AMD Mapping in the Lusatian Region: From Medium to Very

High-Resolution R/S Data

Delira Hanelli

, Enis Sterjo

, Roberto De La Rosa

and Andreas Knobloch

Beak Consultants GmbH, Am St.-Niclas-Schacht 13, 09599 Freiberg, Germany

Keywords: Acid Mine Drainage, Lusatia, Artificial Neural Networks, Remote Sensing, Multispectral Imagery.

Abstract: The Lusatian region is undergoing an extensive landscape rehabilitation program following the closure of

lignite open-pit mines. Under this programme, former open-cast lignite mines are being converted into

artificial water bodies. However, the region faces significant challenges related to the acidification of surface

and groundwater primarily driven by the oxidation of pyrite. Recent geochemical analyses show that, surface

waters exhibit a strong variation of pH and iron concentration. This study aims to elaborate the potential of

free and commercial space- and airborne- multispectral Remote Sensing (R/S) datasets (Sentinel-2,

Worldview-3 and Unmanned Aerial Vehicle (UAV)) for large-scale acid mine drainage (AMD) mapping and

identify the most suitable data sources and approaches for practical case studies. Additionally, cross-sensor

comparisons are performed to gain more insights into the agreement between the spectra from Sentinel-2

images with those from the Worldview-3 and UAV images over surface water. The cross-sensor agreement

of the images is quantified by performing regression analyses between R/S data at different wavelengths.

Finally, dependencies and relationships between AMD constituents and the spectral data are investigated

using artificial neural networks (ANN) of type Multi-Layer Perceptron (MLP).

1 INTRODUCTION

The Lusatian region in Germany, well-known for

lignite mining, is currently undergoing one of the

largest landscape rehabilitation programs in Europe

(LMBV, n.d.). Following the closure of open-pit

lignite mines, the region comprises approximately

176 artificial water bodies, covering a total area of

185 km² (Hanelli, et al., 2023). Leading this large-

scale transformation and rehabilitation, the Lausitzer

und Mitteldeutsche Bergbau-Verwaltung (LMBV) is

monitoring the water quality in these water bodies by

periodic sampling campaigns and geochemical

analyses.

The water quality across the region varies

significantly due to differences in acid mine drainage

(AMD) stages, water treatment methods, and

geological conditions. For instance, newly flooded or

untreated lakes often exhibit highly acidic conditions,

with pH values between 2.5 and 4.5, whereas treated

and naturally neutralized lakes typically range

https://orcid.org/0009-0002-0012-1729

https://orcid.org/0009-0009-0993-0750

https://orcid.org/0000-0002-3004-7104

https://orcid.org/0000-0001-7515-001X

between 6.5 and 8 (LMBV, n.d.). Given its

complexity and large spatial extent, the Lusatian

region presents an ideal case study for developing and

validating cost-effective AMD mapping methods.

Remote sensing (R/S) technologies offer

promising solutions for large-scale AMD monitoring

(Hanelli et al., 2023; Farahnakian et al., 2024;

Kopačková, 2019). This study evaluates the potential

of free and commercial multispectral datasets from

spaceborne and airborne platforms for AMD mapping

in a selected area of the Lusatian region characterized

by strong AMD variations. A key focus is on cross-

sensor comparisons to assess spectral data

consistency across platforms and the transferability of

AMD-related spectral relationships.

The multispectral R/S datasets utilized in this

study include Sentinel-2, WorldView-3, and

Unmanned Aerial Vehicle (UAV) data. These

datasets cover different areas of interest (AOI)

depending on their availability, costs and accessibility

(Figure 1). The free-of-charge Sentinel-2 data cover

Hanelli, D., Sterjo, E., De La Rosa, R. and Knobloch, A.

AMD Mapping in the Lusatian Region: From Medium to Very High-Resolution R/S Data.

DOI: 10.5220/0013497200003935

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 11th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2025), pages 305-316

ISBN: 978-989-758-741-2; ISSN: 2184-500X

305

an extensive area of post-mining water bodies and

several AMD scenarios. In contrast, the commercial

WorldView-3 data were acquired only for a 100 km²

representative area with significant AMD activity.

In most of the post-mining lakes in the Lusatian

region the access is restricted due to geotechnical

instabilities (ground subsidence and landslides) and

very acidic water environments (LMBV, n.d.). Given

these limitations and the objectives of this study, two

pilot sites were selected for UAV surveys: Scheibe

See (685 hectares, no evidence of AMD) and

Bergheider See (325 hectares, evidence of AMD).

