Characteristics of the studied region and satellite images

Our studied region covers a large area of 1.75 million km² divided among 13 countries in the Sub-Tropical West Africa (Fig 1). This region encompasses three distinct biomes [7]: tropical and sub-tropical moist broad-leaf forests (41.1% of our sample region), a mixture of tropical and subtropical grasslands, savannas and shrub-lands (55.1%), and mangroves biome (3.8%) (S1 Fig). To minimize the cloud coverage density and to keep the vegetation colour relatively constant, we make use of satellite images from November to February. The spatial resolution of our images is 10m × 10m (Sentinel-2 satellite, European Space Agency Copernicus program). Sentinel-2 images combine 13 bands in the visible, near infrared, and short wave infrared part of the spectrum, ranging from 440 nm to 2200 nm, with different resolutions (S1 Table in S1 File). These bands perfectly fit our remote sensing application, as recent studies highlight the relevance of the near infrared bands to classify land cover mapping [8], along with the traditional visible bands (red, green and blue). For more information, see S1 Appendix in S1 File.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Studied region visualization.

Mapping of the paper’s studied region (white) and presentation of the dispatched labeled dataset boxes, divided into training set (red), validation set (green) and test set (blue).

https://doi.org/10.1371/journal.pone.0267963.g001

Labelled dataset creation process

A crucial step of our methodology is the creation of a labelled dataset to train, tune and evaluate our model. For instance, during the training step, every time the model predicts a previously labelled region, it compares the ground truth label to the prediction made. The model parameters are then optimized by minimizing the error between these two. Even if qualitative initiatives exist to map mining activities [9], available data contains no information on location at pixel level, or the exact date of the labelling is missing or imprecise. The creation of our labelled dataset is particularly challenging as the identification of asm is related to class imbalance problem: while the number of asm has increased in the last years, they are still sparse and highly dispersed across space, likely representing less than 0.1% of the African continent.

Facing this challenge, we follow a sixth-steps strategy to create our reference dataset. First, we divide the region into a grid, each cell covering 14.5 km × 14.5 km (210 km²: 2.1 million pixels), with an overlap of 1 km on each side (S2 Fig). Second, to get a first idea about the potential asm distribution, we retrieve for each cell the density of “quarry” labels from the Open Street Map contributions. Indeed, pre-identifying boxes with a high density of “quarry” labels helps to overcome the challenge of the positive class’ sparseness. Third, we annotate a mixture of randomly chosen boxes with and without asm, looking for the right compromise between two conflicted purposes: i) including a subsequent number and variety of mining examples, while ii) maintaining our labelled dataset (relatively) representative of the class imbalance reality. After empirical exercises testing various asm proportions, and in coherency with our objective to first minimize the wrong predictions, we finally aimed to reach approximately 1% of asm in our dataset. However, it should not necessarily be the same value in other contexts (e.g. in other regions, or in fine-tuning an already pre-trained algorithm). Fourth, drawing lessons from our recurrent wrong prediction observations, we refine our dataset by manually select boxes containing particular features, such as rivers with sandy shores, railways and highways, and buildings under construction. Note that these three combined steps -2, 3 and 4- were not successive but iteratively operated together. In theory, they might be continually further-improved, by gradually enlarging and refining the current labelled dataset with a wide variety of landscapes, features and asm types. Fifth, for each selected box, given the fine-grained resolution of our images, we thoroughly label the area by imposing strict conditions for a pixel to be marked as an asm: i) limitation to a minimum size for a labelled mine (size > 1, 000 m², i.e. minimum 10 pixels); ii) significant discordance in visible light ground colour compared to surroundings (e.g. to exclude the former asm that started re-vegetated); iii) obvious characteristics of mineral exploitation such as furrows/height holes, presence of pipelines and/or connection to roads; and iv) a typical signature in the temporal evolution of the feature’s shape. Sixth, in addition to these prerequisites, we use images with a resolution up to 60cm × 60cm from Google Earth to cross-validate that every asm location found on Sentinel-2 images is actually tangible asm. For a limited number of boxes—usually for very rural/remote regions -, the images’ resolution available in Google Earth was not sufficient and consequently, these boxes were excluded from the labelled dataset. Note that for each box, the labelling strategy was double-checked. In the end, our labelled dataset covers 23,337 km² over 111 boxes (Fig 1).

