A Deep Learning Approach to the Detection of Gossans in the Canadian Arctic

remote sensing

Article A Deep Learning Approach to the Detection of Gossans in the Canadian Arctic

Étienne Clabaut 1,*, Myriam Lemelin 1, Mickaël Germain 1 , Marie-Claude Williamson 2 and Éloïse Brassard 1

1 Département de Géomatique Appliquée, Université de Sherbrooke, Sherbrooke, QC J1K 2R1, Canada; [email protected] (M.L.); [email protected] (M.G.); [email protected] (É.B.) 2 Natural Resources Canada, Ottawa, ON K1A 0E4, Canada; [email protected] * Correspondence: [email protected]

Received: 28 July 2020; Accepted: 16 September 2020; Published: 23 September 2020

Abstract: Gossans are surficial deposits that form in host bedrock by the alteration of sulphides by acidic and oxidizing fluids. These deposits are typically a few meters to kilometers in size and they constitute important vectors to buried ore deposits. Hundreds of gossans have been mapped by field geologists in sparsely vegetated areas of the Canadian Arctic. However, due to Canada’s vast northern landmass, it is highly probable that many existing occurrences have been missed. In contrast, a variety of remote sensing data has been acquired in recent years, allowing for a broader survey of gossans from orbit. These include band ratioing or methods based on principal component analysis. Spectrally, the 809 gossans used in this study show no significant difference from randomly placed points on the Landsat 8 imageries. To overcome this major issue, we propose a deep learning method based on convolutional neural networks and relying on geo big data (Landsat-8, Arctic digital elevation model lithological maps) that can be used for the detection of gossans. Its application in different regions in the Canadian Arctic shows great promise, with precisions reaching 77%. This first order approach could provide a useful precursor tool to identify gossans prior to more detailed surveys using hyperspectral imaging.

Keywords: gossan; deep learning; convolutional neural network; geo big data; multispectral; Landsat; Canadian Arctic

1. Introduction Supergene weathering and oxidation of massive sulphide deposits often leads to the formation of iron caps between the surface and the water table (Figure1). These gossans are dominated by the presence of silica, goethite, hematite and jarosite [1]. Depending on the relative abundance of these minerals, gossans are characterized by red to yellow hues [2,3].Gossans are of interest to geologists because they constitute vectors to economic mineral deposits (e.g., [4,5]). Also, the oxidation of oxide-sulphide gossans in permafrost constitutes an analogue for natural acid mine drainage [6,7]. Therefore, gossans provide valuable environmental proxies in areas currently aﬀected by climate change (e.g., [8–14]) particularly in fragile ecosystems like the Arctic. Additionally, gossans located in the permafrost regions are used to characterize comparable deposits on Mars and their potential to harbour life [15].

Remote Sens. 2020, 12, 3123; doi:10.3390/rs12193123 www.mdpi.com/journal/remotesensing Remote Sens. 2020, 12, x FOR PEER REVIEW 2 of 16 Remote Sens. 2020, 12, 3123 2 of 16

Ferricrete

Figure 1. IllustrationIllustration of of a a simplified simpliﬁed gossan gossan structure structure adapted from [[3].3].