Figure 1: AOIs for each of the used R/S datasets.

The R/S datasets were utilised as key parameters

for area-wide mapping of AMD in post-mining water

bodies. At the same time, geochemical analyses of

AMD components, such as iron concentration and pH

values obtained from surface water samples, served

as calibration data for the mapping process. The

relationships between the AMD components and the

spectral data are exploited by means of artificial

neural networks (ANNs).

This application requires a synchronization of R/S

datasets with geochemical analysis, in order to ensure

that the measured water quality parameters

correspond to the recorded surface reflectance

information in the R/S datasets. This is particularly

challenging considering the prerequisites of optical

R/S imagery (e.g. cloud- and shadow-free, low-nadir-

angle etc.). To increase the chances of getting suitable

acquisitions, a temporal window of ± 10 – 20 days to

the sampling date was applied, assuming no

significant geochemical changes within this

timeframe (e.g., through neutralization processes).

This study integrates both R/S and geochemical

datasets collected in late June 2024.

The study aims to improve AMD mapping

techniques and identify the most effective and

practical methodologies for environmental

monitoring in post-mining landscapes.

2 DATA ACQUISITION

2.1 Geochemical Analysis

Geochemical analysis results for 31 post-mining

water bodies were provided by LMBV for this study

(Figure 2). The samples were collected from the

epilimnion layer (0–15 m depth). Figure 1 shows the

measured values of iron concentration as graduated

symbols/colours, while the measured pH values are

shown as column chart, whereas small columns

indicate an acidic environment (pH varies between 2

and 3) and the big ones a neutral environment (pH

varies between 6 and 8).

Figure 2: Water monitoring stations and recorded values of

AMD constituents in the Lusatian post-mining water

bodies.

The water bodies in this region are characterized

by strongly varying AMD levels (Table 1) and are

therefore suitable for this research study.

Table 1: Statistics of the water geochemical parameters.

Paramete

Count Min Max Mean St

H 31 2,45 7,96 3,77 1,90

(

)

31 0,07 361 90,65 101,42

Figure 3 shows a plot of dependencies between

the iron concentration and pH values. Water bodies

with high iron concentrations are typically marked by

low pH values (≤4), while water bodies with low iron

concentrations typically have pH values between 6

and 8. Although pH is an optically non-active

parameter and cannot be directly detected using

optical R/S data, in this study we use the observed

dependencies for large-scale mapping of pH values

by leveraging patterns of optically active constituents,

such as iron concentration.

S34I 2025 - Special Session on S34I - From the Sky to the Soil

306

Figure 3: Plot of dependencies between Log (Fe) and pH.

2.2 R/S Datasets

An overview of the used R/S datasets for AMD

mapping follows in Table 2. All data were projected

to WGS1984/UTM 33N (WKID: 32633).

Table 2: Overview of the acquired imagery for AMD

mapping (VNIR: visible and near-infrared, LWIR: Long-

Wave Infrared, SWIR: Short-Wave Infrared).

2.2.1 Sentinel-2 Data

Sentinel-2 data were downloaded from https://code-

de.org/de/. CODE-DE is part of Germany's

geoinformation strategy and offers easy and efficient

access to remote sensing data as well as free cloud

resources for processing. More detailed information

on the Sentinel-2 acquisition resolutions can be found

on the Copernicus Sentinel-2 Mission website

(Copernicus, n.d.).

For the AOI there is a wide archive of historical and

actual data available. Top of Atmosphere (TOA) and

Bottom of Atmosphere (BOA) products were

downloaded respectively. The spatial resolution is set

to 10 m to benefit from the medium spatial resolution

of Sentinel-2 imagery.

The cloud cover of Sentinel-2 acquisition over water

bodies in the area of interest (AOI).

2.2.2 Worldview-3 Data

The commercial high-resolution (HR) Worldview-3

data (VNIR+SWIR) was ordered from European

Space Imaging (https://www.euspaceimaging.com/)

for a 100 km

representative area. More detailed

information on the WorldView-3 data can be found

on the ESA Earth Online website (European Space

Agency [ESA], n.d.). Important prerequisites aspects

to consider for optical analysis are the cloud coverage

and the low nadir angle.

The data has been made available as Ortho-Ready

Standard Product (OR2A), with the spectral data as

Digital Numbers (DN). The DN values are further

processed to TOA Reflectance values using the

radiometric calibration tool supported by NV5

Geospatial's software tools (NV5 Geospatial, n.d.).

