Available via license: CC BY 4.0
Content may be subject to copyright.
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2024.0429000
Automated Detection of Chilean Mine Waste
Storage Facilities Using Advanced Deep Learning
Models and Sentinel-2 Satellite Imagery
MANUEL SILVA1, (Member, IEEE), GABRIEL HERMOSILLA1, GABRIEL VILLAVICENCIO2,
GIOVANNI COCCA-GUARDIA1(Member, IEEE), PIERRE BREUL3, VICENTE APRIGLIANO2
1Escuela de Ingeniería Eléctrica, Pontificia Universidad Católica de Valparaíso, Valparaíso 2374631, Chile
2Escuela de Ingeniería de Construcción y Transporte, Pontificia Universidad Católica de Valparaíso, Valparaíso 2362804, Chile
3Département Génie Civil, Polytech Clermont, Institut Pascal UMR CNRS 6602, Université Clermont Auvergne, Av. Blaise Pascal SA 60206-63178 Aubière,
CEDEX, 63000 Clermont Ferrand, France
Corresponding author: Gabriel Hermosilla (e-mail: gabriel.hermosilla@pucv.cl).
This research was funded and supported by the Vice-Rectorate for Research, Creation, and Innovation (VINCI) at Pontificia Universidad
Católica de Valparaíso (Chile), through the Associative Research Project under grant: 039.301/2024, Centennial Project 2024 under grant:
039.310/2024, and by the Chilean National Agency of Research and Development (ANID) through Fondo Nacional de Desarrollo
Científico y Tecnológico (FONDECYT) under Grant 1240573, and ANID Doctorado Nacional under grant: 2023-21232328
ABSTRACT Mine Waste Storage Facilities (MWSFs) in Chile present substantial environmental and safety
risks due to their extensive scale and the hazardous nature of their contents. This study proposes an automated
detection approach that integrates Sentinel-2 satellite imagery with advanced deep learning models to address
these critical issues. A central contribution of this research is the development of MineWasteCL_DB, a
comprehensive public dataset comprising over 30,000 annotated images and 320,093 labels for diverse
MWSF types, including Tailings Storage Facilities (TSFs), Waste Rock Dumps (WRDs), and Leaching
Waste Dumps (LWDs). The study employs the YOLOv8x-seg model, selected for its high precision, to
validate the presence of 96.15% of officially registered TSFs. Furthermore, it identified 141 WRDs and
112 LWDs in the Antofagasta Refgion, facilities absent from any official national registry. These findings
underscore the methodology’s potential for widespread application and the necessity for routine monitoring
across additional regions. The results provide a robust framework for advancing the understanding and
management of MWSFs, thereby improving regulatory oversight and promoting environmental safety. The
methodology supports not only the efficient monitoring of registered facilities but also the preliminary
identification and prospective registration of unregistered sites. This capability enhances the oversight
capacities of regulatory authorities while fostering the protection of environmental and public safety.
INDEX TERMS automated monitoring, deep learning, detection, instance segmentation, leaching waste
dumps, Sentinel-2, tailings storage facilities, waste rock dumps
I. INTRODUCTION
THE mining industry has experienced substantial trans-
formations, particularly in growth and digitalization,
influencing areas such as mineral exploration, production,
safety, and environmental impact [1]. Between 2018 and
2022, global copper production increased by 6.7% to meet
rising demand, with projections estimating a 45.9% surge in
demand between 2021 and 2040 [2]. This expansion signifies
not only increased production but also a considerable rise in
mining waste generation, posing critical environmental and
safety challenges. Chile, as one of the world’s leading cop-
per producers, is expected to generate approximately 915.4
million tons of tailings and 146 million tons of waste rock
annually by 2026 [3]. These materials are disposed of in Mine
Waste Storage Facilities (MWSFs), which pose significant
environmental risks due to their large volumes and potentially
toxic content.
MWSFs include Waste Rock Dumps (WRDs), Leach-
ing Waste Dumps (LWDs), and Tailings Storage Facilities
(TSFs). WRDs can reach heights between 50 and 500 m,
with projections extending to 1,000 m for large mining
projects, spanning hundreds of hectares. LWDs are smaller,
with heights between 10 and 120 m and areas covering several
hectares, while TSFs may exceed 250 m in height and occupy
VOLUME 11, 2023 1
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
tens of hectares. TSFs, in particular, present heightened risks
of environmental disasters due to physical instability mecha-
nisms that can release stored materials, impacting extensive
downstream areas Studies indicate an increase in TSF failures
globally from 1915 to 2021, emphasizing the urgent need to
address their stability [4]–[8].
Ensuring the physical stability of MWSFs during con-
struction, operation, closure, and post-closure is critical to
minimizing risks to populations and the environment in com-
pliance with regulations [9]–[13]. In Chile, most MWSFs are
located in the regions of Tarapacá, Antofagasta, Atacama, Co-
quimbo, Valparaíso, Metropolitana, O’Higgins, Maule, and
Aysén. However, while 796 TSFs are registered nationally
[14], there are no official registries for WRDs and LWDs,
complicating the oversight responsibilities of the National
Geology and Mining Service (SERNAGEOMIN), especially
given limited resources and personnel. Figure 1 illustrates
examples of MWSFs at mining facilities in the Antofagasta
region.
Remote sensing technologies, particularly satellite im-
agery, have revolutionized MWSF monitoring by enabling
efficient extraction of their geometric characteristics [15].
Traditional detection methods, however, often lack precision
and efficiency when analyzing extensive datasets and com-
plex patterns. Advanced Artificial Intelligence (AI) and Deep
Learning (DL) techniques offer promising solutions to these
limitations. Models such as Swin Transformer, YOLOv8x-
seg, Mask2Former, and Mask R-CNN exhibit exceptional
capabilities for handling large datasets and detecting intri-
cate patterns with high precision. Swin Transformer effec-
tively captures spatial relationships at multiple scales, while
YOLOv8x-seg excels in rapid and accurate object detection.
Mask2Former and Mask R-CNN are particularly effective for
detailed segmentation of MWSFs, enabling comprehensive
analysis of their structures.
This study pursues two primary objectives. First, it aims to
compile a comprehensive and temporal database of MWSFs
in Chile using open-source Sentinel-2 satellite imagery, cre-
ating a resource of over 30,000 images for future research
and for SERNAGEOMIN’s detection and registration efforts.
Second, it develops and optimizes an advanced instance seg-
mentation approach, testing diverse models, backbones, seg-
mentation heads, and configurations. Iterative hyperparame-
ter tuning and evaluation through standard metrics such as
precision, IoU, and recall identified the most effective model
for the dataset.
The trained models were applied to the Antofagasta Re-
gion, one of Chile’s most significant mining zones [16],
to compare automatic segmentation results with SERNA-
GEOMIN’s official TSF cadastre [14]. Additionally, since
no national registry exists for WRDs and LWDs, this work
seeks to generate a new cadastre for these deposits, offering
a preliminary database for future verification by SERNA-
GEOMIN. This contribution enhances monitoring and man-
agement capabilities, providing critical data for risk assess-
ment and environmental planning.
Building on these objectives, this study presents an inte-
grated framework that combines deep learning architectures
with Sentinel-2 satellite imagery for MWSF detection and
segmentation. By integrating automated detection algorithms
with systematic validation against SERNAGEOMIN records,
the framework ensures the comprehensive identification of
both registered and undocumented MWSFs in the Antofa-
gasta region. Empirical results demonstrate the framework’s
ability to augment existing monitoring mechanisms, estab-
lishing a robust analytical tool for improved oversight of
mining waste management systems.
II. RELEATED WORK
The detection and monitoring of MWSFs have been the fo-
cus of numerous studies seeking to enhance precision and
efficiency through advanced technologies. For instance, [17]
ntroduces a dataset derived from Sentinel-2 and Landsat-
8 satellite images, encompassing tailings storage facilities
cataloged by the Brazilian National Mining Agency. This
study integrates data on mine type, construction method, risk
category, and associated potential damage, achieving a preci-
sion of 94.11% in binary TSF classification. While this under-
scores the value of combining geospatial data with machine
learning models, its scope is limited to binary classification,
lacking detailed segmentation. In contrast, our study advances
beyond binary classification by implementing sophisticated
instance segmentation models. These models enable precise
spatial delineation and volume estimation of MWSFs, pro-
viding crucial insights for environmental monitoring and risk
assessment by offering more granular data on facility status
and potential impacts.