Dataset partition

To train and estimate the model performance, the labelled dataset is divided into three different partitions: i) the training set (55% / 13,035 km²) used to train the model; ii) the validation set (20% / 4,625 km²), with images the model has never seen during the training, which is used to tune the model parameters and to avoid over-fitting the training set. Also, the best weights are selected based on validation metrics only and iii) the test set (25% / 5,677 km²) with images the model has never seen before, to estimate the model’s performance. Given the limited amount of labelled data and the scarcity of asm areas in comparison to other landscapes, we select boxes for the test set to ensure that we have sufficient asm—in terms of variety of geological features and surface—to compute coherent metrics (see Results section), while keeping the maximum number of boxes in the training set. Boxes were also selected and split to fulfill several characteristics for each dataset: an equal distribution across regions, a consistent mine coverage proportion and to comprise various typical climate-regrouped landscapes: rivers, cities, agricultural fields, small mountains, coasts and wildlife.

Pre-processing and model description

We start by implementing a pre-processing stage for Sentinel-2 images that includes atmospheric and geometrical corrections, normalization and data augmentation (see S2 and S3 Appendices in S1 File for further details). Following the recent advances in remote sensing, we make use of a modified version of the deep learning model U-Net [6], which is widely used in biomedical segmentation, semantic segmentation and remote-sensing applications [10–13]. Furthermore, a cnn like U-Net performs better than the traditional pixel-based methods [14] since the images’ contextual information is also considered, which is important as some mining activities can only be detected by the relevant understanding of their surroundings. Our model differs from the standard U-Net model by adding batch normalization layers [15] and dropout layers [16] to reduce the tendency toward over-fitting on the training data. In practice, over-fitting means that the model strictly memorizes examples seen during the training instead of learning from them, leading to a difficulty to generalize and usually a worse performance on new data. In S4 Appendix in S1 File, we provide more detailed information on the model and parameters used.

Performance metrics

We start by introducing the precision and recall statistics. They depict, respectively, our ability to i) correctly discriminate the mines—that is precision—and ii) to detect all of them—that is recall -, which are competing abilities though complementary. They are derived from the number of pixels correctly predicted as asm (true positives, TP, see Fig 2a), on the number of pixels that were mistakenly predicted as asm (false positives, FP), and the number of pixels that were mistakenly predicted as not asm (false negatives, FN).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Classification of the predictions: Illustrations.

(a) Illustration of the standard pixels classification derived from the comparison between prediction and reality (TP, FP, FN and TN, together forming the confusion matrix) from a sub-sample of our predicted test set in the region of Bata, Equatorial Guinea. (b) Confusion matrices computed for three distinct scenarios (defined in section Performances along probability threshold value). It comes with an illustration of the binary prediction (red shapes) performed for the corresponding T value, in the region of New Netim (Nigeria). Note that figures are computed for the whole test set, contrary to the sample images that aim to illustrate the effect of the T value variation on a relevant area. Sources: Sentinel-2 (from the Copernicus program) & author calculations.

https://doi.org/10.1371/journal.pone.0267963.g002

As their opposition is inherent to their conflicted purposes, we determine a relevant trade-off between those two statistics considering i) the upstream model’s objectives; ii) the positive class distribution; and iii) its potential ability to be distinguished from other classes of features. Consequently, we introduce the commonly used F(β)_score, defined as the weighted harmonic mean between precision and recall [17]. Given that asm is a highly minority class, easily confused with many other features—such as any landscaping projects, partly due to the spatial resolution of the input satellite data -, we set the β parameter to 0.5, thus putting more emphasis on the predictions’ accuracy rather than their quantity. This ensures that the data provided has satisfying minimum levels of precision, reducing the false positives in priority (this indicator is also used to select the model’s final weights during the validation phase).

However, by ignoring the true negatives (TN) class in its computation, i.e. by focusing mainly on the positive class, the F(β)_score could present an incomplete view of the prediction reality [18]. As such, we suggest an alternative metric with the Matthews Correlation Coefficient (MCC). This standard metric, which evaluates the binary (two-classes) classification quality, has the advantage of often being described as a balanced measure as it can be used even if the classes are of very different sizes, i.e. in imbalanced classification problems such as asm prediction. See S5 Appendix in S1 File for details.

Performances along probability threshold value

A crucial aspect is the mapping of asm for different applications, varying from policy makers aiming to own a complete vision of all potential asm locations, to research use which requires highly accurate data. In this perspective, a major lever stands in the selection of the appropriate probability threshold (T, hereafter) to produce the final binary classification map from the model’s raw probability output.

As a matter of illustration, we distinguish three typical scenarios, illustrated on the region of New Netim, Nigeria (Fig 2b). In each scenario, we display the correctly predicted pixels on the matrix diagonal (TN and TP), which contains the vast majority of the test set surface, leading to an overall prediction accuracy larger than 99%, whereas on the anti-diagonal are counted the wrongly classified pixels (FN and FP). This matrix also highlights the class imbalance we have to deal with, assuming that the vast majority of the areas are not mines—and are classified as such correctly.