The conversion to BOA reflectance resulted in

negative reflectance values in water areas, because of

the low radiance. It is important to highlight that the

atmospheric corrections are designed for land

applications, and not for water bodies. In water

applications they modify the reflectance drastically.

For this reason, TOA reflectance data are used for

further processing.

2.2.3 UAV Data

The very-high resolution (VHR) UAV data were

acquired from Beak Consultants GmbH with the

following equipment:

 UAV: DJI Matrice M300 RTK,

 Multispectral camera: Micasense Altum-PT

(detailed information can be found in the

Altum-PT Integration Guide – MicaSense

Knowledge Base [MicaSense, n.d.]),

 GPS/GNSS System: Emlid RS2 GPS/GNSS

(Global Positioning System / Global

Navigation Satellite System) with NTRIP

(Networked Transport of RTCM via Internet

Protocol) connection to the national CORS

system (Continuously Operating Reference

Station).

The necessary approvals/permissions for the UAV

flights were obtained in advance based on the

regulations of the German Federal Aviation Authority

(LBA, n.d.).

The processing of UAV acquisitions relies on the

Structure from motion (SfM) photogrammetric range

imaging technique. However, applying this technique

over large water bodies presents several challenges:

AMD Mapping in the Lusatian Region: From Medium to Very High-Resolution R/S Data

307

 High reflectivity: The reflective properties of

water create a mirror-like effect.

 Dynamic surface conditions: The continuous

movement of water, influenced by factors such

as wind, introduces discrepancies of

overlapping areas in consecutive scenes.

 Stereo image similarity: The homogeneity of

stereo-image pairs over water surfaces makes it

difficult to identify tie and key points necessary

for accurate image alignment.

To align UAV acquisitions in this case study, we

employed the image block-adjustment by reference

technique, as implemented in Agisoft Metashape

(Agisoft LLC, n.d.). This workflow includes a yaw

estimation process that analyses the drone's flight

path between consecutive images to determine the

camera's horizontal rotation. Assuming zero pitch and

roll, yaw is the only rotational parameter considered.

The drone's movement direction is calculated by a

direction vector, obtained by subtracting the current

camera's location from the next camera's location.

This process effectively determines the camera's

horizontal orientation based on its movement relative

to the previous shot, aiding in initial camera

alignment for photogrammetric processing.

Because of the large areas, UAV flight campaigns

are conducted over multiple days, often under varying

illumination conditions, leading to variations in

ground surface brightness. Figure 4 shows the tiles

representing the flight missions and the weather

conditions, respectively. To mitigate these variations,

the sun sensor correction (DLS) is applied, which

partially compensates for differences in lighting

conditions during data acquisition. However, this

approach assumes a constant irradiance over time and

lacks to develop irradiance series and compensate the

DLS for movement (MicaSense, n.d.)

(a) (b)

Figure 4: The flight missions and respective weather

conditions in a) Bergheider See and b) Scheibe See.

Finally, the recorded values have been divided by

32768 to get the reflectance values for each band

instead of digital numbers (MicaSense, n.d.).

3 METHODS DESRCRIPTION

Dependencies and relationships between spectral

reflectance bands and AMD constituents are

investigated using the supervised machine learning

(ML) algorithm of ANNs of the multilayer perceptron

type (MLP) (Haykin, 1998). Additionally, cross-

sensor comparisons (Chastain, et al., 2019) are

performed to gain more insights into the agreement

between the spectra from Sentinel-2 images with

those from the Worldview-3 and UAV images over

surface water. Finally, transformation parameters are

calculated to harmonize Worldview-3 and UAV

spectral bands to Sentinel-2 over water bodies.

3.1 Prediction Modelling Using

Artificial Neural Networks

MLP ANNs are implemented in the advangeo®

Prediction Software from Beak Consultants GmbH

(www.advangeo.com). The modeling and prediction

software analyses complex relationships between a

wide variety of spatial influencing parameters (in this

case the multispectral data) and given AMD

occurrences, by using methods of artificial

intelligence (AI) within a Geographic Information

System (GIS) environment. The base principle is the

ability of ML algorithms to generalize and learn from

non-linear relationships, and model natural complex

processes and events, which are difficult or

impossible to be described with analytical

mathematics (Noack, et al., 2014).

The aim of the modelling is large-scale mapping

of the iron concentration and pH based on R/S

multispectral data and geochemical data and

elaboration of the influence of spatial and spectral

resolution in the modelling process.