Additionally, [18] demonstrates the use of the Google Earth
Engine platform for processing satellite images with a focus
on mining activities. By integrating machine learning algo-
rithms, it identifies and classifies mining surfaces and TSFs.
Although effective in handling large datasets, this approach is
constrained by limited customization of segmentation models
and the inability to integrate more complex algorithms. Our
methodology addresses these limitations through a flexible
architecture finely tuned for high specificity and sensitivity,
adapting to the diverse and intricate landscape of Chilean
mining regions. By leveraging localized datasets for training
and validation, our models are optimized for regional charac-
teristics, significantly enhancing the operational effectiveness
of MWSF monitoring
This work not only underscores the potential of state-of-
the-art deep learning techniques but also provides practical
solutions for monitoring MWSF physical stability in regions
with extensive mining activities. By minimizing potential
environmental risks and safeguarding nearby populations, it
contributes to sustainable mining practices and environmental
safety
Recent advancements in deep learning architectures, par-
ticularly convolutional neural networks (CNNs), have revo-
lutionized fields like medical image analysis by extracting
hierarchical features from complex datasets. These models
2VOLUME 11, 2023
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
(a) (b) (c)
FIGURE 1: Examples of MWSFs in the Antofagasta region. (a) WRD at Chuquicamata mining facility. (b) LWD at Spence
mining facility. (c) TSF at Escondida mining facility.
have demonstrated exceptional performance in critical ap-
plications, such as left ventricle segmentation in echocar-
diographic imaging, detection of dopaminergic patterns in
DaTscan imagery for Parkinson’s disease, and privacy-
preserving federated learning for epistasis detection. Stud-
ies validating these advancements include the application of
Mask R-CNN in cardiac imaging, DenseNet-121 with soft at-
tention for neurological diagnostics, and secure collaborative
learning strategies in genomics [19]–[21].
In mining appications, [22] eveloped a high-precision
method for extracting information on TSFs using the Faster
R-CNN model with high-resolution images. This approach
achieved a precision of 85.7%, demonstrating the efficacy
of deep learning models in the precise detection of MWSFs.
Building on this, [23] introduced a combination of YOLOv7
object detection with vision transformer classifiers for detect-
ing mine waste deposits, achieving 81% precision in detection
tasks. While these studies mark significant advancements
in detection capabilities, they also highlight the persistent
challenge of creating accurate and efficient systems for mon-
itoring diverse mining waste facilities.
The construction of high-resolution databases for MWSFs
was addressed in [24], where images from various satellite
sources were compiled, and spatial analyses of land use
within these facilities were conducted. Despite these advance-
ments, the study emphasized the ongoing issue of outdated
information, underscoring the need for comprehensive and
current databases to enhance environmental and safety assess-
ments of MWSFs.
In the field of image segmentation, the use of backbones
has been extensively studied. For example, [25], analyzed
various models used as backbones to extract image features
during the initial stages of training detection and segmenta-
tion models. Although these backbones have proven effective,
they involve high computational costs, and their performance
often depends on the nature of the data, limiting their appli-
cability in resource-constrained environments.
In [26], he integration of multiple backbones into a U-Net
Ensemble architecture was evaluated for segmenting images
across different datasets. The results demonstrated significant
improvements compared to baseline models, with backbones
such as MaxViT, ConvFormer, and EfficientNet outperform-
ing others by 8.32%, 6.96%, and 7.23% in mIoU, respectively.
This study highlights the potential of ensemble architectures
to enhance segmentation performance in complex scenarios.
Further research, such as [27] and [28], has shown
the widespread use of models like U-Net, ResNet50, and
DeepLabv3 for object detection and segmentation in satellite
images. These models are particularly valued for their ver-
satility in incorporating various backbones. Similarly, [29],
demonstrated the effectiveness of these methodologies for
deforestation detection, illustrating their adaptability and ef-
ficacy in a range of environmental monitoring applications.
In [30], 2,000 satellite images of MWSFs obtained from the
Sentinel-2 satellite were used to train a U-Net segmentation
model with ResNet50 as its backbone, utilizing a combination
of satellite bands. The model achieved a precision of 98%
with RGB band combinations and 96% on the test set. By
incorporating temporal samples for validation, this approach
demonstrated the high effectiveness of the methodology for
MWSF segmentation.
In contrast to previous studies centered on traditional ar-
chitectures, our work evaluates the performance of three
complementary families of state-of-the-art models for MWSF
detection: advanced convolutional neural networks (Mask R-
CNN, ConvNeXt, and ResNet variants), transformer-based
architectures (Mask2Former and Swin Transformer), and
fast detection models (YOLOv8x-seg, YOLOv9e-seg, and
YOLOv11x-seg). The integration of these diverse approaches
with an extensive dataset comprising over 30,000 labeled
satellite images from Chile’s primary mining regions enables
a comprehensive assessment of each model family’s capabil-
VOLUME 11, 2023 3
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
ities. CNNs excel in capturing edge precision and intricate
details, transformers are highly effective in managing irregu-
lar geometries and multi-scale features, and the YOLO family
demonstrates superior computational efficiency.
By combining these modern architectures with a robust
dataset, this study establishes a new benchmark framework
for the automated monitoring of mining waste, advancing the
precision and efficiency of MWSF detection and segmenta-
tion.
III. METHODOLOGY
This study developed a methodology for the automatic de-
tection of MWSFs using high-resolution satellite imagery
combined with advanced deep learning-based detection and
segmentation techniques. The Sentinel-2 satellite, with its
optimal spatial and temporal resolution, effectively captures
the dynamic nature of mining areas in Chile. As illustrated
in Figure 2, the process involved location identification, im-
age acquisition, manual labeling, training and evaluation of
segmentation models, and applying the trained models for
a comparative analysis of results against official records.
The methodology emphasized precise and reproducible seg-
mentation to provide detailed insights into the location and
characteristics of the facilities of interest.
A. GEOREFERENCING AND DATA SOURCES OF MWSF
Two primary data sources were utilized for georeferencing
MWSFs in Chile. The first source, the "2023 Cadastre of
Tailings Deposits in Chile" (CDR) maintained by SERNA-
GEOMIN, offers technical data and precise geographic co-
ordinates for each TSF. This resource facilitates exact spatial
localization and monitoring of TSFs using satellite imagery.
The second source, the "Mining Sites Atlas," is a Chilean
government tool containing information on all mining oper-
ations approved up to 2020 [31]. However, unlike the CDR,
the Atlas does not provide georeferenced locations for other
deposit types, such as WRDs and LWDs, limiting its utility
for precise mapping of these facilities.
To address this limitation, a broad search area was defined
around mining facilities listed in the Atlas to account for
potential WRD and LWD locations. This initial analysis zone
was further refined during the segmentation process, ensur-
ing comprehensive coverage of various MWSF types. This
approach enabled the inclusion of previously unregistered
deposits, enhancing the accuracy and completeness of the
analysis.
B. DATA ACQUISITION
To construct a comprehensive database of mining facilities
in Chile, 49,115 RGB satellite images (bands 04, 03, and
02) from the Sentinel-2 MSI L2A satellite were downloaded,
spanning the period from January 2017 to April 2024. These
images, with a resolution of 10 m/pixel, cover the main
mining sites from the Tarapacá Region (I) to the O’Higgins
Region (VI). A Python script automated the download pro-
cess, using coordinates from the CDR [14] and the Mining
Sites Atlas of Chile [31] to target specific areas efficiently.
Image selection adhered to strict quality criteria. Using
the Copernicus Data Space Environment API, at least two
monthly captures of each area were obtained, with a maxi-
mum cloudiness threshold of 20% per image. Images with
cloud cover obstructing mining facilities, graphic distortions,
or black areas were excluded. After filtering, 30,433 high-
quality images (61.96% of the initial set) were retained for
labeling and training segmentation models, as detailed in
Table 1.
TABLE 1: Description of the Downloaded Database
Name Images Size %
Downloaded dataset 49,115 294 GB 100
Labeled images 30,433 182.16 GB 61.96
Unlabeled images 18,682 111.84 GB 38.04
To capture both registered deposits and potentially unreg-
istered ones, such as WRDs and LWDs, a bounding box of
200×200 km was defined around each mining site, centered
at the midpoint of locations registered by SERNAGEOMIN.
This configuration ensures coverage not only of specific TSF
areas but also of surrounding zones where other MWSFs
associated with operations might be located. This approach
effectively groups discharge areas near mining facilities while
leveraging the spatial resolution of Sentinel-2. It also adheres
to the 2,500×2,500-pixel limit set by the Copernicus Data
Space Environment API, ensuring full coverage of MWSFs
without unnecessary overlaps.
