm.gov.in/writereaddata/files/1714040600662a2f181f02eIMIG_202021_Final_Copy_dt.18.04.2024.pdf
- 34. Bongale P, Kshirsagar MA. Geology of the Western Dharwar Craton Ms. 2015. https://www.semanticscholar.org/paper/Geology-of-the-Western-Dharwar-Craton-Ms-Bongale-Kshirsagar/86f564e8f9115c3e15b57e5c5f88d08efbf5f6fe
- 35. Münzel T, Hahad O, Daiber A, Landrigan PJ. Soil and water pollution and human health: what should cardiologists worry about?. Cardiovasc Res. 2023;119(2):440–9. pmid:35772469
- 36. Ganiyah H-GO, Onoshagbe OB, Joseph AO. Assessment of Heavy Metal Pollution from Granite Quarrying in Akure, Nigeria. JOGEE. 2023;19(3–4):78–86.
- 37. Molders N. Landscape changes over a region in east Germany and their impact upon the processes of its atmospheric water-cycle. Meteorl Atmos Phys. 1998;68(1–2):79–98.
- 38. Phillips J. Climate change and surface mining: A review of environment-human interactions & their spatial dynamics. Applied Geography. 2016;74:95–108.
- 39. Vandana M, John SE, Maya K, Sunny S, Padmalal D. Environmental impact assessment (EIA) of hard rock quarrying in a tropical river basin-study from the SW India. Environ Monit Assess. 2020;192(9):580. pmid:32785784
- 40. Manzoor J, Khan M. Environmental Perspective of Stone Crushing and Quarrying: A Case Study from Poonch District of Jammu and Kashmir, India. Current World Environment. 2020;15(1):68–74.
- 41. Uuemaa E, Ahi S, Montibeller B, Muru M, Kmoch A. Vertical Accuracy of Freely Available Global Digital Elevation Models (ASTER, AW3D30, MERIT, TanDEM-X, SRTM, and NASADEM). Remote Sensing. 2020;12(21):3482.
- 42. Meadows M, Jones S, Reinke K. Vertical accuracy assessment of freely available global DEMs (FABDEM, Copernicus DEM, NASADEM, AW3D30 and SRTM) in flood-prone environments. International Journal of Digital Earth. 2024;17(1).
- 43. Carabajal CC, Harding DJ. ICESat validation of SRTM C‐band digital elevation models. Geophysical Research Letters. 2005;32(22).
- 44. O’Loughlin FE, Paiva RCD, Durand M, Alsdorf DE, Bates PD. A multi-sensor approach towards a global vegetation corrected SRTM DEM product. Remote Sensing of Environment. 2016;182:49–59.
- 45. Jan N, Minallah N, Gohar N, Jan N, Khan S, Khan S, et al. Granite Exposure Mapping Through Sentinel‐2 Visible and Short Wave Infrared Bands. Radio Science. 2024;59(2). https://doi.orCg/10.1029/2023rs007864
- 46. Zupanc A. Improving cloud detection with machine learning. Planet Stories. https://medium.com/sentinel-hub/improving-cloud-detection-with-machine-learning-c09dc5d7cf13. 2020. Accessed 2025 August 30
- 47. Vaiopoulos AD, Karantzalos K. Pansharpening On The Narrow Vnir And Swir Spectral Bands Of Sentinel-2. Int Arch Photogramm Remote Sens Spatial Inf Sci. 2016;XLI-B7:723–30.
- 48. Turgeon-Pelchat M, Papasodoro C, Moisan Y, Gravel P. Toward the use of deep learning and high-resolution digital elevation models for extracting geological landforms: an example for eskers. https://www.semanticscholar.org/paper/Toward-the-use-of-deep-learning-and-high-resolution-Turgeon-Pelchat-Papasodoro/2e613b536933945dd33847f8b93d5a6ca6545182. 2019.
- 49. Chen Y, Dong Y, Wang Y, Zhang F, Liu G, Sun P. Machine learning algorithms for lithological mapping using Sentinel-2 and SRTM DEM in highly vegetated areas. Front Ecol Evol. 2023;11.
- 50. Optical Imagery. https://www.maxar.com/maxar-intelligence/products/optical-imagery. Accessed 2025 August 30
- 51. Anwar H, Zambanini S, Kampel M. Efficient Scale- and Rotation-Invariant Encoding of Visual Words for Image Classification. IEEE Signal Process Lett. 2015;22(10):1762–5.