Remote predictivepredictive mappingmapping of of gossans gossans is is widely widely used used in in mineral mineral exploration exploration surveys surveys carried carried out outin remote, in remote, sparsely sparsely vegetated vegetated areas. areas. Some Some gossans gossans have alreadyhave already been studied been studied from a remotefrom a sensingremote perspective.sensing perspective. Their identification Their identification is generally is possible generally by thepossible observation by the of observation absorption features of absorption in their featuresreflectance in spectratheir reflectance by assuming spectra that by they assuming are characterized that they byare yellowish characterized or reddish by yellowish colors. Asor reddish gossans colors.are Fe-oxides As gossans/hydroxides are Fe-oxides/hydroxides rich, the main absorption rich, the peaks main are absorption causes by peaks electronic are causes processes by electronic between processes0.45 µm and between 0.9 µm 0.45 for µm Fe2 and+, between0.9 µm for 0.9 Fe2+,µm between and 1.2 µ 0.9m µm for Fe3and+ 1.2, and µm vibrational for Fe3+, and processes vibrational for processesFe-OH bond for around Fe-OH 2.2 bondµm around [16–18]. 2.2 Mapping µm [16–18]. Fe rich Mapping formation Fe or rich gossans formation has been or donegossans for overhas been four donedecades for using over Landsatfour decades multispectral using Landsat scanners multispe [19,20],ctral or more scanners recently, [19,20], using or Landsat more 7recently, [3] or Landsat using Landsat8 [18,21, 227 ].[3] The or Landsat methods 8 used [18,21,22]. in these The studies methods consist used mostly in these of Principal studies consist Components mostly Analysis of Principal and Componentsbands ratios. Analysis For example and thebands ratio ratios. 660/490 For nm example (B3/B1) the of Landsat ratio 660/490 7 imagery nm clearly(B3/B1) highlightsof Landsat the 7 presenceimagery clearly of the Highhighlights Lake gossanthe presence in Canada of the [3 ].Hi Thegh 1610Lake/ 865gossan nm (B6in Canada/B5) of Landsat [3]. The 8 1610/865 imagery hasnm (B6/B5)also been of successfullyLandsat 8 imagery used to has highlight also been gossans successf inully Saudi used Arabia to highlight [21]. This gossans study in also Saudi shows Arabia that [21].gossans This can study behighlighted also shows that in Landsat gossans 8 can imagery be highlighted following ain principal Landsat 8 component imagery following (PC) analysis, a principal using componenta color composite (PC) analysis, of PC2 as using red, PC3a color as green,composite and of PC4 PC2 as as blue red, or PC3 by using as green, the components and PC4 as blue 2, 3, andor by 4 usingof a minimum the components noise fraction 2, 3, and (MNF) 4 of a minimum analysis. Annoise iron fraction feature (MNF) depth analysis. (IFD) index, An iron built feature on diff deptherent (IFD)ratios index, around built Fe2+ onand different Fe3+ absorption ratios around features, Fe2+ was and developed Fe3+ absorp totion map features, gossans [was22,23 developed]. to map However,gossans [22,23]. while these different techniques have shown to be useful to detect gossans in specific settings,However, none ofwhile them these appears different suitable techniques to identify have gossans shown with to be various useful to compositions detect gossans or associatedin specific withsettings, diff erentnone bedrockof them appears compositions. suitable A to test identify of these gossans methods with to various detect gossanscompositions mapped or associated in various withlocations different across bedrock the Canadian compositions. Arctic yielded A test of a verythese low methods success to rate. detect Indeed, gossans our mapped experiments in various based locationson band ratioingacross the and Canadian iron feature Arctic depth yielded (IFD) a indexvery low demonstrate success rate. that Indeed, gossans our are experiments undistinguishable based onfrom band randomly ratioing placed and iron points feature (see depth Section (IFD)4). index demonstrate that gossans are undistinguishable fromOn randomly the other placed hand, points machine (see Section learning 4). techniques can perform a task without being explicitly programmedOn the other to do hand, so. The machine most useful learning methods techniques derive from can neuralperform networks, a task without especially being deep explicitly learning programmedalgorithms, that to outperform do so. The other most machine useful learningmethods techniques. derive from Convolutional neural netw neuralorks, networksespecially (CNN) deep learninghave been algorithms, widely used that ever outperform since the initial othersuccesses machine obtainedlearning intechniques computer. visionConvolutional [24]. CNNs neural can networksextract specific (CNN) features have been from widely an image used with ever very since little the supervision. initial successes This type obtained of network in computer has reached vision a [24].success CNNs rate thatcan extract even surpasses specific humanfeatures recognition from an image capacity with [25 very]. Thanks little tosupervision. CNNs, companies This type have of networkdeveloped has powerful reached applicationsa success rate in that fields even such surpasses as image human recognition recogn [25ition,26], capacity object detection [25]. Thanks [27,28 to], CNNs,semantic companies segmentation have [ 27developed,29] amongst powerful others. applications In remote sensing in fields applications, such as image CNNs recognition has proved [25,26], to be objectefficient detection to identify [27,28], different semantic object classes segmentati (e.g., roads,on [27,29] buildings, amongst forest, orothers. water bodies)In remote or to segmentsensing applications,those within imagesCNNs [30has,31 proved]. to be efficient to identify different object classes (e.g., roads, buildings,In this forest, study, or wewater propose bodies) a or new to segment method those with within a few images steps to[30,31]. identify gossans of various compositionsIn this study, based we on apropose simple CNNa new architecture method with and a Landsat few steps 8 data. to identify A CNN isgossans trained of for various binary compositionsimage classification, based on and a simple the result CNN enables architecture the user and to determineLandsat 8 data. if a gossan A CNN is is present trained or for absent binary in imageeach Landsat classification, 8 image. and The the method result wasenables tested the using user to the determine locations ofif a hundreds gossan is of present gossans or previously absent in