All images were stored using a standardized naming
convention: "[y_min, x_min, y_max, x_max] – (date) –
band.png", indicating the georeferenced boundaries of each
image. This system ensured systematic data management,
providing an organized structure for efficient access and anal-
ysis.
C. DATA LABELING
To train the detection and segmentation models, all im-
ages acquired in the previous section were labeled to en-
sure accurate identification of mining facilities. This pro-
cess distinguished mining facilities from natural geographic
formations and other structures, such as solar plants or
nearby urban areas. The result was the creation of a new
dataset, MineWasteCL_DB, comprising 30,433 labeled im-
ages and 320,093 annotations of MWSFs, distributed as fol-
lows: LWD (155,743; 48.66%), WRD (96,789; 30.24%), and
TSF (67,561; 21.11%).
The annotations were exported in two formats: COCO
(Common Objects in Context) [32] and YOLO [33]. COCO
was chosen for its versatility in segmentation tasks, while
YOLO provides compatibility with models requiring this
format, broadening the range of applicable segmentation op-
tions. This dual-format approach ensures the interoperability
and adaptability of the labels across various models and ar-
4VOLUME 11, 2023
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
Regions of Chile
Arica y Parinacota
XV
Tarapacá
I
Antofagasta
II
Atacama
III
Coquimbo
IV
Valparaíso
V
Metropolitana
RM
O’Higgins
VI
Maule
VII
Biobío
VIII
Araucanía
IX
Los Ríos
XIV
Los Lagos
X
Aysén
XI
Magallanes
XII
Chilean Large-scale Mining
Facilities
Chile
MWSFs
Data Adquisition
MWSFs
Data Labeling
WMSF
Data Storage
Total
Satellite
Images
Downloads
(49115)
Sentinel-2
Highest Resolution Available
(10 𝑚!/𝑝𝑖𝑥𝑒𝑙)
0: WRD
1: LWD
2: TSF
Download Period
(2017-2024)
Classes:
-WRD: Waste Rock Dumps
-LWD: Leaching Waste Dumps
-TSF: Tailings Storage Facilities
AI Training
and Validation
TSFs Automatic Detection
(Antofagasta)
Segmentations
Models
Official CDR 2023
Best Segm mAP/AR
(IoU=0.50:0.95)
55.7%/46.4%
Our Best Trained Model
(96.15% Precision)
1
2
3
5
6
MWSFs
Labeled Data
Storage
Format:
-COCO
-YOLO
4
Total Labeled
Satellite Images
(30433)
ANTOFAGASTA
Legend
TSFs detected
TSFs official
FIGURE 2: Methodology workflow diagram.
chitectures, enhancing the dataset’s utility for diverse deep
learning applications.
D. OBJECT DETECTION AND INSTANCE SEGMENTATION
MODELS
Instance segmentation and object detection are computer vi-
sion techniques used to identify, detect, and delineate indi-
vidual objects in an image by providing bounding boxes and
masks for each instance of the same class. These methods are
particularly suited for analyzing MWSFs, where it is essential
to distinguish between deposits of the same category, which
are often located in close proximity and exhibit variations in
shape and size. The ability to treat each deposit as a distinct
object is critical for the precise and detailed analysis of these
structures.
To tackle the complexity of this task, three categories of
models were selected, tailored to specific requirements of
MWSF analysis. Convolutional Neural Networks (CNNs)
were chosen for their capability to achieve precise segmen-
tation of contours and complex objects, capturing edges and
shapes with high fidelity [34]–[36]. Transformer-based mod-
els were selected for their capacity to represent complex
spatial relationships and handle scale variations [37], which
are vital for identifying deposits of varying sizes and shapes
in close proximity. Additionally, the YOLO family of models
was included for its efficiency and speed in real-time object
detection [38], [39], offering a rapid and effective solution for
detecting multiple deposits simultaneously.
By combining these diverse architectures, this study ad-
dresses the visual and geometric variability of MWSFs, pro-
viding a robust framework to identify the most effective
approach for detecting and segmenting these facilities in a
challenging and complex environment.
1) CNN-based Models
1) Mask R-CNN: Built upon Faster R-CNN, this model
adds a dedicated branch for instance segmentation, en-
hancing its ability to detect and delineate individual
objects. The incorporation of RoIAlign significantly
improves spatial precision in edge detection, which
is critical for the accurate delineation of satellite im-
ages [40], [41]. With backbones like ResNet and FPN,
Mask R-CNN effectively balances precision and com-
putational efficiency, making it well-suited for detailed
MWSF segmentation.
2) ConvNeXt: A modern evolution of CNNs, ConvNeXt
integrates transformer-inspired concepts into its convo-
lutional architecture, achieving improved performance
in computer vision tasks [42]. It utilizes deep convo-
lutions for feature extraction and is designed to handle
high-resolution images. This capability is ideal for the
detailed segmentation of MWSFs, particularly when
precision in capturing spatial and geometric variations
is required.
3) ResNet50, ResNet101, and ResNet101 with Small An-
chors: These convolutional networks employ residual
connections to facilitate the training of deep archi-
tectures by mitigating gradient vanishing issues [43].
ResNet50 offers a balance between precision and com-
putational efficiency, while ResNet101 excels in de-
tecting edges and contours, which is essential for de-
lineating complex deposits. The specialized version of
ResNet101 with small anchors is optimized for detect-
ing smaller objects, making it particularly effective for
segmenting small-sized MWSFs.
VOLUME 11, 2023 5
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
2) Transformer-based Models:
1) Mask2Former: This model incorporates a masked at-
tention mechanism in its transformer decoder, enhanc-
ing segmentation by selectively focusing on relevant
areas [44]. Its architecture combines a backbone for
feature extraction, a pixel decoder, and a transformer
decoder, making it particularly effective for struc-
tures with irregular geometries, which are common in
MWSFs.
2) Swin Transformer: Utilizing a hierarchical approach
with shifted window attention, this model efficiently
identifies elements at various scales [45]. This capa-
bility is especially valuable for analyzing large-scale
satellite images, such as those of MWSFs, which ex-
hibit considerable size and complexity variations.
3) Fast Detection Models (YOLO Family)
1) YOLOv8x-seg: This version extends YOLO’s compu-
tational efficiency to instance segmentation, integrating
fast detection with precise mask generation [46]. Its
multi-scale fusion capability is particularly effective for
processing Sentinel-2 images, enabling the detection of
mining deposits of varying sizes.
2) YOLOv9e-seg: This model introduces advancements
such as Programmable Gradient Information (PGI) and
the Generalized Efficient Layer Aggregation Network
(GELAN), optimizing information flow between layers
to ensure high efficiency and adaptability [47].
3) YOLOv11x-seg: Designed for higher precision with
fewer parameters, this model achieves high mean Av-
erage Precision (mAP) on datasets like COCO while
using 22% fewer parameters than YOLOv8m [48]. Its
efficiency makes it well-suited for devices with limited
computational resources without sacrificing precision.
The implementation of these models enables the evalu-
ation of advanced segmentation and detection methods on
the MineWasteCL_DB dataset. The architectural diversity of
these models effectively addresses the visual and geometric
variability in Sentinel-2 RGB images, ensuring the identifi-
cation of the most suitable approach for this context.
IV. EVALUATION OF SEGMENTATION AND DETECTION
ARCHITECTURES
This section evaluates various segmentation and detection ar-
chitectures to identify the most suitable model for automated
analysis of MWSFs in Chile. Selecting the optimal model is
critical for validating the CDR and generating a preliminary
cadastre of WRDs and LWDs in the Antofagasta Region.
To achieve a balance between precision and computa-
tional efficiency, multiple models with different backbones
and input resolutions (1024 ×1024 and 512 ×512 pixels)
were tested. These resolutions were chosen to ensure high-
precision results in satellite imagery, aligning with the re-
quirements of mining management tasks conducted by en-
tities like SERNAGEOMIN. The training configuration was
consistently designed to maximize comparability across mod-
els.
Images were normalized using descriptive statistics from
the complete dataset (see Table 2), facilitating transfer learn-
ing from pre-trained models. Data augmentation techniques,
including Random Resize, Flip, Crop, and Pad, were applied
using frameworks like MMDetection, Ultralytics YOLO, and
PyTorch. This enhanced the models’ generalization capacity
across diverse geographical and visual conditions in satellite
images.
