Detection and mapping abandoned areas of artisanal and small-scale gold mining (ASGM) using multi-sensor data on Google Earth Engine: A case study of Kuantan Singingi, Riau

. Artisanal and small-scale gold mining (ASGM) activities in Kuantan Singingi, Riau have been operating over a decade without proper permits and using unsafe procedures for the environment. Mercury releases and degraded land have been the leading factors in the decreased environmental functions. ASGM activities are nomadic and secluded, posing a considerable challenge in detecting their location and extent. The aims of this study are to provide a method for detecting and mapping ASGM footprints utilizing multi-sensor data on cloud computing platforms. The detection method is performed using a supervised random forest algorithm. The result successfully mapped an ASGM footprints, estimating an area of 10,044.38 ha with 89.23% accuracy through Sentinel-1 data and an area of 12,308.57 ha with 87.25% accuracy through Sentinel-2 data. The spatial distribution of ASGM footprints is scattered over the streams and tributaries across all regions. These maps are pivotal in establishing regulatory measures for environmental restoration and preventing further expansion of degraded land.


Introduction
Artisanal and small-scale gold mining (ASGM) is widespread throughout Indonesia and has been operating for decades.Generally, ASGM activities in Indonesia are conducted illegally and lack environmentally safe procedures [1].ASGM activities are carried out by dredging massive sediments and using large quantities of mercury, leading to land degradation and pollution of water and soil [2].This has caused a significant impact on the environment, such as reduced watershed function, damage to freshwater ecosystems, loss of biodiversity, and an increased frequency of flooding events.Releasing mercury into the environment presents a significant hazard, as it can lead to substantial human exposure and result in severe health consequences for individuals [3].
ASGM activities in Kuantan Singingi Regency, Riau, have been active for more than a decade.Abandoned mining sites have left extensive stretches of degraded land polluted by mercury chemicals.Miners operate in remote locations that are difficult to reach and can swiftly relocate in a matter of days, making it challenging to pinpoint the exact locations of ASGM activities.In some cases, authorities have obtained information about ASGM from indigenous communities who have witnessed mining activities around their settlements.However, most of these affected lands have not been identified.Meanwhile, the mapping and * Corresponding author: ikhwan.amri@mail.ugm.ac.id monitoring of ASGM activities in Kuantan Singingi are still limited and have received less attention.The remaining locations and spatial extents need to be determined to establish a framework for rehabilitation and reclamation policies aimed at preventing further land damage.When developing a rehabilitation policy framework, it is crucial to map ASGM sites to detect spatial patterns, assess the extent of affected areas, and plan landscape management [4].
Remote sensing data can be relied upon to detect and map land affected by illegal mining activities in near real-time, rapidly, and reach areas that are difficult to access [5].The synoptic overview capability of remote sensing data relies on several approaches to analyze the inter-relationship between spatial patterns, regional characteristics, and environmental issues [1,5].Remote sensing data is composed of a large number of pixels that contain digital values representing geographic features' response (soil, water, vegetation, and built-up) to electromagnetic energy [5].The rapid development of digital processing platforms has enabled the provision of cloud computing services such as Google Earth Engine, which provides spatial big data at multiple scales, sensors, and temporal resolutions.Cloud computing enables the processing of large-volume data and cloudfree images, which are useful for detecting ASGM footprints in tropical regions.
Detection and mapping of degraded land due to ASGM based on remote sensing data have been conducted and developed in various studies [1,4,6,7].Pixel-based classification and segmented pixel methods are commonly used techniques.Both of these methods require training data from users for knowledge transfer purposes.Several guided classification techniques have been developed into machine learning algorithms.Machine learning algorithms such as random forest, decision trees, artificial neural networks, and support vector machines are suitable for handling complex spectral signatures [8], such as separating ASGM land use from non-ASGM open land and built-up land.
ASGM is indeed one of the significant mining sectors, but it also has substantial negative impacts on the environment and human health.However, up to the present era, monitoring of ASGM-affected land in Indonesia has received relatively little attention [9].Further studies comparing machine learning-based methods for detecting ASGM footprints from multisensor images are still in development in Indonesia.Accurate identification of illegal mining plays a pivotal role in monitoring natural resource management and environmental protection.This research aims to provide a method for the detection and mapping of ASGM areas in Kuantan Singingi Regency, Riau Province, Indonesia, by utilizing multi-sensor remote sensing data and the Google Earth Engine cloud computing platform.A supervised classification approach with machine learning algorithms was implemented to extract the location and spatial extent of ASGM areas.This method can be adapted for application in other regions addressing illegal mining issues, especially ASGM.The map of the ASGM-affected area can be utilized as input data for drafting regulations related to environmental conservation and rehabilitation based on regional landscape planning.