Remote Sens. 2020, 12, 3123 3 of 16

Remote Sens. 2020, 12, x FOR PEER REVIEW 3 of 16 identiﬁed and mapped in the Canadian Arctic. Based on previous studies [18,21,32], Landsat 8 data may beeach used Landsat because 8 image. of itThe is spectrallymethod was rich tested and using there the is locations broad coverage of hundreds of Canada’sof gossans previously Arctic regions. The resultsidentified demonstrate and mapped that in the geo Canadian big data Arctic. [33] could Based be on used previous to identify studies gossans[18,21,32], in Landsat poorly-vegetated 8 data may be used because of it is spectrally rich and there is broad coverage of Canada’s Arctic regions. areas of the Canadian Arctic and other parts of the circum-Arctic landmass. The results demonstrate that geo big data [33] could be used to identify gossans in poorly-vegetated 2. Materialsareas of the and Canadian Methods Arctic and other parts of the circum-Arctic landmass.

2.1. Region2. Materials of Study and Methods Natural2.1. Region Resources of Study Canada has made considerable efforts to map Canada’s northern geology in the past decades.Natural Resources For example, Canada since has 2008, made the considerable geo-mapping efforts for to energy map Canada’s and minerals northern was geology developed in to help tothe unlock past decades. the full For mineral example, and since energy 2008, the potential geo-mapping in the for North, energy and and to minerals promote was responsible developed land developmentto help to [unlock34]. Consequently, the full mineral geological and energy mapping potential has in the been North, conducted and to promote in the field responsible by various land teams overdevelopment the years, and [34]. abundant Consequently, gossaneous geological exposures mapping werehas been identified. conducted However, in the field these by gossaneousvarious exposuresteams areover currently the years, reported and abundant in different gossaneous maps, reports, exposures and scientific were identified. publications. However, The firstthese step in our studygossaneous was to exposures identify theare locationcurrently ofreported gossanous in different deposits maps, in the reports, Canadian and scientific Arctic that publications. could be used The first step in our study was to identify the location of gossanous deposits in the Canadian Arctic to train and validate the model. This resulted in the identification of 809 exposures which had the that could be used to train and validate the model. This resulted in the identification of 809 exposures most precise coordinates for our study. These exposures were either reported with coordinates in a which had the most precise coordinates for our study. These exposures were either reported with publication,coordinates consisted in a publication, of shapefiles consisted in the GEOSCANof shapefilesdatabase, in the GEOSCAN or they coulddatabase, beeasily or they recognizable could be in Googleeasily Earth recognizable Pro or Earth in Google Explorer Earth based Pro onor theEarth figures Explorer included based on in eachthe figures geological included report. in each The 809 gossansgeological used inreport. this studyThe 809 are gossans grouped used into in this six study clusters are grouped as can be into seen six clusters in Figure as2 can. The be seen geological in mapFigure of the Arctic2. The geological bedrock compiledmap of the byArctic [35] bedroc was selectedk compiled to provideby [35] was context selected for to our provide study. context This map showsfor that our the study. gossans This frommap ourshows study that canthe begossan founds from on 35 our di ffstudyerent can lithologies. be found This on 35 wide different variety of geologicallithologies. contexts Thisis wide part variety of the diof ffigeologicalculty to identifycontexts them.is part Forof the instance, difficulty on to the identify AxelHeiberg them. For Island (clusterinstance, 6), most on ofthe the Axel gossans Heiberg are Island embedded (cluster in 6), sandstones, most of the shales,gossans orare anhydrites embedded fromin sandstones, the Mesozoic shales, or anhydrites from the Mesozoic era. The gossans from the cluster 5 are mostly in Neoarchean era. The gossans from the cluster 5 are mostly in Neoarchean mafic to intermediate volcanic flows mafic to intermediate volcanic flows and greywackes, some of them related to felsic intrusions. Most and greywackes, some of them related to felsic intrusions. Most of the gossans from the cluster 1 are of the gossans from the cluster 1 are in metamorphosed siliciclastic rocks dated from the in metamorphosedPaleoproterozoic. siliciclastic For more detailed rocks dated geological from setting the Paleoproterozoic.s, please read the description For more of detailed the geological geological settings,map please from [35]. read the description of the geological map from [35].

Figure 2. Location of the gossans used in this study with their associate cluster number to highlight

their geographical repartition (colors used for better visualization). Remote Sens. 2020, 12, 3123 4 of 16

2.2. Satellite Imagery Landsat 8 was launched on 11 February 2013. This satellite has 2 instruments called the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS). The OLI acquires data in nine spectral channels between 0.43 and 2.29 µm with a spatial resolution varying between 15 m (panchromatic channel) and 30 m (multispectral channels) per pixel (Table1). The TIRS acquires data in two spectral channels between 10.60 and 12.51 µm with a spatial resolution of 100 m per pixel. The data acquired by OLI were preferred because of their higher spatial resolution.