TABLE 2: Descriptive statistics for image normalization
Channel Mean Standard Deviation
R 0.4012 0.1224
G 0.3228 0.1068
B 0.2638 0.0970
The dataset was randomly divided into training, validation,
and test subsets, as shown in Table 3, ensuring comprehen-
sive representation of the dataset’s conditions. This division
allowed for robust evaluation of the models’ ability to gener-
alize to unseen data. Standard performance metrics, including
IoU, AP, mAP, and recall, were used to assess each model,
focusing on their effectiveness in detecting and segmenting
MWSFs.
Given the Sentinel-2 images’ resolution of 10 m/pixel,
objects were classified by size based on their area in square
meters, following COCO dataset specifications:
•Small objects: Area less than 32 ×32 pixels, which
translates to an area less than 102,400 m2.
•Medium objects: Area between 32 ×32 and 96 ×96
pixels, equivalent to an area between 102,400 m2and
921,600 m2.
•Large objects: Area greater than or equal to 96 ×96
pixels, corresponding to 921,600 m2or more.
This classification enabled spatial correlation of model
precision with segmented object size, providing a detailed
analysis of performance in detecting and segmenting mining
structures.
Training was conducted on a server equipped with two
NVIDIA H100 GPUs, using the MineWasteCL_DB dataset
across the three MWSF categories. for each model were
initially selected based on values reported in their original
implementations and relevant literature. These were itera-
tively adjusted using a heuristic approach to optimize per-
formance on the dataset. Precision, recall, and training loss
curves were monitored to ensure convergence and stability.
TABLE 3: Distribution of the Labeled Database
Name % Images Annotations
Full Dataset 100% 30,433 320,093
Training Set 74.93% 22,805 238,827
Validation Set 16.54% 5,035 52,488
Test Set 8.52% 2,593 28,778
6VOLUME 11, 2023
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
The final hyperparameters, presented in Table 4,were fine-
tuned to address the specific characteristics of MWSFs while
maintaining computational efficiency. This approach ensured
that the models were both effective and adaptable to the
unique challenges posed by the detection and segmentation
of mining waste storage facilities.
TABLE 4: Training hyperparameters for each model
Common Parameters
Hardware NVIDIA A100 GPU, 40GB VRAM
Input Resolution 1024 ×1024 and 512 ×512 pixels
Classes WRD, TSF, LWD
Data Augmentation Flip, Rotate 90°, Scale ±20%
Evaluation Metrics AP, AP50, AP75
Normalization µ(RGB): [0.4012, 0.3228, 0.2638]
σ(RGB): [0.1224, 0.1068, 0.0970]
CNN Family
Models - Mask R-CNN (Normal & Small Anchors): ResNet-101
- ConvNeXt-Small
Batch Size 16
Epochs 36
Specific Config
- ConvNeXt: AdamW (lr: 0.0002, weight decay: 0.05)
- MaskRCNN: SGD (lr: 0.015/0.01, momentum: 0.9)
- FPN out channels: 256
- Anchor ratios: [0.5, 1.0, 2.0]
- Anchor scales: [8] / [2,4,8,16] (small)
Transformer Family
Models Mask2Former-Swin-Tiny
Batch Size 1
Epochs 36
Specific Config
- AdamW (lr: 0.0001, weight decay: 0.05)
- Swin backbone: [2,2,6,2] depths
- DeformAttn decoder: 6 layers
YOLO Family
Models YOLOv[8,9,11]-seg
Batch Size 15
Epochs 50
Specific Config - Loss weights (box: 7.5, cls: 0.5, dfl: 1.5)
- IoU thresh: 0.7, NMS disabled
The results presented in Table 5 highlight the superior-
ity of the YOLO family, particularly YOLOv8x-seg with a
resolution of 1024 ×1024, which achieved the best overall
performance. All architectures showed improvements with
increased input resolution, with notable gains in models such
as ConvNeXt. The size-based performance analysis (Table 6)
demonstrated that YOLOv8x-seg at 1024 ×1024 resolution
performed optimally across all scales, achieving a mAP of
0.975 for large deposits (>921,600 m2) and 0.936 for medium
deposits (102,400–921,600 m2). However, its performance
significantly declined for small structures (<102,400 m2),
with a mAP of 0.448. This suggests limitations in the model’s
ability to accurately identify smaller structures, potentially
due to the spatial resolution constraints of Sentinel-2 imagery.
Performance analysis by MWSF category (Table 7)urther
confirmed the dominance of the YOLO family. YOLOv8x-
seg achieved the highest results for TSFs (mAP 0.572) and
LWDs (mAP 0.590), while YOLOv11x-seg performed best
for WRDs (mAP 0.508). Traditional architectures like Con-
vNeXt and ResNet delivered lower performance in compari-
son.
Figure 3 illustrates the practical application of the trained
models at 1024 ×1024 resolution on the Minera Centinela
site. The segmentation results showcase the models’ ability
to accurately identify and delineate the contours of various
MWSFs, effectively differentiating between TSFs, WRDs,
and LWDs, while assigning confidence levels to each de-
tected structure. Based on these findings, YOLOv8x-seg with
1024 ×1024 resolution was selected for subsequent experi-
ments due to its optimal balance of precision and computa-
tional efficiency, making it highly suitable for generating an
automated MWSF cadastre.
V. EXPERIMENTS AND RESULTS
The experimental validation of the proposed system consists
of three systematic evaluations. First, the performance of
the best segmentation model was analyzed on the test set
to validate its accuracy and robustness. Second, the model’s
generalization capacity was evaluated using an unseen tempo-
ral dataset (2023–2024), ensuring its applicability to newly
acquired data. Finally, the model was applied to generate a
preliminary WRD and LWD cadastre through a case study in
the Antofagasta Region, demonstrating its practical utility in
identifying and mapping mining waste facilities.
A. PERFORMANCE EVALUATION ON TEST DATASET
The first experiment systematically evaluated the perfor-
mance of the YOLOv8x-seg model using the standard met-
rics of the COCO dataset on the temporal test set (June
2023–April 2024). This standardized evaluation calculated
precision metrics for both bounding box detection and mask
segmentation, ensuring a robust performance assessment.
The evaluation was conducted at three complementary
levels: global performance, stratification by object size, and
analysis by MWSF category. As shown in Table 8 YOLOv8x-
seg with a resolution of 1024 ×1024 achieved optimal perfor-
mance, recording a mAP/AR of 0.69/0.46 for detection and
0.54/0.44 for segmentation. Under stricter IoU criteria, the
model maintained consistent performance, achieving mAP
values of 0.73 and 0.72 at thresholds of 0.50 and 0.75, re-
spectively.
Segmentation performance across categories and object
sizes highlighted specific challenges faced by the YOLOv8x-
seg model. Lower mAP values for smaller objects, as shown
in Table 9, revealed difficulties in detecting features with re-
duced scale. For instance, small WRD objects at a resolution
of 1024 ×1024 achieved an mAP of only 0.312, indicating a
notable decline in detection capability. This limitation corre-
lates with an increased likelihood of false negatives and false
positives in such scenarios.
Additionally, the comparison between input resolutions of
1024 ×1024 and 512 ×512 underscored the importance of
higher resolutions in enhancing detection accuracy, especially
in complex scenes with closely spaced or overlapping struc-
tures. These findings demonstrate that while the YOLOv8x-
VOLUME 11, 2023 7
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
TABLE 5: Performance metrics for each model
Model Res. Bbox mAP/AR
(IoU=0.50:0.95)
Bbox mAP
(IoU=0.50)
Bbox mAP
(IoU=0.75)
Segm mAP/AR
(IoU=0.50:0.95)
Segm mAP
(IoU=0.50)
Segm mAP
(IoU=0.75)
ConvNeXt 1024 0.395 / 0.405 0.430 0.423 0.323 / 0.340 0.429 0.383
512 0.194 / 0.211 0.228 0.217 0.154 / 0.172 0.225 0.182
ResNet50 1024 0.257 / 0.287 0.378 0.287 0.203 / 0.236 0.361 0.209
512 0.158 / 0.177 0.213 0.186 0.125 / 0.143 0.207 0.136
ResNet101 1024 0.358 / 0.377 0.384 0.378 0.289 / 0.301 0.382 0.352
512 0.188 / 0.203 0.209 0.201 0.148 / 0.167 0.207 0.179
ResNet101
(small anchors)
1024 0.346 / 0.366 0.530 0.367 0.249 / 0.266 0.482 0.236
512 0.151 / 0.176 0.258 0.150 0.104 / 0.133 0.222 0.087
Mask2Former
(Swin-S)
1024 0.297 / 0.322 0.476 0.314 0.275 / 0.291 0.478 0.278
512 0.121 / 0.138 0.243 0.144 0.098 / 0.113 0.201 0.105
YOLOv8x-seg 1024 0.719 /0.482 0.740 0.739 0.557 /0.464 0.718 0.599
512 0.599 / 0.259 0.629 0.628 0.432 / 0.245 0.599 0.468
YOLOv9e-seg 1024 0.713 / 0.476 0.737 0.736 0.552 / 0.461 0.718 0.595
512 0.591 / 0.238 0.618 0.617 0.441 / 0.230 0.601 0.483
YOLO11x-seg 1024 0.713 / 0.478 0.738 0.737 0.555 / 0.461 0.717 0.594
512 0.588 / 0.230 0.614 0.614 0.438 / 0.223 0.598 0.478
Best results are highlighted in bold.