The accuracy and robustness of the trained

network is assessed by:

 Statistical evaluation: A comparison plot of the

modelling results with the measured values of

iron concentration and pH.

 The network mean squared error (MSE): A

converging and stable model error indicates

that the network is learning effectively.

 The model distribution weights: Balanced

model distribution weights indicate that the

network has appropriately distributed

importance across all spectral bands without

overemphasizing or neglecting any specific

band.

 The distribution map: Predictions should

closely align with the actual target values and

present uniform and logical AMD clustering.

S34I 2025 - Special Session on S34I - From the Sky to the Soil

308

3.2 Cross-Sensor Comparison of

Sentinel-2 and Worldview-3 TOA

Products

The calibration of a reliable training network requires

adequate and sufficient sampling data, covering a

wide range of AMD scenarios. Given the free

availability of Sentinel-2 data and the typically high

costs of high-resolution commercial R/S datasets, we

propose a methodology where ANNs are trained

using Sentinel-2 data over a large area in conjunction

with extensive geochemical monitoring data. The

established dependencies and relationships are then

applied to commercial high-resolution datasets for

targeted identification of AMD in specific areas. This

approach requires that the training and application

models are provided with similar controlling

parameters.

Though the R/S multispectral datasets used in this

study can provide “similar” observations (VNIR-

SWIR for Sentinel-2 and Worldview-3 data, and

VNIR for UAV data), they differ in the field of view,

spatial resolution, spectral bandwidth, and spectral

response function. While the difference introduced

by different field of view and spatial resolution can be

reduced and solved by the orthorectification and data

resampling, respectively, the difference caused by

different spectral bandwidth and spectral response

function (the so-called reflectance difference) is a

more complex problem.

In this study, we apply the linear regression

approach at different wavelengths to minimize

reflectance difference between two similar satellite

observations over water bodies. Figure 5 shows the

cross-sensor agreement analyses between the

Sentinel-2 and Worldview-3 TOA data. The Sentinel-

2 bands B5, B7, B8A, B11 and B12 are paired to

Worldview-3, based on the introduced concept of

synthesised bands by (Gasparovic, et al., 2018).

The weakest correlations are observed in the low-

resolution Sentinel-2 bands, specifically B1 (443

nm), B9 (940 nm), and the blue spectral band B2 (490

nm). The low correlation in the blue wavelength

range can be addressed to the fact that the reflection

in this part of the spectrum is more susceptible to

atmospheric scattering, which can drastically affect

measurements.

Figure 5: Cross-sensor agreement analyses between the

Sentinel-2 and Worldview-3 TOA data.

Figure 6 shows examples of cross-sensor

comparison of the median spectra for three lakes:

Kleinleipischer See (Fe = 195 mg/L, pH = 2,56),

Bergheider See (Fe = 50,2 mg/L, pH = 2,74) and

Poleysee (Fe = 0,15 mg/L, pH = 3,9). In all lakes, the

Sentinel-2 and Worldview-3 spectra are in good

agreement in terms of shape.

Additionally, we use the transformation

parameters from the linear regression model to

harmonize the reflectance of Worldview-3 to

Sentinel-2. The Sentinel-2 and adjusted Worldview-3

spectra are in good agreement in terms of shape and

magnitude (Figure 6). However, the adjusted spectra

are slightly brighter than those of Sentinel-2 within

the near- and short wavelength infrared (>700-

nanometers).

(a)

AMD Mapping in the Lusatian Region: From Medium to Very High-Resolution R/S Data

309

(b)

(c)

Figure 6: TOA reflectance spectra for a) Kleinleipischer

See (Fe = 195 mg/L, pH = 2,56), b) Bergheider See (Fe =

50,2 mg/L, pH = 2,74) and c) Poleysee (Fe = 0,15 mg/L, pH

= 3,9).

3.3 Cross-Sensor Comparison of

Sentinel-2 BOA Products and UAV

Similar to section 3.2, cross-sensor agreement

analyses were performed between Sentinel-2 BOA

data and UAV data. The results of these analyses are

presented in Figure 7. The Sentinel-2 bands B5 and

B7 are paired to UAV, based on the introduced

concept of synthesised bands by (Gasparovic, et al.,

2018). The weakest correlations are observed in the

visible part of the spectrum corresponding to

Sentinel-2 B2 (490 nm), B3 (560 nm), B4 (665 nm).

Figure 7: Cross-sensor agreement analyses between the

Sentinel-2 BOA and UAV data.