Table 1. Summary of the OLI bands from Landsat 8.

Spectral Range Spatial Resolution Band 1 Costal aerosol 0.43–0.45 µm 30 m Band 2 Blue 0.450–0.51 µm 30 m Band 3 Green 0.53–0.59 µm 30 m Band 4 Red 0.64–0.67 µm 30 m Band 5 Near-Infrared 0.85–0.88 µm 30 m Band 6 SWIR 1 1.57–1.65 µm 30 m Band 7 SWIR 2 2.11–2.29 µm 30 m Band 8 Panchromatic 0.50–0.68 µm 15 m Band 9 Cirrus 1.36–1.38 µm 30 m

All the OLI data used in our study belongs to the United States Geological Survey (USGS) Landsat 8 Surface Reflectance. The process to obtain surface reflectance images from raw data is described in the Landsat 8 Surface Reflectance Code (LASRC) product guide [36]. At high latitude, the atmospheric model used to correct the reflectance could be inappropriate because of high solar zenith angle. Consequently, objects with very high or very low reflectance can be affected and snow results in reflectance values higher than 1. It is a known computational artifact acknowledged by the USGS. Nonetheless, the spectra of other objects like rocks or vegetation seem correct. Landsat 8 was used as the data covered all our areas of study with no cloud and very small amount of snow. In Google Earth Engine (GEE), a script can be written to filter the dates of acquisition for the data. As the region of interest is far North, only images acquired between June 30 and September 30 from 2014 to 2019 were selected to minimize the presence of snow cover. A cloud filter allows to keep only the data with less than 2% of cloud cover. Finally, we chose 6 areas covering the 6 clusters of gossans. The smallest image (cluster 6) covers 9000 km2 and the largest (cluster 2) covers 97,000 km2. To highlight the need of a new method of gossaneous detection in our context, some results were compared with known examples of the literature [3,21]. Hence, we also downloaded OLI imageries of the High lake gossan in Canada and the Khunayqiyah gossans in Saudi Arabia. A Digital Elevation Model of Canada from GEE provided by Natural Resources Canada [37] was also downloaded. This DEM was derived by integrating various sources of elevation products between 1945 and 2011. The spatial resolution depends on the latitude but is in the same order as the OLI data (~30 m). The vertical resolution varies between 2 and 16 m and is not always known.

2.3. Methodology CNNs are built in layers, themselves made of neurons. The more the layers, the deeper the network. The training is the iterative adjustment of the weight and the bias for each neuron that minimizes the loss between the predicted output and the given label [38,39]. Each update is called an iteration and an epoch is all the iterations required to see every training samples. The number of epochs can vary depending on the obtained precision. The batch size controls the number of samples “seen” by the model at each iteration before it updates its parameters. The amount of change in Remote Sens. 2020, 12, 3123 5 of 16

Remote Sens. 2020, 12, x FOR PEER REVIEW 5 of 16 the weights and biases at each iteration is called learning rate. In some cases, overfitting can occur duringweights the and training, biases at especially each iteration with is a smallcalled traininglearning dataset,rate. In some so the cases, model overfitting won’t generalize can occur on during new data.the training, A way toespecially prevent overfittingwith a small is training by adding data a dropoutset, so the layer model [40 ].won’t This generalize method makes on new the data. layers A beway randomly to prevent treated-like overfitting layers is by with adding a different a dropou numbert layer of neurons [40]. This and method connectivity makes to the the previouslayers be layers.randomly The treated-like number of epochs,layers with the learninga different rate, numb theer batch of neurons size, and and the connectivity probability forto the dropoutprevious layerlayers. are The some number of the of hyperparameters epochs, the learning encountered rate, the in batch CNNs. size, The and values the probability for these hyperparameters for the dropout mustlayer beare empirically some of the found hyperparameters because they encountere are dependentd in ofCNNs. the task The to values be accomplished. for these hyperparameters Table2 shows themust range be empirically of values tested found in because the frame they of are our dependen work. t of the task to be accomplished. Table 2 shows the range of values tested in the frame of our work. Table 2. Hyperparameters values tested in the frame of our work.