- 52. Yu X, Wu X, Luo C, Ren P. Deep learning in remote sensing scene classification: a data augmentation enhanced convolutional neural network framework. GIScience & Remote Sensing. 2017;54(5):741–58.
- 53. Nesteruk S, Shadrin D, Pukalchik M. Image augmentation for multitask few-shot learning: Agricultural domain use-case. In: scite.ai, 2021.
- 54. Dronelink Inc. Dronelink. https://www.dronelink.com/. 2024.
- 55. Eisenbeiss H, Sauerbier M. Investigation of uav systems and flight modes for photogrammetric applications. The Photogrammetric Record. 2011;26(136):400–21.
- 56. Rosnell T, Honkavaara E. Point cloud generation from aerial image data acquired by a quadrocopter type micro unmanned aerial vehicle and a digital still camera. Sensors (Basel). 2012;12(1):453–80. pmid:22368479
- 57. Ai M, Hu Q, Li J, Wang M, Yuan H, Wang S. A Robust Photogrammetric Processing Method of Low-Altitude UAV Images. Remote Sensing. 2015;7(3):2302–33.
- 58. Pix4D SA. Pix4Dmapper. https://www.pix4d.com/pix4dmapper. 2023.
- 59. Minh NQ, Huong NTT, Khanh PQ, Hien LP, Bui DT. Impacts of Resampling and Downscaling Digital Elevation Model and Its Morphometric Factors: A Comparison of Hopfield Neural Network, Bilinear, Bicubic, and Kriging Interpolations. Remote Sensing. 2024;16(5):819.
- 60. Saputra MRU, Bhaswara ID, Nasution BI, Ern MAL, Husna NLR, Witra T, et al. Multi-modal deep learning approaches to semantic segmentation of mining footprints with multispectral satellite imagery. Remote Sensing of Environment. 2025;318:114584.
- 61. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Cham: Springer International Publishing; 2015:234–241.
- 62. Zhao H, Shi J, Qi X, Wang X, Jia J. Pyramid scene parsing metwork. In: 2017. 2881–90. https://openaccess.thecvf.com/content_cvpr_2017/html/Zhao_Pyramid_Scene_Parsing_CVPR_2017_paper.html
- 63. Chen LC, Papandreou G, Schroff F, Adam H. Rethinking atrous convolution for semantic image segmentation. arXiv. 2017.
- 64. Chen L-C, Zhu Y, Papandreou G, Schroff F, Adam H. Encoder-Decoder with atrous separable convolution for semantic image segmentation. arXiv; 2018.
- 65. Li S, Wang R, Wang L, Liu S, Ye J, Xu H, et al. An Approach for Monitoring Shallow Surface Outcrop Mining Activities Based on Multisource Satellite Remote Sensing Data. Remote Sensing. 2023;15(16):4062.
- 66. Long J, Shelhamer E, Darrell T. Fully Convolutional Networks for Semantic Segmentation. arXiv. 2015.
- 67. Li X, Zhong Z, Wu J, Yang Y, Lin Z, Liu H. Expectation-maximization attention networks for semantic segmentation. 2019.
- 68. Guo B, Chen S, Hong Z, Xu G. Pattern Recognition and Analysis: Neural Network using Weighted Cross Entropy. J Phys: Conf Ser. 2022;2218(1):012043.
- 69. Zhou C, Wang H, Liu Y, Ni X, Liu Y. Green Plums Surface Defect Detection Based on Deep Learning Methods. IEEE Access. 2022;10:100397–407.
- 70. Hoffer E, Hubara I, Soudry D. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. In: scite.ai, 2017.
- 71. Wang Z, Li X, Yao M, Li J, Jiang Q, Yan B. A new detection model of microaneurysms based on improved FC-DenseNet. Sci Rep. 2022;12(1):950. pmid:35046432
- 72. Brown LD, Cai TT, DasGupta A. Interval Estimation for a Binomial Proportion. Statist Sci. 2001;16(2).
- 73. McNemar Q. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika. 1947;12(2):153–7. pmid:20254758
- 74. Belgiu M, Dr Guţ L, Strobl J. Quantitative evaluation of variations in rule-based classifications of land cover in urban neighbourhoods using WorldView-2 imagery. ISPRS J Photogramm Remote Sens. 2014;87(100):205–15. pmid:24623959
- 75. Olofsson P, Foody GM, Herold M, Stehman SV, Woodcock CE, Wulder MA. Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment. 2014;148:42–57.