T = 0.2 (Fig 2b-i): In this scenario, we maximize the number of predicted locations where asm are likely to take place, while placing less emphasis on the precision of the information. This translates into an overview of at least 50% of the actual mines, but with a relatively low precision (around 60%). This scenario is particularly suitable when the information can be cross-checked locally, for instance by local policy makers or non-governmental organizations, as the data provided will contain false positives. As depicted on Fig 2b-i, mining areas are all detected as long with a mispredicted small part of a village.

T = 0.4 (Fig 2b-ii): This scenario stands as the middle ground situation between the maximum values for the F(β = 0.5)_score and the mcc. For instance, this scenario is perfectly suitable for the purpose of mapping overall asm activities when it is not possible to cross-check the information locally. Note that the small part of the village wrongly predicted in the previous scenario is not anymore.

T = 0.6 (Fig 2b-iii): Finally, this scenario corresponds to a maximum precision (≈90%) without a recall under 30%. For instance, this scenario is well suited to research projects focusing on the socioeconomic impacts induced by asm activities, for which the veracity of the asm locations stands as crucial, because it minimizes the false positives.

Ultimately, Fig 3a) illustrates the evolution of the metrics with the T variation, hence demonstrating that choosing an appropriate T value goes along with relevantly selecting a specific set of metrics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Main performance metrics computed on the test set.

(a) Overall performance metrics (Precision, Recall, F(0.5) score and MCC) depending on the probability threshold value. (b) Recall metric along asm size classification depending on the probability threshold value.

https://doi.org/10.1371/journal.pone.0267963.g003

Performances along ASM size classification

Another important feature of our model is its ability to detect asm independently from their ground surface range (in hectares). In our labelled dataset, the average asm size is around 13 hectares, while the standard deviation is around 70 hectares and 75% of the labelled asm cover less than 7 hectares (S3 Fig). This indicates that our model is trained on a wide range of asm sizes and not only large-scale (or industrial) mines. To investigate its ability to effectively detect this wide range of asm, we make use of the labelled dataset to create equally distributed mine-sized classes and compute the recall for each of them. We also add a last class to separately represent the industrial mining activities, since their size and features are distinct from asm. The resulting classes are: micro asm from 0.1 to 1 ha; small asm from 1 to 2.5 ha; medium asm from 2.5 to 7 ha; macro asm from 7 to 2000 ha; and separately the industrial mines.

Fig 3b) displays the results. Consistent with the previous performances, asm are more likely to be detected for low T values, and the recall decreases progressively as T increases, whatever the mine-sized class. Additionally, the larger the asm ground surface, the more easily it is identified by the algorithm, even though the large-scale mines are under-represented. Putting in perspective the results to pixels amounts to explain this phenomenon, micro asm only represent a few pixels, making the recognition essentially based on the colour bands, whereas macro asm also provide a typically recognizable mine shape and contain extra features—such as tailing ponds or large visible ditches via the variations in the pixels’ colours. Besides, as depicted on Fig 3b), the difficulty to detect the micro-asm (recall ≈ 12% for T = 0.4) supports our decision to favour the model’s predictions only for asm larger than 0.1 ha.

Taken together, the results point out that our model performs well in detecting all kinds of asm—even with the smallest asm representing only a few pixels within images of 100 million pixels—with varying proportions depending on the selected scenario. This diversity is illustrated on Fig 4. Furthermore, selecting a specific scenario also implies querying specific mine-sized classes, which is another way to adapt the resulting data to the user’s particular objectives.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Timeline examples of various regions predicted by the model.

Four examples of predictions performed by our model, from various parts of the studied region. Each vertical timeline chronologically presents i) a sample of Sentinel-2 (S2) model’s input image as reference, ii) the ground truth asm contouring label (in black) for this S2 sample, iii) the corresponding probability prediction performed by our model, and iv) the binary prediction obtained after applying a probability threshold T = 0.5. Sample a) comes from Yaoundé city in Cameroon (3°58’N: 11°30’E), sample b) comes from Kenema District in Sierra Leone (8°10’N: 10°56’W), sample c) comes from southern Nigeria (6°44’N: 5°09’E) and sample d) comes from Koinadugu District, Sierra Leone (9°09’N: 11°27’W). Sources: Sentinel-2 (from the Copernicus program) & author calculations.