Table 2. HyperparametersMin Values Tested values testedKept in the Values frame of our work. Max Values Tested Number of epochs Min 1 Values Tested Kept 5–35Values Max Values Tested 500 5 4 3 1 LearningNumber rate of epochs 10− 1 10 5–35− –10 − 500 10− −5 −4 −3 −1 Batch sizeLearning rate 16 10 10 –10 6410 256 Batch size 16 64 256 Dropout probability 0.1 0.4–0.7 0.9 Dropout probability 0.1 0.4–0.7 0.9

TheThe architecturearchitecture usedused inin thisthis study,study, builtbuilt onon pytorch, pytorch, is is inspired inspired from from [ 41[41].]. TheThe latterlatter obtainedobtained 95.13%95.13% ofof precisionprecision inin thethe classificationclassification ofof thethe SAT-6SAT-6 datasetdataset (barren(barren land,land, trees,trees, grassland,grassland, roads,roads, buildings,buildings, and and water water bodies). bodies). We We chose chose 2-D 2-D convolutional convolutional layers layers with with a kernela kernel size size of 3of and 3 and a stride a stride of 1.of Pooling 1. Pooling layers layers lie after lie after the first the two first convolution two convolution layer. Afterlayer. a After third convolutionala third convolutional layer, a dropoutlayer, a layerdropout is placed layer tois mitigateplaced to overfitting mitigate overfitting during the duri training.ng the After training. the flattening After the operation, flattening 3 operation, linear fully 3 connectedlinear fully layers connected were used.layers Thewere last used. one The has last a softmax one has binary a softmax exit forbinary the presenceexit for the or presence absence ofor gossanabsence on of the gossan tile, so on the the output tile, so is the a number output betweenis a number 0 and between 1 for each 0 and case. 1 for All each the other case. layersAll the have other a rectifiedlayers have linear a rectified unit (RELU) linear activation unit (RELU) function activation that outputs function directly that outputs the input directly value the if positive input value or 0 ifif negative.positive or The 0 if optimizer negative. is The the adaptativeoptimizer is moment the adaptative estimation moment (Adam) estimation [39] which (Adam) is a type [39] of which stochastic is a gradienttype of stochastic descent algorithm gradient anddescent the lossalgorithm criterion and is thethe rootloss meancriterion squared is the error root betweenmean squared predictions error andbetween labels. predictions The architecture and labels. used in The our studyarchitecture possesses used 93.954 in our trainable study parameterspossesses and93.954 is showntrainable in theparameters Figure3. and is shown in the Figure 3.

FigureFigure 3.3. RepresentationRepresentation ofof thethe convolutionalconvolutional neuralneural networknetwork architecturearchitecture usedused ininthis this work. work. ThisThis architecturearchitecture comprisescomprises 93.95493.954 trainingtraining parameters.parameters. SeeSee mainmain texttext forfor detailsdetails (colors(colors usedused forfor betterbetter visualization).visualization). To detect gossans in a variety of contexts, our goal was to develop a new and effective method To detect gossans in a variety of contexts, our goal was to develop a new and effective method involving few steps and a simple CNN architecture. As the quality of the output is dependent of the involving few steps and a simple CNN architecture. As the quality of the output is dependent of the quality of the input, we decided to test four different datasets as inputs. Here we present the main steps quality of the input, we decided to test four different datasets as inputs. Here we present the main that we undertook, following the method depicted in Figure3. Our method was tested using iteratively steps that we undertook, following the method depicted in Figure 3. Our method was tested using four different datasets as input in the CNN. The first one is the multispectral OLI images converted iteratively four different datasets as input in the CNN. The first one is the multispectral OLI images into reflectance values by the USGS [36]. We kept 7 bands as the cirrus band and the panchromatic converted into reflectance values by the USGS [36]. We kept 7 bands as the cirrus band and the band were excluded. The second dataset is the OLI images stacked with the DEM, as gossans are more panchromatic band were excluded. The second dataset is the OLI images stacked with the DEM, as resistant to erosion if a ferricrete is preserved, than the surrounding terrains. This suggests that a DEM gossans are more resistant to erosion if a ferricrete is preserved, than the surrounding terrains. This could provide a relevant information to discriminate the presence of gossans. The third dataset is an suggests that a DEM could provide a relevant information to discriminate the presence of gossans. The third dataset is an MNF dataset created from the 7 kept bands of the OLI images in ENVI (Environment for Visualizing Images). A MNF consists of two rotations of principal components analysis [42], which separate the information from the noise. A plot of the eigen values versus the