(a) (b) (c) (d)
(e) (f) (g) (h)
WRD
TSF
LWD
WRD
TSF
LWD
WRD
TSF
LWD
WRD
TSF
LWD
WRD
TSF
LWD
WRD
TSF
LWD
WRD
TSF
LWD
WRD
TSF
LWD
Legend Legend Legend Legend
Legend Legend Legend Legend
FIGURE 3: Detection and segmentation results of the AI models at Centinela mining facility with 0.9 confidence. a) ConvNeXt,
b) ResNet50, c) ResNet101, d) ResNet101 (small anchors), e) Mask2Former (Swin-s), f) YOLOv8x-seg, g) YOLOv9e-seg, h)
YOLOv11x-seg.
seg model excels in broader scenarios, its precision is signifi-
cantly challenged by smaller, less distinct features in densely
configured environments.
B. VALIDATION OF TSFS WITH CDR IN ANTOFAGASTA
REGION
The second experiment assessed the practical applicability
of the proposed methodology by cross-validating automatic
detections with the CDR in the Antofagasta Region. For this
analysis, the YOLOv8x-seg model at 1024 ×1024 resolution
was selected due to its superior performance in previous
evaluations, with a confidence threshold of 90% applied to
ensure detection reliability.
The systematic comparison was conducted using Sentinel-
2 images from the 2023–2024 period, contrasting automatic
detections with the 54 active deposits registered in the CDR
for the region (see Figure 4). To ensure robust validation, a
300 m radius around each point listed in the CDR was used
as the detection criterion. Validation results demonstrated
the system’s effectiveness, achieving a detection precision of
96.15% (calculated as true positives divided by total identi-
fied polygons) and a low false positive rate of 1.89% (false
positives divided by total detections). The system’s coverage
rate, defined as true positives divided by total CDR entries,
8VOLUME 11, 2023
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
TABLE 6: Performance Analysis of Segmentation by Object
Size Detected
Model Res. Type mAP
Small
mAP
Medium
mAP
Large
ConvNeXt
1024 Bbox 0.147 0.896 0.943
Segm 0.097 0.756 0.845
512 Bbox 0.004 0.436 0.879
Segm 0.003 0.330 0.747
ResNet50
1024 Bbox 0.078 0.599 0.725
Segm 0.040 0.468 0.641
512 Bbox 0.006 0.342 0.751
Segm 0.005 0.246 0.639
ResNet101
1024 Bbox 0.094 0.909 0.923
Segm 0.060 0.747 0.817
512 Bbox 0.003 0.386 0.890
Segm 0.003 0.295 0.749
ResNet101
(small anchors)
1024 Bbox 0.174 0.646 0.800
Segm 0.094 0.484 0.668
512 Bbox 0.019 0.273 0.668
Segm 0.007 0.155 0.525
Mask2Former
(Swin-S)
1024 Bbox 0.105 0.596 0.815
Segm 0.086 0.575 0.814
512 Bbox 0.005 0.455 0.602
Segm 0.002 0.364 0.487
yolov8x-seg
1024 Bbox 0.433 0.928 0.974
Segm 0.234 0.717 0.881
512 Bbox 0.154 0.776 0.939
Segm 0.061 0.508 0.769
Small objects: area <322pixels, Medium objects: 322
≤area <962pixels, Large objects: area ≥962pixels.
Best results are highlighted in bold.
Legend
TSFs detected
TSFs official
FIGURE 4: Comparison of detected TSFs using YOLOv8x-
seg (red) vs CDR (blue) in the Antofagasta Region. This
visualization underscores the precision and applicability of
our detection methods.
reached 47.17% (see Table 10).
Notably, 24 unidentified deposits corresponded to aban-
FIGURE 5: Automatic detection of a TSF at Mantos Blancos
mining facility.
doned or inactive facilities, highlighting the system’s inherent
ability to differentiate between operational and abandoned
TSFs. This discrimination, combined with the absence of
false negatives within the established 300 m radius and the
high precision in detecting active facilities, strongly supports
the system’s viability for automated TSF monitoring.
Figure 5 provides an example of the automatic detection
of a TSF at the Mantos Blancos mining facility, validated
against the CDR. Sentinel-2’s spatial resolution of 10 m/pixel
enables area estimation of detected deposits; in this case,
the TSF’s estimated area was 37.5 hectares. This capacity
to measure TSF surfaces enhances spatial analysis, offering
an effective tool for prioritizing inspections in high-risk areas
and optimizing the management of mining facilities.
C. GENERATION OF PRELIMINARY WRD AND LWD
CADASTRE
The third experiment focused on generating a preliminary
automated cadastre of WRDs and LWDs in the Antofagasta
Region, which currently lack an official registry. Using the
YOLOv8x-seg model, Sentinel-2 images from the 2023–2024
period were analyzed to detect these facilities. The results,
presented in Table 11, show that the system successfully
identified 141 WRD deposits and 112 LWD deposits. This
application of automated detection represents a significant
contribution by creating a preliminary cadastre for WRDs and
LWDs, deposits that have not been officially registered by
VOLUME 11, 2023 9
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
TABLE 7: Performance Analysis of Segmentation by MWSFs Category, Evaluated Using Various Metrics and Object Sizes
Category Model Res. mAP
(IoU=0.50:0.95)
mAP
IoU=0.50
mAP
IoU=0.75
mAP
Small
mAP
Medium
mAP
Large
WRD
ConvNeXt 1024 0.342 0.371 0.363 0.127 0.902 0.951
512 0.144 0.169 0.159 0.010 0.405 0.884
ResNet50 1024 0.217 0.322 0.242 0.067 0.605 0.729
512 0.109 0.143 0.130 0.010 0.297 0.754
ResNet101 1024 0.304 0.325 0.322 0.074 0.915 0.926
512 0.140 0.156 0.147 0.010 0.355 0.900
ResNet101
(small anchors)
1024 0.292 0.449 0.305 0.139 0.663 0.802
512 0.112 0.199 0.108 0.018 0.256 0.654
Mask2Former
(Swin-S)
1024 0.285 0.459 0.296 0.087 0.597 0.799
512 0.105 0.227 0.143 0.005 0.454 0.589
yolov8x-seg 1024 0.508 0.677 0.548 0.154 0.776 0.939
512 0.373 0.523 0.397 0.154 0.776 0.939
yolov9e-seg 1024 0.505 0.677 0.550 0.434 0.932 0.973
512 0.369 0.518 0.390 0.151 0.766 0.935
yolov11x-seg 1024 0.508 0.678 0.544 0.448 0.936 0.975
512 0.394 0.571 0.423 0.152 0.764 0.937
TSF
ConvNeXt 1024 0.432 0.470 0.465 0.183 0.887 0.935
512 0.216 0.254 0.242 0.000 0.472 0.878
ResNet50 1024 0.291 0.424 0.324 0.111 0.610 0.731
512 0.175 0.236 0.205 0.000 0.356 0.760
ResNet101 1024 0.388 0.414 0.409 0.120 0.896 0.923
512 0.214 0.237 0.233 0.000 0.454 0.891
ResNet101
(small anchors)
1024 0.403 0.617 0.429 0.246 0.646 0.802
512 0.181 0.310 0.178 0.034 0.304 0.705
Mask2Former
(Swin-S)
1024 0.297 0.476 0.314 0.104 0.579 0.816
512 0.121 0.243 0.127 0.006 0.439 0.603
yolov8x-seg 1024 0.572 0.739 0.627 0.061 0.508 0.769
512 0.237 0.453 0.210 0.061 0.508 0.769
yolov9e-seg 1024 0.562 0.741 0.612 0.235 0.719 0.873
512 0.234 0.453 0.205 0.059 0.499 0.758
yolov11x-seg 1024 0.568 0.737 0.613 0.234 0.719 0.878
512 0.445 0.615 0.483 0.058 0.494 0.764
LWD
ConvNeXt 1024 0.412 0.449 0.441 0.131 0.898 0.942
512 0.223 0.260 0.250 0.001 0.431 0.876
ResNet50 1024 0.262 0.389 0.294 0.055 0.582 0.716
512 0.189 0.261 0.223 0.008 0.373 0.739
ResNet101 1024 0.382 0.414 0.405 0.087 0.917 0.919
512 0.209 0.235 0.225 0.000 0.348 0.878
ResNet101
(small anchors)
1024 0.344 0.524 0.366 0.137 0.629 0.797
512 0.158 0.264 0.163 0.005 0.260 0.644
Mask2Former
(Swin-S)
1024 0.309 0.493 0.332 0.118 0.612 0.830
512 0.137 0.259 0.162 0.004 0.472 0.614
yolov8x-seg 1024 0.590 0.739 0.622 0.154 0.776 0.939
512 0.373 0.523 0.397 0.154 0.776 0.939
yolov9e-seg 1024 0.588 0.735 0.622 0.434 0.932 0.973
512 0.369 0.518 0.390 0.151 0.766 0.935
yolov11x-seg 1024 0.588 0.736 0.625 0.448 0.936 0.975
512 0.474 0.607 0.529 0.152 0.764 0.937
Best results are highlighted in bold.