- 76. Gong K, Fritsch D. A Detailed Study About Digital Surface Model Generation Using High Resolution Satellite Stereo Imagery. ISPRS Ann Photogramm Remote Sens Spatial Inf Sci. 2016;III–1:69–76.
- 77. Gong K, Fritsch D. Point Cloud And Digital Surface Model Generation From High Resolution Multiple View Stereo Satellite Imagery. Int Arch Photogramm Remote Sens Spatial Inf Sci. 2018;XLII–2:363–70.
- 78. Nuth C, Kääb A. Co-registration and bias corrections of satellite elevation data sets for quantifying glacier thickness change. The Cryosphere. 2011;5(1):271–90.
- 79. Lane SN, Westaway RM, Murray Hicks D. Estimation of erosion and deposition volumes in a large, gravel‐bed, braided river using synoptic remote sensing. Earth Surf Processes Landf. 2003;28(3):249–71.
- 80. Wheaton JM, Brasington J, Darby SE, Sear DA. Accounting for uncertainty in DEMs from repeat topographic surveys: improved sediment budgets. Earth Surf Processes Landf. 2009;35(2):136–56.
- 81. James MR, Robson S, d’Oleire-Oltmanns S, Niethammer U. Optimising UAV topographic surveys processed with structure-from-motion: Ground control quality, quantity and bundle adjustment. Geomorphology. 2017;280:51–66.
- 82. Long J, Shelhamer E, Darrell T. Fully Convolutional Networks for Semantic Segmentation. 2015.
- 83. Balaniuk R, Isupova O, Reece S. Mining and Tailings Dam Detection in Satellite Imagery Using Deep Learning. Sensors (Basel). 2020;20(23):6936. pmid:33291634
- 84. Couttenier M, Di Rollo S, Inguere L, Mohand M, Schmidt L. Mapping artisanal and small-scale mines at large scale from space with deep learning. PLoS One. 2022;17(9):e0267963. pmid:36136980
- 85. Jabłońska K, Maksymowicz M, Tanajewski D, Kaczan W, Zięba M, Wilgucki M. MineCam: Application of Combined Remote Sensing and Machine Learning for Segmentation and Change Detection of Mining Areas Enabling Multi-Purpose Monitoring. Remote Sensing. 2024;16(6):955.
- 86. Wang Y, Qin K, Zhang Z, He Q, Cohen J. Mapping open-pit mining area in complex mining and mixed land cover zone using Landsat imagery. International Journal of Applied Earth Observation and Geoinformation. 2024;129:103782.
- 87. Feng X, Wei C, Xue X, Zhang Q, Liu X. RST-DeepLabv3+: Multi-Scale Attention for Tailings Pond Identification with DeepLab. Remote Sensing. 2025;17(3):411.
- 88. Wang H, Zhang G, Cao H, Hu K, Wang Q, Deng Y, et al. Geometry‐Aware 3D Point Cloud Learning for Precise Cutting‐Point Detection in Unstructured Field Environments. Journal of Field Robotics. 2025;42(7):3063–76.
- 89. Przybilla H-J, Lindstaedt M, Kersten T. Investigations Into The Quality Of Image-based Point Clouds From Uav Imagery. Int Arch Photogramm Remote Sens Spatial Inf Sci. 2019;XLII-2/W13:539–45.
- 90. Densities and magnetic susceptibilities of precambrian rocks of different metamorphic grade (Southern Indian shield). International Journal of Rock Mechanics and Mining Sciences & Geomechanics Abstracts. 1981;18(4):67.
- 91. Chen Y, Yang Z, Zhang L, Cai W. A semi-supervised boundary segmentation network for remote sensing images. Sci Rep. 2025;15(1):2007. pmid:39814791
- 92. Kurian V, Jacob V, Kuruvilla J. Approach of Transfer Learning in Remote Sensing Image Classification. In: 2024 1st International Conference on Trends in Engineering Systems and Technologies (ICTEST), 2024:1–3.
- 93. Dlamini SM, Ouma YO. Large-Scale Topographic Mapping Using RTK-GNSS and Multispectral UAV Drone Photogrammetric Surveys: Comparative Evaluation of Experimental Results. Geomatics. 2025;5(2):25.
- 94. Aziz A, Schelén O, Bodin U. A Study on Industrial IoT for the Mining Industry: Synthesized Architecture and Open Research Directions. IoT. 2020;1(2):529–50.