Remote Sens. 2020, 12, 3123 6 of 16

MNF dataset created from the 7 kept bands of the OLI images in ENVI (Environment for Visualizing Images). A MNF consists of two rotations of principal components analysis [42], which separate the information from the noise. A plot of the eigen values versus the eigen values numbers allows the user to select the number of significant bands that minimize noise and redundant information. Following a visual inspection, we decided to keep the first five MNF bands as the two others were noisy with eigen values close to one. The fourth dataset consists of an MNF dataset created from the OLI images and DEM. In this case, we also kept the first five bands for the same reason. All these data were masked for water and snow areas with ENVI. To train a CNN, one must construct matrices with their associated labels (presence or absence of gossan). Hence, each OLI image has to be cut in tiles. [43] express how important the context could be to classify a tile extracted from high resolution satellite imagery. This is also discussed in [44] and [31]. The authors explain that a tile should be large enough to give a context to the target, but small enough to not have different objects in it. Gossans can greatly vary in size from a few meters to kilometers. We decided to cut our images in 28 28 non-overlapping tiles (840 840 m). This area should be large × × enough to cover a gossan and its immediate surroundings. When a tile is built from the image, it is considered as positive if a known gossan is either on it or at less than 4 pixels from it. For instance, if a gossan is very close to the border of a tile, the neighboring one could present features corresponding to the presence of a gossan too. Not all the geographical range of the images is cut in tiles because they cover tens of thousands of squared kilometers. As 809 gossans are used, which represents a small fractional area of the OLI images, most of the tiles would be labeled as “negative”. For a balanced training dataset, an equal or nearly equal amount of positive and negative tiles should be identified. To achieve this, each time a positive tile is built, a negative tile is randomly built with no reported gossan on it (or closer than 4 pixels, or 120 m, from the tile). This method raises the question of the false negative tiles. Indeed, a gossan could be present and not being yet detected on a negative tile. We decided to test two approaches. The first approach considers that the chance to have a false negative is small. Indeed, the area covered by the images is large and gossans should be rare enough. Hence, the proportion of false negative in the training dataset should be acceptable and not confusing for the trained model. The second approach decreases the risk of false negative by using a lithological map. For the area of an image, all the known gossans are intersected with their underlying lithology. A histogram is built from the count of the gossans per lithology. Hence, all the negative tiles belonging to the lithology with the highest count are discarded. This should lower the risk of having false negative in the dataset. In the positive dataset, we noticed that gossans being often close to water bodies or sometimes snow, could be associated with no data values (masks). For this case, we also tried two approaches. The first approach considers that these cases occur rarely enough to not confuse the model. The second approach prevents tiles with more than a certain proportion of no data values to be included in the positive dataset. We tried different thresholds for the no data proportion accepted on a tile (25%, 50%, 100%) with and without lithological “filtering”. The size of the dataset to train a model is of a crucial importance. Indeed, a neural network needs a lot of examples to “learn” the characteristic features of the object and limit overfitting. Different ways are often used to “artificially” increase the amount of data such as cropping, rotating and flipping [45]. In our study, flipped tiles were used (up and down, left, and right) and 180◦ rotated tiles to multiply by 4 our original dataset. No cropping was done as we thought that tiles of 28 28 pixels were already × the lower limit. The normalization is also a commonly used process for the data preparation. To do so, the absolute minimum of the matrix is first added, so there are no more negative values as encountered in MNF images. The matrix is then divided by its maximum value, resulting in a matrix that ranges from 0 to 1. It is then subtracted by its mean and divided by its standard deviation. Hence, the matrix has a mean of 0 and a unit variance. Remote Sens. 2020, 12, 3123 7 of 16

Our entire dataset contains around 8000 (depending on filters choices) tiles after the augmentation. It is separated in 3 non equal datasets: the training, the validation, and the test datasets. The training dataset that is used to adjust the weights and the biases of the model. The validation dataset is used to evaluate the model between each epoch. The test dataset has never been seen by the model during the training. It allows, at the end of the training, an objective measure of the effectiveness of the obtained Remotemodel Sens. to perform 2020, 12, x the FOR assigned PEER REVIEW task. As these data are new to the model, it is equivalent to finding7 of new 16 gossans in unexplored areas never and verifying their presence or absence. It is like having the ground model, it is equivalent to finding new gossans in unexplored areas never and verifying their presence truth data collected before performing the prediction. In our study, we used 10%, 15%, and 20% of all or absence. It is like having the ground truth data collected before performing the prediction. In our the non-augmented data for the test dataset. The remaining tiles were separated as 80% and 20% for study, we used 10%, 15%, and 20% of all the non-augmented data for the test dataset. The remaining the training and the validation datasets respectively. tiles were separated as 80% and 20% for the training and the validation datasets respectively. To evaluate the model, the precision metric only was used. Indeed, the recall or the F1 score To evaluate the model, the precision metric only was used. Indeed, the recall or the F1 score need need the false negatives to be accounted. As discussed in the methodology section, the false negatives the false negatives to be accounted. As discussed in the methodology section, the false negatives cannot be identified so their use for evaluation could be misleading. Our method is summarized in cannot be identified so their use for evaluation could be misleading. Our method is summarized in the Figure4. the Figure 4.

FigureFigure 4. 4. FlowFlow chart chart summarizing summarizing the the method methodologyology followed followed in in this this study. study.