A detailed view of the individual bands revealed

that the visible bands are more susceptible to the

ground surface brightness and reflectance differences

coming from dynamically changing weather

conditions.

These discrepancies are confirmed also when

comparing the Sentinel-2 BOA and UAV median

spectra for Scheibe See (Figure 8). Generally, both

spectres in the two lakes (Bergheider and Scheibe

See) are in good agreement both in terms of shape and

magnitude. Due to their similarity, in this case no

transformation is used to harmonize the reflectance of

UAV data to Sentinel-2 BOA.

(a)

(b)

Figure 8: BOA reflectance spectra for a) Bergheider See (Fe

= 50,2 mg/L, pH = 2,74), b) Scheibe See (Fe = 0,17 mg/L,

pH = 7,38)

4 RESULTS

Two MLP training models were developed using

Sentinel-2 and WorldView-3 data as controlling

parameters. These models aim to evaluate the impact

of WorldView-3's higher spatial and spectral

resolution on AMD mapping. Furthermore, an

application model leverages the established

dependencies derived from Sentinel-2 training model

and applies them to harmonized WorldView-3 data to

assess the transferability of the knowledge gained in

MLPs across similar remote sensing datasets.

S34I 2025 - Special Session on S34I - From the Sky to the Soil

310

On the other hand, the development of training

models based on UAV-derived data for two lakes was

deemed impractical due to the limited availability of

training data. However, an application model was

implemented to evaluate the transferability of the

established dependencies for AMD mapping and was

validated by the geochemical sampling in the two

lakes.

4.1 Training Scenario Using Sentinel-2

Controlling parameters include Sentinel-2

multispectral bands of Level-1C and Level-2A

products. Figure 9 shows a comparison plot of the

modelling results with the measured values of iron

concentration and pH values. The trained neural

network has been able to reproduce the calibration

data in case of the iron concentration, as an optically

active parameter. In case of pH values, there are no

clearly established dependencies between the

controlling parameters and the calibration data,

however there is a significant differentiation of the

acidic from the neutral waters.

(a) (b)

Figure 9: Plot of given and modelled a) iron concentration

and b) pH values based on Sentinel-2 Level-2A BOA

products.

The MSE in both cases shows systematic

convergence and remains stable, confirming the

neural network's accuracy and robustness (Figure 10).

(a) (b)

Figure 10: Plot of MSE for the MLP for a) iron

concentration and b) pH values based on Sentinel-2 Level-

2A BOA products.

The model parameter weights revealed the

Sentinel-2 Level-2A green (B03) and SWIR (B11 and

B12) spectral bands to have the highest contribution

for the modelling of iron concentration and B08

(NIR) and SWIR (B11 and B12) for pH values.

The result is a distribution map of iron

concentration (Figure 11) and pH values (Figure 12)

in the value ranges of input calibration data (0 – 361

mg/L and 2–8, respectively) over the water bodies in

the AOI. The typical patterns of high iron

concentrations and low pH values in the shores are

mostly due to the mixed pixel information in shallow

waters. Generally, the distribution map reflects the

AMD severity as measured from the geochemical

analysis.

Figure 11: Distribution map of iron concentration over the

water bodies in the AOI; additionally, the measured Fe

values are shown as a column chart.

Figure 12: Distribution map of pH values over the water

bodies in the AOI; additionally, the measured pH values are

shown as a column chart.

AMD Mapping in the Lusatian Region: From Medium to Very High-Resolution R/S Data

311

4.2 Training Scenario Using

Worldview-3

Controlling parameters include Worldview-3 TOA

multispectral bands. In this case, the geochemical

analytic results are available for about 10 post-mining

water bodies in the AOI. Figure 13 shows a

comparison plot of the modelling results with the

measured values of iron concentration and pH values.

In this case, the trained neural network has been able

to better reproduce the calibration data.

(a) (b)

Figure 13: Plot of given and modelled a) iron concentration

and b) pH values based on Worldview-3 TOA products.

The MSE in both cases shows systematic

convergence and remains stable, confirming the

neural network's accuracy and robustness (Figure 14).

Figure 14: Plot of MSE for the MLP for a) iron

concentration and b) pH values based on Worldview-3

TOA products.

Similar to the training model in 4.1, the training

model weights confirmed the Worldview-3 TOA

green and SWIR (from SWIR2 to SWIR7) spectral

bands to have the highest contribution for the

modelling of iron concentration and NIR1/2 and

SWIR (SWIR6 to SWIR8) for pH values.