TABLE 8: Global performance metrics of YOLOv8x-seg on
the test set
Model Res Bbox mAP/AR Bbox mAP Segm mAP/AR Segm mAP
mAP AR @.50 @.75 mAP AR @.50 @.75
yolov8x 1024 0.69 0.46 0.73 0.72 0.54 0.44 0.70 0.58
512 0.58 0.24 0.62 0.61 0.42 0.23 0.59 0.46
SERNAGEOMIN.
Although WRDs and LWDs do not pose the same level of
risk as TSFs, their proper identification is essential due to
their potential environmental impact, particularly in regions
of high mining activity such as Antofagasta.
Figure 6 illustrates examples of WRD and LWD detections,
highlighting the areas of interest containing these deposits.
TABLE 9: Segmentation performance by category and object
size
Cat. Res Segm mAP mAP Small Med Large
mAP @.50 @.75
WRD 1024 0.49 0.66 0.53 0.312 0.687 0.704
512 0.38 0.57 0.40 0.162 0.631 0.758
TSF 1024 0.56 0.73 0.61 0.394 0.785 0.858
512 0.44 0.60 0.48 0.187 0.683 0.792
LWD 1024 0.56 0.72 0.61 0.364 0.764 0.851
512 0.45 0.60 0.49 0.131 0.616 0.831
10 VOLUME 11, 2023
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
FIGURE 6: Examples of mining deposit detection: (a) WRD at Tal Tal mining facility (104 ha), (b) LWD at Franke mining
facility (29.7 ha), showing deposit contour, bounding box, and georeferenced location.
TABLE 10: Validation results with the CDR in the Antofa-
gasta Region
Metric Value
Total entries in CDR 53
Total identified polygons 26
True Positives (Correctly located detections) 25
False Positives (Incorrectly located detections) 1
False Negatives (No detection at polygon) 0
CDR Coverage Rate 47.17%
Precision 96.15%
False positive rate 1.89%
The generation of this cadastre provides a foundation for
future validations and audits by SERNAGEOMIN, support-
ing improved monitoring and environmental management of
these structures. This preliminary database will serve as a
critical tool for enhancing oversight and mitigating envi-
ronmental risks associated with unregistered mining waste
facilities.
The experimental design effectively addressed real-world
challenges in monitoring MWSFs. The YOLOv8x-seg model,
validated using Sentinel-2 imagery, achieved a precision of
96.15% and a false positive rate of 1.89% (Table 10).The
model was subsequently applied in the Antofagasta Region
to generate a preliminary cadastre, identifying 141 WRD and
112 LWD deposits (Table 11). These results demonstrate the
system’s potential to enhance SERNAGEOMIN’s operational
efficiency by facilitating the detection of previously unregis-
tered deposits, thereby improving environmental monitoring
and risk management.
However, applying the YOLOv8x-seg model in practical
TABLE 11: Distribution of Detected MWSFs by Category
and Size Class
Clase Small Medium Large Total
WRD 26 65 50 141
LWD 13 35 64 112
contexts revealed key considerations for its use in environ-
mental monitoring. While the model performed robustly for
large and medium-sized mining waste deposits, its perfor-
mance was notably impacted when detecting smaller struc-
tures, as indicated in Table 9. This limitation is linked to the
resolution of Sentinel-2 imagery, which constrains the ability
to discern finer details in smaller objects.
Despite this limitation, the practical application in the
Antofagasta Region confirmed the model’s efficacy in creat-
ing a preliminary cadastre, identifying 141 WRDs and 112
LWDs. These detections are vital for expanding periodic
monitoring and inspection by SERNAGEOMIN. The model’s
high precision rate of 96.15% and minimal false positive
rate of 1.89% (Table 10) highlight its operational viabil-
ity. These findings lay the groundwork for integrating auto-
mated systems into SERNAGEOMIN’s workflows, offering
an effective tool to enhance periodic monitoring, improve
environmental oversight, and ensure greater safety in mining
operations.
VI. CONCLUSION
The findings of this study highlight the effectiveness of
the proposed methodology in detecting and characterizing
MWSFs using Sentinel-2 satellite imagery and advanced deep
learning models. A key contribution is the development of the
MineWasteCL_DB dataset, a significant innovation compris-
ing over 30,000 annotated images and 320,093 labels. This
comprehensive dataset provides a robust benchmark for eval-
uating detection and segmentation models, advancing regu-
latory monitoring of mining waste through the integration of
Sentinel-2 satellite imagery and deep learning techniques.
The methodology, specifically designed for MWSF de-
tection and characterization, underwent rigorous validation
to ensure its precision and scalability. The system achieved
a high precision rate of 96.15% and a minimal false pos-
itive rate of 1.89% in detecting officially registered TSFs
within the CDR in the Antofagasta Region. Furthermore, the
VOLUME 11, 2023 11
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
model successfully identified 141 unregistered WRDs and
112 LWDs, demonstrating its potential to support prelimi-
nary registrations and enhance regulatory oversight of mining
waste facilities.
The performance of the best algorithm, YOLOv8x-seg,
was rigorously evaluated using a test dataset, achieving a
mAP of 0.69 and an AR of 0.46 for detection, and 0.54/0.44
for segmentation. These results highlight the algorithm’s ca-
pability to accurately segment and detect MWSFs at a high
resolution of 1024 ×1024 pixels.
By leveraging the periodic revisit schedule of Sentinel-2
satellites, which capture imagery approximately every five
days, our system enables regular monitoring of MWSFs. This
ongoing analysis is essential for evaluating critical parameters
affecting the physical and chemical stability of these facilities.
The use of open-source tools within our methodology pro-
vides a cost-effective solution, promoting wider accessibility
to MWSF monitoring technologies. This framework holds
potential to evolve into a comprehensive digital platform that
enhances environmental risk assessments and supports more
informed decision-making. Additionally, it lays a solid foun-
dation for improving evaluations of both physical and chem-
ical stability of MWSFs, contributing to safeguarding human
health and protecting the environment. These objectives align
with current national legislation and the OECD’s sustainable
development goals, emphasizing industry, innovation, and
infrastructure.
While the methodology has demonstrated effectiveness in
the Antofagasta Region, its adaptability to diverse mining
regions worldwide and its potential applicability to various
types of MWSFs present significant opportunities for broader
implementation. Current efforts are focused on expanding the
MineWasteCL_DB dataset to encompass all MWSFs across
Chile’s mining regions, aiming to achieve a comprehensive
national scope.
Challenges remain, particularly in differentiating between
metallic and non-metallic mining waste and in detect-
ing unregistered or abandoned facilities in regions with
limited georeferenced data. Future research will address
these issues by refining detection algorithms and integrating
higher-resolution imagery, extending the applicability of the
MineWasteCL_DB framework. These ongoing developments
promise to advance mining waste management practices, en-
hance regulatory oversight, and bolster global environmental
safety.