Study area
The Kuantan Singingi region was formed through fluvial processes involving several streams and tributaries (Figure 1).The name of this region is derived from the major rivers that flow through it, namely the Kuantan River and the Singingi River.The accumulation of alluvial sediments containing goldbearing materials has attracted miners to exploit these minerals [10].The exploration of gold-bearing materials in Kuantan Singingi began in 2005 and has continued to accelerate up to the present day [11].
Kuantan Singingi is one of the areas with the largest ASGM activity in Riau Province.In 2021, Kuantan Singingi was designated as a pilot project for artisanal gold mining in Riau Province to minimize environmental damage.However, illegal ASGM operations persist, and the remaining mining land has not been rehabilitated, resulting in an abandoned wasteland.

Data
The image dataset was collected from the Google Earth Engine database.Google Earth Engine provides a number of collection remote sensing data that can be accessed easily, quickly, and free of charge.Data retrieval was performed using the JavaScript programming language on the Google Earth Engine code editor.
This study used Sentinel-2 Multispectral Instrument (MSI) and Sentinel-1 C-band Synthetic Aperture Radar Ground Range Detected (SAR GRD) datasets.The latest 2022 data was retrieved to determine the current location and extent of the ASGM footprint.
To obtain high-quality and uniform Sentinel-2 MSI data, various filters were employed, including cloud masking and median filtering.For creating a uniform subset of Sentinel-1 SAR GRD data, it was necessary to apply filters using metadata properties.Additionally, a speckle filter was applied to reduce salt and pepper noise at the pixel level.
Several spectral indices from Sentinel-2 data were generated as additional input analysis.Spectral indices are able to enhance specific object by modified digital number.In this study, spectral indices were used are the normalized difference vegetation index (NDVI) (Equation ( 1)), the modified normalized difference water index (MNDWI) (Equation ( 2)), and the normalized difference built-up index (NDBI) (Equation ( 3)).In Sentinel-2 data, the spectral bands collected included green, red, near infrared (NIR), and short-wave infrared (SWIR).The equation of these indices can be found in below [12][13][14].

Supervised classification
The supervised classification was performed using a random forest algorithm to classify different objects from the subset images (Figure 2).Random Forest is an algorithm that determines classification results by averaging the results of a set of decision trees.This algorithm has the advantages of reducing dataset overfitting and increasing detection accuracy.The Random Forest algorithm successfully detected and separated surface features with over 80% accuracy [7,8,15].
In this research, ASGM footprints were differentiated from other features, which included non-ASGM objects and water bodies.These features were identified based on recognizable characteristics such as pattern, shape, size, and color, which were visualized in the image.The characteristics of these features are described in Table 1.The model was subsequently tested for accuracy using testing data collected by utilizing the dataset contained in Google Earth, which involved identifying existing objects based on high-resolution satellite imagery and local knowledge of the study area.A total of 250 points were distributed, with 50 points allocated to each land cover class.The determination of the sample size followed the approach outlined by Congalton and Green [16].The testing data was then utilized to construct a comparison matrix for calculating the values of producer accuracy, user accuracy, overall accuracy, and the kappa coefficient (Equations ( 4), ( 5), (6), and ( 7)).
PA (producer accuracy) is the probability that a pixel is correctly classified in a particular class.Meanwhile, UA (user accuracy) is the probability that a pixel classified in the model represents the actual class on the ground.OA (overall accuracy) represents the percentage of correct classification of the validation data.K (kappa coefficient) measures the agreement between the classification produced by the model and the actual values.

Accuracy assessment
The detection of ASGM footprints using the supervised random forest algorithm resulted in satisfactory accuracy.The overall accuracy of the classification results for Sentinel-1 and Sentinel-2 was 89.23% and 87.25%, respectively (Table 2 and Table 3).Sentinel-1 achieved a higher overall accuracy than Sentinel-2 MSI, which can be attributed to the combination of single polarizations (VV) and dual polarization (VH) as the primary data in the classification inputs.The classification result of Sentinel-2 had an overall accuracy 1.98 points lower than the overall accuracy of Sentinel-1.These lower accuracies were a result of misclassifications, as indicated by the user's accuracies being lower than the producer's accuracies in the non-ASGM class.The misclassifications could be attributed to the presence of bare land features in certain areas, which were classified as ASGM features.Additionally, the side-looking effect further compromised the accuracy mapping due to the fact that it produced a shadow effect on the opposite slope.
However, Sentinel-2 performed better in detecting fewer ASGM features compared to Sentinel-1.The PA percentage for ASGM features in Sentinel-2 was higher than that in Sentinel-1.The lower PA percentage in Sentinel-1 indicated a higher number of false negative errors for ASGM footprints, as evidenced by the lower detection rate of ASGM footprints.Sentinel-2 demonstrated greater sensitivity in detecting ASGM; it detected ASGM in certain areas that were not identified as such by Sentinel-1.This inconsistency suggested challenges in accurately distinguishing between ASGM and other bare land areas.Nevertheless, the classification results of Sentinel-1 exhibited a more compact pattern compared to Sentinel-2, which appeared disjointed and fragmented.