The resulting distribution maps of iron

concentration (Figure 15) and pH values (Figure 16)

reflect the AMD severity as measured from the

geochemical analysis, taking into consideration only

those part of the WV3-image that are free of cirrus

clouds.

Figure 15: Distribution map of iron concentration over the

water bodies in the AOI; additionally, the measured Fe

values are shown as a column chart.

Figure 16: Distribution map of pH values over the water

bodies in the AOI; additionally, the measured pH values are

shown as a column chart.

4.3 Application Scenario Using

Harmonized Worldview-3

This scenario is useful when there are no sufficient

calibration data inside the AOI of commercial

Worldview-3 data. In this case, the harmonized

Worldview-3 image bands to Sentinel-2 are used as

controlling parameters and the established

dependencies from the training scenario in 4.1 are

used for AMD mapping. This approach does not

S34I 2025 - Special Session on S34I - From the Sky to the Soil

312

exploit the full potential of SWIR in the Worldview-

3 data, since they are harmonized to SWIR bands of

Sentinel-2. However, it enables AMD mapping in the

shores and in small/narrow water bodies, which

cannot be represented properly in medium resolution

images.

Figure 17 shows a comparison plot of the

application results with the measured values of iron

concentration and pH values. The predicted iron

concentration values resemble to the trend of

measured values, but they are obviously

overestimated, showing higher AMD-levels then the

ones from geochemical analysis. This is reflected also

in the modelled pH value, where all the water bodies

in the AOI are predicted as very acidic.

(a) (b)

Figure 17: Plot of given and modelled a) iron concentration

and b) pH values based on harmonized Worldview-3 TOA

products.

The same observations are confirmed from the

distribution maps of iron concentration (Figure 18)

and pH values (Figure 19).

Figure 18: Distribution map of iron concentration over the

water bodies in the AOI; additionally, the measured Fe

values are shown as a column chart.

Figure 19: Distribution map of pH values over the water

bodies in the AOI; additionally, the measured pH values are

shown as a column chart.

4.4 Application Scenario Using UAV

Data

In this scenario, a new neural network was trained by

Sentinel-2 data using only bands in the VIS-VNIR.

This network confirmed the Sentinel-2 Level-2A

green band (B03) to have the highest contribution for

modelling of iron concentration and NIR band (B08)

for the pH value. The established dependencies in the

trained network were used for large-scale mapping of

AMD in Scheibe See and Bergheider See using UAV

data.

Figure 20 and 21 show the distribution map of iron

concentration and pH value over Bergheider See and

Scheibe See. The median values of AMD parameters

over both lakes and results of geochemical analysis

are presented for comparison in Table 3.

(a) (b)

Figure 20: Distribution map of a) iron concentration and b)

pH value over the Bergheider See.

AMD Mapping in the Lusatian Region: From Medium to Very High-Resolution R/S Data

313

(a) (b)

Figure 21: Distribution map of a) iron concentration and b)

pH value over the Scheibe See.

Table 3: Comparison of geochemistry and modelling results

for Bergheider See and Scheibe See.

Ber

heide

See Scheibe See

Measure

Model Measure

Model

H 2,74 3,84 7,38 6,62

(

)

50,2 39,3 0,17 9,4

The variating weather conditions between the mission

flights seem to have a very small effect in the modelling of

iron concentration, but show a considerable influence in the

modelling of pH values. This issue is further elaborated in

chapter 6.

5 DISCUSSION

This study assesses the feasibility of using free and

commercial multispectral R/S datasets in

combination with supervised ML algorithms for the

automatic mapping of AMD in water bodies.

Supervised ML algorithms require large and

diverse training datasets that represent a wide range

of AMD scenarios. However, in practical

applications, such comprehensive datasets are often

unavailable. To address this limitation, the study

developed training models using a large variety of

water bodies exhibiting different AMD levels. The

transferability of the established models was further

evaluated using commercial R/S datasets.

The application of optical R/S analyses requires

cloud- and cirrus-free acquisitions. Sentinel-2 data

are generally more available for this purpose due to

their high temporal resolution. In contrast, acquiring

commercial datasets such as WorldView-3 imagery

often requires careful planning, as constraints such as

low cloud coverage and low nadir angles significantly

influence the availability of suitable acquisitions.

UAV-flight campaigns also present logistical

The median pH values for Scheibe See were derived only

from tiles captured under cloudy weather conditions.

challenges, especially for large water bodies, as they

require several days of data collection under stable

weather conditions to ensure consistent reflectance

values across adjacent flight paths.