VII. ACKNOWLEDGEMENTS
The authors express their gratitude to SERNAGEOMIN
for their invaluable contribution in validating the informa-
tion used and providing critical feedback on the evalua-
tion methodology proposed in this work. We also sincerely
thank the undergraduate students from the School of Elec-
trical Engineering at Pontificia Universidad Católica de Val-
paraíso—Rodrigo Pereira, Gonzalo Caballero, and Felipe Ar-
avena—for their dedicated efforts in downloading and label-
ing the database.
REFERENCES
[1] M. Onifade, J. A. Adebisi, A. P. Shivute, and B. Genc, ‘‘Challenges and
applications of digital technology in the mineral industry,’’ Resources
Policy, vol. 85, p. 103978, Aug. 2023.
[2] N. E. Idoine, E. R. Raycraft, S. F. Hobbs, P. Everett, E. J. Evans, A. J. Mills,
D. Currie, S. Horn, and R. A. Shaw, ‘‘World mineral production 2018-22,’’
tech. rep., British Geological Survey, Keyworth, Nottingham, 2024.
[3] International Energy Agency, ‘‘Copper.’’
https://www.iea.org/reports/copper, 2024. Accessed: Jan. 20, 2024.
Licence: CC BY 4.0.
[4] F. Perlatti, E. P. Martins, D. P. de Oliveira, F. Ruiz, V. Asensio, C. F.
Rezende, X. L. Otero, and T. O. Ferreira, ‘‘Copper release from waste
rocks in an abandoned mine (NE, Brazil) and its impacts on ecosystem
environmental quality,’’ Chemosphere, vol. 262, p. 127843, Jan. 2021. doi
:10.1016/j.chemosphere.2020.127843.
[5] J. C. Quezada and G. Villavicencio, ‘‘Characterisation of the internal fric-
tion angle of waste rock material from large triaxial tests using the contact
dynamics method,’’ International Journal of Mining, Reclamation and
Environment, pp. 1–20, Jun 2024. doi: 10.1080/17480930.2024.2367977.
[6] G. Villavicencio, R. Espinace, J. Palma, A. Fourie, and P. Valenzuela,
‘‘Failures of sand tailings dams in a highly seismic country,’’ Canadian
Geotechnical Journal, vol. 51, no. 4, pp. 449–464, 2014.
[7] F. F. Garcia, C. F. Camilo Cotrim, S. S. Caramori, E. F. L. C. Bailão, J. C.
Nabout, G. de Farias Neves Gitirana Junior, and L. M. Almeida, ‘‘Mine
tailings dams’ failures: serious environmental impacts, remote solutions,’’
Environment, Development and Sustainability, Mar. 2024.
[8] L. Piciullo, E. B. Storrøsten, Z. Liu, F. Nadim, and S. Lacasse, ‘‘A new look
at the statistics of tailings dam failures,’’ Engineering Geology, vol. 303,
p. 106657, 2022.
[9] I. A. Okewale and H. Grobler, ‘‘Assessment of heavy metals in tailings
and their implications on human health,’’ Geosystems and Geoenvironment,
vol. 2, p. 100203, Nov. 2023. doi : 10.1016/j.geogeo.2023.100203.
[10] D. M. Franks, M. Stringer, L. A. Torres-Cruz, E. Baker, R. Valenta,
K. Thygesen, A. Matthews, J. Howchin, and S. Barrie, ‘‘Tailings facility
disclosures reveal stability risks,’’ Scientific Reports, vol. 11, p. 5353, Mar.
2021.
[11] M. Hoy, C. B. Doan, S. Horpibulsuk, A. Suddeepong, A. Udomchai, A. Bu-
ritatum, A. Chaiwan, P. Doncommul, and A. Arulrajah, ‘‘Investigation
of a large-scale waste dump failure at the Mae Moh mine in Thailand,’’
Engineering Geology, vol. 329, p. 107400, 2024.
[12] Ministerio de Minería, Chile, ‘‘Ley N°20.551: Regula el cierre de faenas
e instalaciones mineras,’’ 2020. [Online; accessed 28-October 2024.
Available online: https://bcn.cl/2f7ka].
[13] SERNAGEOMIN, ‘‘Guía metodológica para evaluación de la estabil-
idad física de instalaciones mineras remanentes,’’ 2018. Available
online: https://www.sernageomin.cl/wp-content/uploads/2019/06/GUIA-
METODOLOGICA.pdf.
[14] SERNAGEOMIN, ‘‘Datos públicos depósito de relaves,’’ 2023. Data
available for download as an Excel file: https://www.sernageomin.cl/wp-
content/uploads/2024/04/CDRCHILE2023.xlsx.
[15] K. Islam and S. Murakami, ‘‘Global-scale impact analysis of mine tail-
ings dam failures: 1915–2020,’’ Global Environmental Change, vol. 70,
p. 102361, 2021.
[16] Consejo Minero de Chile A.G., Minería en números. Consejo
Minero de Chile A.G., 7 ed., 2023. Available online:
https://consejominero.cl/en/mining-in-chile/relevance-of-the-mining-
industry-in-chile/.
[17] E. Ferreira, M. Brito, R. Balaniuk, M. S. Alvim, and J. A. dos Santos,
‘‘BrazilDAM: A Benchmark Dataset For Tailings Dam Detection,’’ The
International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences, vol. XLII-3/W12-2020, pp. 343–348, 2020. doi:
10.5194/isprs-archives-XLII-3-W12-2020-343-2020.
[18] R. Balaniuk, O. Isupova, and S. Reece, ‘‘Mining and Tailings Dam Detec-
tion in Satellite Imagery Using Deep Learning,’’ Sensors, vol. 20, no. 23,
2020.
[19] X. Wu, Y.-T. Zhang, K.-W. Lai, M.-Z. Yang, G.-L. Yang, and H.-H.
Wang, ‘‘A novel centralized federated deep fuzzy neural network with
multi-objectives neural architecture search for epistatic detection,’’ IEEE
Transactions on Fuzzy Systems, pp. 1–13, 2024.
[20] M. A. Shoaib, K. W. Lai, J. H. Chuah, Y. C. Hum, R. Ali, S. Dhanalakshmi,
H. Wang, and X. Wu, ‘‘Comparative studies of deep learning segmentation
models for left ventricle segmentation,’’ Frontiers in Public Health, vol. 10,
2022.
12 VOLUME 11, 2023
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
[21] M. Thakur, H. Kuresan, S. Dhanalakshmi, K. W. Lai, and X. Wu, ‘‘Soft
Attention Based DenseNet Model for Parkinson’s Disease Classification
Using SPECT Images,’’ Frontiers in Aging Neuroscience, vol. 14, 2022.
[22] D. Yan, G. Li, X. Li, H. Zhang, H. Lei, K. Lu, M. Cheng, and F. Zhu,
‘‘An Improved Faster R-CNN Method to Detect Tailings Ponds from High-
Resolution Remote Sensing Images,’’ Remote Sensing, vol. 13, no. 11,
2021.
[23] M. Silva, G. Hermosilla, G. Villavicencio, and P. Breul, ‘‘Automated
Detection and Analysis of Massive Mining Waste Deposits Using Sentinel-
2 Satellite Imagery and Artificial Intelligence,’’ Remote Sensing, vol. 15,
no. 20, 2023.
[24] L. Tang and T. T. Werner, ‘‘Global mining footprint mapped from high-
resolution satellite imagery,’’ Communications Earth & Environment,
vol. 4, p. 134, Apr. 2023.
[25] O. Elharrouss, Y. Akbari, N. Almadeed, and S. Al-Maadeed, ‘‘Backbones-
review: Feature extractor networks for deep learning and deep reinforce-
ment learning approaches in computer vision,’’ Computer Science Review,
vol. 53, p. 100645, 2024.
[26] I. Dimitrovski, V. Spasev, S. Loshkovska, and I. Kitanovski, ‘‘U-Net
Ensemble for Enhanced Semantic Segmentation in Remote Sensing Im-
agery,’’ Remote Sensing, vol. 16, no. 12, 2024.
[27] B. Amirgan and A. Erener, ‘‘Semantic segmentation of satellite images
with different building types using deep learning methods,’’ Remote Sens-
ing Applications: Society and Environment, vol. 34, p. 101176, 2024.
[28] A. Maurya, Akashdeep, P. Mittal, and R. Kumar, ‘‘A modified U-net-
based architecture for segmentation of satellite images on a novel dataset,’’
Ecological Informatics, vol. 75, p. 102078, 2023.