ASGM abandonment land footprint
The spatial distribution of the ASGM footprint in Kuantan Singingi is visually depicted in Figure 3. Based on the classification results from both Sentinel-1 and Sentinel-2 data, the pattern of ASGM locations is predominantly associated with the presence of stream networks, further accentuated by their linear shapes.Specifically, the natural morphology of the Singingi River has undergone significant alterations due to ASGM activities, serving as a clear indication of substantial environmental degradation resulting from uncontrolled mineral resource exploitation.
The findings in our study align with previous research conducted in Kuantan Singingi.Focusing on the Singingi River, Mailendra and Buchori confirmed that gold mining activities are commonly found along the river, often operating without proper permits [11].The extent of land degradation along the riverbanks is so substantial that distinguishing the river's course from mining excavations becomes challenging.Additionally, within a more limited area, we identified ASGM activities around the Kuantan River and other tributaries.Some characteristics of the ASGM footprint are irregularly scattered amidst non-ASGM land cover.This could potentially be a result of land conversion, affecting both natural ecosystems (e.g., bareland and shrubland) and agrogenic ecosystems (e.g., rubber and oil palm plantations) [11].
In addition, we conducted calculations to determine the number of pixels and the area covered by ASGM footprints based on the classifications using Sentinel-1 and Sentinel-2 data (as shown in Table 4).There were slight variations in the results obtained from these calculations.The ASGM footprint features identified through the Sentinel-1 classification amounted to 257,438 pixels, which was roughly equivalent to 10,044.38 ha.In comparison, the ASGM footprint features identified through the Sentinel-2 classification totaled 329,164 pixels, corresponding to approximately 12,308.57ha.
Figure 4 presents a comparative view of ASGM footprint sizes in each sub-district of Kuantan Singingi Regency.Mining activities have been identified across all sub-districts.Both Sentinel-1 and Sentinel-2 consistently classify Singingi Hilir, Singingi, and Kuantan Tengah as the top three sub-districts with the largest ASGM footprint areas, each covering over a thousand hectares.This information holds vital importance for decision-makers, emphasizing the need to prioritize extensive post-mining environmental restoration efforts in these specific regions.Conversely, Inuman, Hulukuantan, and Kuantan Hilir sub-districts unanimously exhibit the smallest ASGM footprint areas.Our study effectively mapped ASGM footprints in Kuantan Singingi using multi-sensor remote sensing data and the GEE platform, achieving an acceptable level of accuracy.By employing machine learning algorithms, ASGM sites could be swiftly identified without the need for extensive field observations.The resultant mapping outcomes for ASGM serve as a foundation for further investigating potential repercussions, such as alterations in river water quality due to mining activities and the socioeconomic impacts on local communities [2].Moreover, it is essential to implement tighter monitoring of mining operations to prevent their unauthorized expansion into different areas.Failure to address this matter optimally could potentially lead to heightened environmental crises.The challenging part of detecting ASGM is separating ASGM features from bare area and other mine areas.Improve understanding the nature and state of ASGM activities aids to establish knowledge transfer in machine learning.Nevertheless, the developed random forest model has a limited ability to separate ASGM from current coal mining activities.To get a reliable map for actual conditions, an intervention through gathering knowledge of ASGM activities should continue to be conducted.
Unpredictable and unmonitored ASGM activities could potentially cause changes over time.A time series identification of ASGM extent potentially becomes key understanding the nature of ASGM activities and how mining activities have been transformed to contribute land cover change and land degradation.
In addition, data input should be further explored to determine the difference in spectral patterns produced.Integration optical and radar sensors could be considered as data input for future ASGM studies, since the results of this study show that radar sensors are more uniform at detecting, but the sensitivity of optical sensors is more delicate in detection particular areas.To fully harness the potential of remote sensing techniques in enhancing the understanding of artisanal mining impacts, it is crucial to adopt interdisciplinary and collaborative approaches, which can effectively address the multifaceted challenges linked to mining sites [18].

Conclusion
The detection of ASGM activities using multi-sensor data and machine learning algorithms has effectively mapped the spatial extent and distribution of ASGM footprints, achieving an overall accuracy of 89.23% through Sentinel-1 and 87.25% through Sentinel-2.A significant presence of ASGM footprints was identified, covering approximately 10,044.38 ha through Sentinel-1 and 12,308.57ha through Sentinel-2.Most of the ASGM footprints in Kuantan Singingi are located along streams and tributaries, resulting in extensive abandoned and degraded wastelands.Sentinel-1 produced maps that displayed relatively compact ASGM features, while Sentinel-2, with its multispectral sensor capabilities, exhibited slightly higher sensitivity in detecting ASGM features.The ASGM footprint maps hold vital importance for emphasizing the regulation framework of post-mining restoration and rehabilitation to overcome the environmental consequences due to illegal mining activities.

Table 1 .
Characteristics of the features class.

Table 4 .
Number of pixel and area of ASGM footprints based on Sentinel-1 and Sentinel-2 classification.