The high spatial and spectral resolution of

WorldView-3 imagery, particularly in the SWIR

bands, proved to be highly effective for AMD

mapping. This may be related to the ability of SWIR

bands to detect high concentrations of heavy metals

typically associated with severely acidic conditions.

Future research could further explore AMD patterns

within the SWIR region using hyperspectral datasets,

such as those provided by EnMAP or PRISMA.

Additionally, the green spectral band was identified

as an important feature, potentially due to the absence

of vegetation or algae in water bodies with high

acidity levels.

The MLP models developed in this study

demonstrated robust performance but are limited by

the concentration ranges defined by the calibration

data. Consequently, quantitative predictions cannot

be reliably extrapolated beyond the range of the

calibration data. In this context, discrete sampling

remains of critical importance for properly calibrating

or validating the algorithms. However, the proposed

approach enables large-scale AMD mapping of water

bodies by significantly reducing the need on

extensive sampling campaigns.

The flight campaign for a complete survey of

large water bodies can take several days due to the

European Union Aviation Safety Agency (EASA)

restrictions for UAV (such as a maximum flight

height of 120-meters). In practice, it is almost

impossible to have constant weather conditions

during such campaigns. The variating weather

conditions (cloud, cirrus, haze, shadows) have a big

influence on the surface reflectance of water bodies,

leading to difficulties for balancing of the reflectance

values and modelling inconsistencies over a water

body.

To improve UAV-based monitoring, fixed-wing

UAVs are recommended for their ability to cover

larger areas efficiently, reducing weather-induced

variability and resulting imbalances across flight

missions. Additionally, multispectral cameras with

wider spectral bands would enhance the detection of

subtle water quality variations. On the other hand,

increasing the UAV flight altitude would

considerably reduce flight time and also improve the

accuracy of photogrammetric reconstructions by

enhancing feature variations between consecutive

images.

S34I 2025 - Special Session on S34I - From the Sky to the Soil

314

6 CONCLUSIONS

The novelty of this study is the cross-sensor

comparison of free and commercial space- and

airborne- multispectral R/S datasets (Sentinel-2,

Worldview-3 and UAV) with a focus on assessing the

transferability of established dependencies between

AMD parameters and spectral data across several

datasets.

The cross-sensor analysis identified spectral

discrepancies coming mainly from differences in

spectral bandwidth and spectral response functions.

To address these variations, transformation

parameters were derived to align the spectral

characteristics of commercial datasets with those of

Sentinel-2, which was used as a reference due to its

free availability and high temporal resolution. This

makes Sentinel-2 a valuable dataset for training ML

algorithms.

Results indicate that adjusted WorldView-3 data

appear slightly brighter than Sentinel-2 data in the

NIR and SWIR (>700 nm) regions. Consequently, the

transferred neural network exhibited a tendency to

overestimate AMD levels. Future research can focus

on optimizing transformation parameters using larger

and more diverse datasets, including time-series data

and broader spatial coverage. Nevertheless, the

correct relative distribution of iron concentrations

suggests that the established dependencies from the

training model remain transferable across these

datasets. This approach fully elaborates the high

spatial resolution of WV3-datasets and enables AMD

mapping even in small-scale or narrow water bodies,

offering a more efficient and cost-effective

alternative, as running extensive training models on

commercial datasets.

The training scenario with the best results was

obtained when using Worldview-3 datasets as

controlling parameters, due to their high spatial and

spectral resolution, particularly in the SWIR bands.

However, the trained network in this case is relied in

a few number of water bodies and AMD scenarios.

The transferred neural network for UAV-based

monitoring has shown also very promising results.

While clear-sky and sunny conditions offer optimal

reflectance, they can introduce sun-glint effects in

UAV-based monitoring. The large-scale pH

distribution map of Scheibe See (Figure 21)

highlighted the significant impact of weather

conditions on the modelling process. In Bergheider

See, flight missions occurred under more consistent

conditions, resulting in minimal weather-related

influences. These findings suggest that bright,

diffused sunlight represents the ideal weather

conditions for UAV-based water quality monitoring.

Finally, despite not being included in any training

scenarios, Scheibe See was correctly classified as a

lake with no evidence of AMD, demonstrating the

applicability of the trained neural network beyond the

AOI. This demonstrates the robustness and

application of the developed approach for large-scale

mapping of the water quality in post-mining water

bodies.