3.3. Results Results ToTo exploreexplore thethe variability variability in in the the predictions predictions based ba onsed di onfferent different model model configurations configurations (the 4 datasets (the 4 datasetswith the with different the different filtering conditions),filtering conditions), it was necessary it was necessary to choose to a relevantchoose a range relevant for therange number for the of numberepochs andof epochs the learning and the rates. learning Our trialsrates. showed Our tria thatls showed no significant that no improvementsignificant improvement in the precision in the of precisionthe model of could the model be achieved could after be achieved ~30 to 40 after epochs ~30 (Figure to 40 epochs4). For the(Figure learning 4). For rate, the we learning note that rate, values we between 10 10 4 and 1.2 10 3 allowed to get better results (Figure5). The batch size was set to 64 note that values× −between 10e× −4 and− 1.2e−3 allowed to get better results (Figure 5). The batch size was setas greaterto 64 as values greater decreased values precision decreased and precision smaller valuesand smaller considerably values increased considerably the computing increased time.the computingConsequently, time. we Consequently, focus the rest we of ourfocus analysis the rest on of the our CNN analysis results on the obtained CNN usingresults hyperparameter obtained using hyperparametervalues within these values ranges. within these ranges.

Remote Sens. 2020, 12, x FOR PEER REVIEW 8 of 16 RemoteRemote Sens. Sens.2020 2020, 12, 12, 3123, x FOR PEER REVIEW 8 of8 of 16 16

a b a b

Figure 5. Variability in the precision of the model using different configurations. (a) When the trained FigureFigure 5. 5.Variability Variability in in the the precision precision of of the the model model using using di differentfferent configurations. configurations. (a ()a When) When the the trained trained model is applied on the test dataset, one can see that there is no significant improvement in the modelmodel is appliedis applied on theon testthe dataset,test dataset, one can one see can that see there that is nothere significant is no signif improvementicant improvement in the precision in the precision after ~30 to 40 epochs. (b) The same is done while varying the learning rate values. The best afterprecision ~30 to after 40 epochs. ~30 to (40b) epochs. The same (b is) The done same while is varyingdone while the learningvarying the rate learning values. Therate bestvalues. results The are best −4 −3 results are obtained between4 103 −4 and 1.2 −3. obtainedresults are between obtained 10− betweenand 1.2− 10. and 1.2 .

TheTheThe Figure FigureFigure5 presents 55 presentspresents an anan example exampleexample of ofof a CNNaa CNNCNN result resultresult obtained obtainedobtained using usingusing hyperparameter hyperparameterhyperparameter values valuesvalues within withinwithin thethethe ranges rangesranges mentioned mentionedmentioned above. above.above. This ThisThis precise preciseprecise example exampleexample concerns concernsconcerns the thethe MNF MNFMNF dataset datasetdataset used usedused as as input,as input,input, without withoutwithout a lithologicalaa lithologicallithological ﬁlter, filter,filter, for aforfor “no aa “no“no data datadata threshold” threshold”threshold” of 25%, ofof 25%,25%, a probability aa probabilityprobability of 0.6 ofof for 0.60.6 the forfor dropout thethe dropoutdropout layer, layer,layer, and a andand test aa datasettesttest datasetdataset representing representingrepresenting 10% of 10%10% the ofof non-augmented thethe non-augmentednon-augmented data (228 datadata tiles). (228(228 tiles).tiles). Indeed, Indeed,Indeed, augmented augmentedaugmented data indatadata our inin test ourour datasettesttest datasetdataset would wouldwould have have beenhave been useless.been useless.useless. Figure FigureFigure6 shows 66 shsh theowsows variability thethe variabilityvariability of the ofof precisions, thethe precisions,precisions, depending dependingdepending on the onon chosenthethe chosenchosen hyperparameters. hyperparameters.hyperparameters.