[29] I. Md Jelas, M. A. Zulkifley, M. Abdullah, and M. Spraggon, ‘‘Defor-
estation detection using deep learning-based semantic segmentation tech-
niques: a systematic review,’’ Frontiers in Forests and Global Change,
vol. 7, 2024.
[30] K. Jabłońska, M. Maksymowicz, D. Tanajewski, W. Kaczan, M. Zięba,
and M. Wilgucki, ‘‘MineCam: Application of Combined Remote Sensing
and Machine Learning for Segmentation and Change Detection of Mining
Areas Enabling Multi-Purpose Monitoring,’’ Remote Sensing, vol. 16,
no. 6, 2024.
[31] SERNAGEOMIN, ‘‘Nuevo catastro minero online,’’ 2023. Available
online: https://www.sernageomin.cl/catastro-minero/.
[32] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘‘ImageNet classification
with deep convolutional neural networks,’’ Commun. ACM, vol. 60, pp. 84–
90, May 2017. Place: New York, NY, USA Publisher: Association for
Computing Machinery.
[33] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, ‘‘You only look once:
Unified, real-time object detection,’’ 2016.
[34] L. Hu, M. Qin, F. Zhang, Z. Du, and R. Liu, ‘‘Rscnn:A cnn-based method to
enhance low-light remote-sensing images,’’ Remote Sensing, vol. 13, no. 1,
2021.
[35] J. Yang, J. Guo, H. Yue, Z. Liu, H. Hu, and K. Li, ‘‘Cdnet: Cnn-based cloud
detection for remote sensing imagery,’’ IEEE Transactions on Geoscience
and Remote Sensing, vol. 57, no. 8, pp. 6195–6211, 2019.
[36] T. Kattenborn, J. Leitloff, F. Schiefer, and S. Hinz, ‘‘Review on Convo-
lutional Neural Networks (CNN) in vegetation remote sensing,’’ ISPRS
Journal of Photogrammetry and Remote Sensing, vol. 173, pp. 24–49,
2021.
[37] A. A. Aleissaee, A. Kumar, R. M. Anwer, S. Khan, H. Cholakkal, G.-S.
Xia, and F. S. Khan, ‘‘Transformers in remote sensing: A survey,’’ Remote
Sensing, vol. 15, no. 7, 2023.
[38] M.-T. Pham, L. Courtrai, C. Friguet, S. Lefèvre, and A. Baussard, ‘‘Yolo-
fine: One-stage detector of small objects under various backgrounds in
remote sensing images,’’ Remote Sensing, vol. 12, no. 15, 2020.
[39] L. Cheng, J. Li, P. Duan, and M. Wang, ‘‘A small attentional YOLO model
for landslide detection from satellite remote sensing images,’’ Landslides,
vol. 18, pp. 2751–2765, Aug. 2021.
[40] X. Nie, M. Duan, H. Ding, B. Hu, and E. K. Wong, ‘‘Attention mask r-cnn
for ship detection and segmentation from remote sensing images,’’ Ieee
Access, vol. 8, pp. 9325–9334, 2020.
[41] H. Su, S. Wei, M. Yan, C. Wang, J. Shi, and X. Zhang, ‘‘Object detection
and instance segmentation in remote sensing imagery based on precise
mask r-cnn,’’ in IGARSS 2019-2019 IEEE International Geoscience and
Remote Sensing Symposium, pp. 1454–1457, IEEE, 2019.
[42] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, ‘‘A
convnet for the 2020s,’’in 2022 IEEE/CVF Conference on Computer Vision
and Pattern Recognition (CVPR), pp. 11966–11976, 2022.
[43] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for image
recognition,’’ in 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pp. 770–778, 2016.
[44] B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, ‘‘Masked-
attention mask transformer for universal image segmentation,’’ 2022.
[45] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, ‘‘Swin
Transformer: Hierarchical Vision Transformer using Shifted Windows,’’
2021. _eprint: 2103.14030.
[46] R. Varghese and S. M., ‘‘Yolov8: A novel object detection algorithm with
enhanced performance and robustness,’’ in 2024 International Conference
on Advances in Data Engineering and Intelligent Computing Systems
(ADICS), pp. 1–6, 2024.
[47] C.-Y. Wang, I.-H. Yeh, and H.-Y. Mark Liao, ‘‘Yolov9: Learning what you
want to learn using programmable gradient information,’’ in Computer
Vision – ECCV 2024 (A. Leonardis, E. Ricci, S. Roth, O. Russakovsky,
T. Sattler, and G. Varol, eds.), (Cham), pp. 1–21, Springer Nature Switzer-
land, 2025.
[48] G. Jocher, J. Qiu, and A. Chaurasia, ‘‘Ultralytics YOLO,’’ 2023. Available
online: https://github.com/ultralytics/ultralytics.
MANUEL SILVA VEGA (Member, IEEE) was born
in Valparaíso, Chile, in 1990. He received his BS
in Electronics Engineering in 2022 and his MSc in
Electrical Engineering in 2023 from the School of
Electrical Engineering at the Pontificia Universi-
dad Católica de Valparaiso.
He is currently a Ph.D candidate in Electrical
Engineering in the Robotics and Vision Lab at
the same university, where his research focuses on
computer vision, remote sensing, machine learn-
ing, and robotics.
GABRIEL HERMOSILLA VIGNEAU was born in
Chillan, Chile, in 1982. He received the degree in
electronic engineering from the University of La
Frontera, Temuco, Chile, in 2007, and the Ph.D.
degree in electric engineering from the University
of Chile, Santiago, Chile, in 2012. Currently, he
is an Associate Professor with the School of Elec-
trical Engineering, Pontificia Universidad Católica
de Valparaiso (PUCV), Valparaiso, Chile. His main
areas of research interest are thermal face recogni-
tion, pattern recognition, computer vision, and deep learning.
GABRIEL VILLAVICENCIO ARANCIBIA was
born in Valparaíso, Chile in 1977. He received
the Ph.D. degree in civil engineering from Ecole
Doctorale des Sciences pour l’Ingénieur. Univer-
sité Blaise Pascal, Clermont II, France in 2009.
From 2004 to 2009, he was a temporary teaching
and research associates with the civil engineering
department, Polytech Clermont-Ferrand, France.
Since 2010 to date, he has been an assistant profes-
sor with the Construction and Transportation En-
gineering of the Pontificia Universidad Católica de Valparaíso, and geotech-
nical engineer with the LEPUCV laboratory. His research interests include
geotechnical engineering applications for tailings storage facilities (TSF),
wasted rocks dumps and leaching waste dumps. In addition, in topics such as
physical stability TSF, evaluation of liquefaction potential, slope stability and
geotechnical modeling of urban sites. He is an active member of the Chilean
Geotechnical Society.
VOLUME 11, 2023 13
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
GIOVANNI COCCA GUARDIA (Member, IEEE)
was born in Calama, Chile. He received the B.Sc.
degree in engineering sciences and the Civil Elec-
trical Engineering degree from Pontificia Univer-
sidad Católica de Valparaíso, where he is currently
pursuing the M.Sc. and Ph.D. degrees in electrical
engineering. His research interests include graph
neural networks, electrical power systems, genera-
tive AI, and efficient learning AI.
PIERRE BREUL was born in France, in 1972.
He received the degree in civil engineering from
the University Blaise Pascal, Clermont-Ferrand,
France, in 1995, and the Ph.D. degree in civil
engineering from the University Blaise Pascal,
Clermont-Ferrand, France, in 1999. After having
been engineer for 5 years in a geotechnical en-
gineering company, he is currently Professor at
the National Polytechnic Institute of Clermont Au-
vergne University and head of the engineering
school Polytech Clermont. His research interests include geotechnics, soils
and granular materials mechanics, images and data analysis for soils recog-
nition and identification, infrastructure diagnosis and risk analysis.
VICENTE APRIGLIANO Professor at the School
of Construction and Transportation Engineering at
PUCV, Valparaíso, Chile. He completed a post-
doctoral fellowship at the Center for Sustainable
Urban Development (PUC-Chile) and holds a PhD
in Human Geography from Universität Tübingen,
Germany. With a Master’s in Transportation Engi-
neering (COPPE/UFRJ) and a Bachelor’s in Ge-
ography (UFRJ), his research focuses on transport
geography, economic geography, urban geography,
and urban mobility planning.
14 VOLUME 11, 2023
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2025.3546150
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