https://doi.org/10.1371/journal.pone.0267963.g004

Performances across climate and landscape variations

To envision extending the asm prediction to the whole African continent, the model’s robustness to landscape and climate diversity is the most crucial limiting factor. The two main biomes identified in our studied region (S1 Fig) respectively cover 54.2% (savannas, grasslands and shrub-lands) and 11.7% (moist broad-leaf forests) of the whole continent surface, and hence constitute a significant sample to determine the robustness of our model. Thus, we investigate the performances on each biome separately, by dividing the 27 boxes contained in our test set according to their biome affiliation. We observe coherent performances on the two major biomes: for T = 0.4, 85% precision while retrieving 46% of the asm on the moist broad-leaf forests biome (F(β = 0.5)_score = 0.73), and 60% precision while retrieving 28% of the asm on the savannas, grasslands and shrub-lands biome (F(β = 0.5)_score = 0.48). Note that the recall gap can be explained by i) the asm smaller average size and ii) by the lower contrasts between asm and ground vegetation in more arid regions such as savannas. Alternatively, the performances on the remaining biome—mangroves, less than 4% of the studied region and ≈ 0.1% of the whole Africa—are still too inaccurate, which can mainly be explained by its poor representation in our region. Still, the remarkable consistency of the results over the two major sub-Saharan eco-regions, which present very distinct and challenging landscape features (partly illustrated on Fig 4), indicates a comforting ability to generalize the procedure at very large scales.

Comparison to literature

In this paper we use a convolutional neural network (cnn), which is widely adopted in the literature [10–13]. Alternatively, some authors use pixel-based classifiers to map asm activity in Ghana [19–21], Burkina Faso [22] and Brazil [23]. In opposition to cnn, pixel-based methods do not leverage the pixels’ context but only make use of their inherent properties limiting the detection of asm. Hence, their reported omission errors range from 8 to 40% for the mining class, indicating a large variability in the accuracy of this method [19–21]. In regards to the recent literature on mines detection using deep learning on the same spatial resolution (10m), our accuracy (> 99%) is similar but the performance of our model differs in some dimensions. First, covering 1.75 million km² has the drawback to lower our segmentation metric in comparison to papers covering smaller regions, especially in the case of regions renowned for their rich content of asm, like Ghana with a region of 63,000 km² covered [13]. Second, in contrast, our classification metrics are very close to recent coal mines classification [24], while detecting a larger array of mine’s types (Fig 4), which implies dealing with more heterogeneous data. Third, using very high resolution images (0.81m) to detect open-pit mines on a single region in China [25], they display a F-Score of 0.67 and a precision of 0.80. Our model’s performance is similar on the precision metric, but lower on the detection’s accuracy for mines smaller than 1ha. Besides the literature on mine detection, there is a growing literature that uses machine learning techniques and satellite imagery to measure various human or environmental outcomes with a very high accuracy [26]. For instance, with very high resolution images (1m) brick kilns are detected with a precision of 88% and with a very low probability to be missed [27], as well as trees in Sahel (0,5m) [28] that are detected up to a size of 3m² with a very high recall of 95%. Last, our classifier achieves an Area Under Curve (AUC) of 0.86 on the test set (S4 Fig), which is similar and coherent with respect to the score of 0.84 obtained in [29] when detecting buildings’ destruction in conflict zones. To conclude, drawing a comparison to existing work highlights the necessity of making trade-offs between the area covered, the satellite resolution, the accuracy of the detection and the minimum size that can be detected.

Results synthesis

By aggregating the results along the different dimensions, our model performs remarkably well compared to similar work that has been completed in this area, which reflects the relevant compromise made between cost, scale and performance criteria. Considering the case of a probability threshold of 0.4, the model predicts 1.75 million km² for which 6,970 km² are correctly reported as asm. First, we achieve a precision of 70.4% while reaching at the same time a recall of 42.2%. Second, detected asm activities range from 0.1 ha to around 2,000 ha, with for instance one-third of the small asm (1 ha < surface < 2.5 ha) retrieved, thus satisfying our primary ambition to detect the largest permissible range of asm. Third, we crucially achieve coherent performances on the two major biomes represented in this study. Furthermore, as the Mangroves area is not well predicted by the algorithm, if the user’s region of interest stands within the Tropical and sub-Tropical moist broad-leaf forests biome, predictions even get much greater: a precision of 85% while retrieving 46% of the asm and a satisfying MCC of 0.62. This last finding is particularly reassuring, suggesting our model should perform well on larger African regions or other continents. Additionally, we compare our predictions to the most exhaustive existing dataset on the large-scale mines locations (S&P Global—SNL Metals and Mining). We detect all of the 265 listed large-scale mines located within our region of interest. Moreover, we detect 700 clusters of asm, originally unrecorded, in a radius of 5km around the locations of the large-scale mines included in the S&P Global dataset.