ACKNOWLEDGEMENTS

This research has been done within the framework of

the project S34I—Secure and sustainable supply of

raw materials for EU industry¬, coordinated by Ana

C. Teodoro. This project has received funding from

the European Union’s HORIZON Research and

Innovation. Grant Agreement No. 101091616

(https://doi.org/10.3030/101091616). Portuguese

National Funds also support this work through the

FCT – Fundação para a Ciência e a Tecnologia, I.P.

(Portugal), projects UIDB/04683/2020

(https://doi.org/ 10.54499/UIDB/04683/2020) and

UIDP/04683/2020 (https://doi.org/10.54499/UIDP/

04683/2020).

The authors are grateful to LMBV for providing the

water monitoring data, which made this research

study possible. The authors gratefully acknowledge

also the comments and suggestions of three

anonymous reviewers, which led to a substantial

improvement in the manuscript.

REFERENCES

Agisoft LLC. (n.d.). Metashape scripts [GitHub

repository]. https://github.com/agisoft-llc/metashape-

scripts

Chastain, R., Housman, I., Goldstein, J., Finco, M.,

Tenneson, K. (2019). Empirical cross sensor

comparison of Sentinel-2A and 2B MSI, Landsat-8

OLI, and Landsat-7 ETM+ top of atmosphere spectral

characteristics over the conterminous United States,

Remote Sensing of Environment, Volume 221, 2019,

Pages 274-285, ISSN 0034-4257, https://doi.org/

10.1016/j.rse.2018.11.012.

Copernicus. (n.d.). Sentinel-2 mission. Retrieved from

https://sentiwiki.copernicus.eu/web/s2-mission

European Space Agency (ESA). (n.d.). WorldView-3

mission. Retrieved from https://earth.esa.int/

eogateway/missions/worldview-3

Farahnakian, F., Luodes, N., Karlsson, T. (2024). “A

Comparative Study of Machine Learning Models for

Pixel-wise Acid Mine Drainage Classification Using

Sentinel-2”. IGARSS 2024 - IEEE International

Geoscience and Remote. 2024.

AMD Mapping in the Lusatian Region: From Medium to Very High-Resolution R/S Data

315

Gasparovic, M., Medak, D., Pilaš, I., Jurjevic, L.,

Balenović, I. (2018). Fusion of Sentinel-2 and

PlanetScope Imagery for Vegetation Detection and

Monitoring. ISPRS - International Archives of the

Photogrammetry, Remote Sensing and Spatial

Information Sciences. XLII-1. 155-160. 10.5194/isprs-

archives-XLII-1-155-2018.

Hanelli, D., Barth, A., Volkmer, G., Köhler, M. (2023).

Modelling of acid mine drainage in open pit lakes using

Sentinel-2 time-series: A case study from Lusatia,

Germany. Minerals 2023, 13, 271.

https://doi.org/10.3390/min13020271.

Haykin, S. (1998). Neural Networks: A Comprehensive

Foundation. United States: Prentice Hall PTR, Upper

Saddle River, NJ, United States; ISBN 978-0-13-

273350-2.

Kopačková, V. (2019). “Mapping acid mine drainage (amd)

and acid sulfate soils using sentinel-2 data”. IGARSS

2019 - 2019 IEEE International Geoscience and

Remote Sensing Symposium, S. pp. 5682–5685.

Lausitzer und Mitteldeutsche Bergbau-

Verwaltungsgesellschaft (LMBV). (n.d.). Homepage.

https://www.lmbv.de/

LBA (n.d.) Legal basis for drone operations. Available at:

https://www.lba.de/DE/Drohnen/Allgemeine_Informat

ionen/Rechtliche_Grundlagen/Rechtliche_Grundlagen

_node.html

MicaSense. (n.d.). MicaSense image processing [Website].

https://micasense.github.io/imageprocessing/

MicaSense. (n.d.). Altum-PT integration guide – MicaSense

knowledge base. Retrieved from https://support.

micasense.com/hc/en-us/articles/Altum-PT-

Integration-Guide

MicaSense. (n.d.). MicaSense support center. Retrieved

from https://support.micasense.com/

Noack, S., Knobloch, A., Etzold, S. H., Barth, A.,

Kallmeier, E. (2014). Spatial predictive mapping using

artificial neural networks. The International Archives of

the Photogrammetry, Remote Sensing and Spatial

Information Sciences, XL–2, 79–86.

NV5 Geospatial. (n.d.). Homepage.

https://www.nv5geospatialsoftware.com/

S34I 2025 - Special Session on S34I - From the Sky to the Soil

316