Figure 6. Example of the overall CNN precisions obtained using number of epochs and lerning rates Figure 6. Example of the overall CNN precisions obtained using number of epochs and lerning rates valuesFigure within 6. Example the optimal of the ranges overall identified. CNN precisions obtained using number of epochs and lerning rates values within the optimal ranges identified. values within the optimal ranges identified. Overall, we produced 288 figures similar to Figure6 in order to calculate the precisions obtained Overall, we produced 288 figures similar to Figure 6 in order to calculate the precisions obtained using eachOverall, configuration we produced for 288 the figures datasets. similar These to configurations Figure 6 in order represent to calculate the four the precisions different datasets obtained using each configuration for the datasets. These configurations represent the four different datasets usedusing as each input, configuration the six filter configurationsfor the datasets. (no These lithological configurations filter, lithological represent the filter four each different using no datasets data used as input, the six filter configurations (no lithological filter, lithological filter each using no data thresholdused as input, of 100%, the nosix datafilter thresholdconfigurations of 50%, (no andlitho nological data filter, threshold lithological of 25%), filter 36 each combinations using no data of threshold of 100%, no data threshold of 50%, and no data threshold of 25%), 36 combinations of numbersthreshold of epochsof 100%, and no learning data threshold rates, three of 50%, sizes and (10%, no 15%, data and threshold 20% of theof 25%), non-augmented 36 combinations data) of of numbers of epochs and learning rates, three sizes (10%, 15%, and 20% of the non-augmented data) of thenumbers test dataset of epochs and four and values learning for rates, the dropout three sizes layer (10%, (0.4, 0.5,15%, 0.6, and 0.7). 20% To of summarize the non-augmented this information, data) of thethe testtest datasetdataset andand fourfour valuesvalues forfor thethe dropoutdropout layerlayer (0.4,(0.4, 0.5,0.5, 0.6,0.6, 0.7).0.7). ToTo summarizesummarize thisthis information,information,

Remote Sens. 2020, 12, x FOR PEER REVIEW 9 of 16 RemoteRemote Sens. Sens.2020 2020, 12, ,12 3123, x FOR PEER REVIEW 9 of9 of 16 16 we present the best and the worst precision obtained for each dataset considering all the different configurationswe present the as best a whole and the (Figure worst 7). precision This illustrates obtained the for variability each dataset in the considering precisions retrieved all the different during we present the best and the worst precision obtained for each dataset considering all the different theconfigurations process. Figure as a 8whole shows (Figure the same 7). Thisfor the illustrates six different the variability configurations. in the precisions retrieved during configurations as a whole (Figure7). This illustrates the variability in the precisions retrieved during the process. Figure 8 shows the same for the six different configurations. the process. Figure8 shows the same for the six di fferent configurations.

Figure 7. Precision obtained for each input dataset considering all the different model configurations FigureasFigure a whole. 7. 7.Precision Precision “Overall obtained obtained max” foris for the each each highest input input datasetprecisio dataset consideringn cons obtained.idering all“Overall all the the di differentﬀ mean”erent model ismodel the conﬁgurationsmean configurations precision asobtainedas a whole.a whole. with “Overall “Overall its standard max” max” is deviation. is the the highest highest “Ove precision precisiorall min” obtained.n obtained.is the lowest “Overall “Overall precision mean” mean” obtained. is is the the mean mean precision precision obtainedobtained with with its its standard standard deviation. deviation. “Overall “Overall min” min” is is the the lowest lowest precision precision obtained. obtained.

a b a b

c d c d

FigureFigure 8. 8.Precisions Precisions obtained obtained for for each each input input dataset dataset using using six six di differentfferent filter filter configurations. configurations. “Overall “Overall max”max”Figure is theis 8. the highestPrecisions highest Precision obtainedPrecision obtained. for obtained. each “Overall input “Overall dataset mean” mean” using is the mean sixis differentthe precision mean filter precision obtained configurations. withobtained its standard “Overallwith its deviation.standardmax” is “Overallthedeviation. highest min” “Overall Precision is the min” lowest obtained. isprecision the “Overalllowest obtained. pr mean”ecision NLF: isobtained. the no lithologicalmean NLF: precision no filter, lithological obtained LF: lithological filter, with LF: its filter,lithologicalstandard 1: no datadeviation. filter, threshold 1: no “Overall data 100%, threshold min” 0.5: no isdata 100%,the thresholdlowest 0.5: no pr dataecision 50%, threshold 0.25: obtained. no data 50%, NLF: threshold 0.25: no no lithological 25%.data threshold (a) Precisions filter, 25%. LF: obtained(lithologicala) Precisions on the filter, OLIobtained 1: dataset no dataon only. thethreshold (OLIb) Precisions dataset 100%, only.0.5: obtained no (b data) Precisions when threshold the OLIobtained 50%, dataset 0.25: when is no stacked data the thresholdOLI with dataset a DEM. 25%. is (cstacked)(a Precisions) Precisions with obtained a obtainedDEM. on (c the) onPrecisions OLIthe datasetOLI obtaineddataset after using only.on the an(b OLI) MNF Precisions data transform.set afterobtained ( dusing) Precisions when an MNF the obtained OLItransform. dataset on the ( dis) OLIPrecisionsstacked dataset with stackedobtained a DEM. with on ( cthe a) DEMPrecisions OLI afterdataset performingobtained stacked on withan the MNF a OLI DEM transform.data afterset performingafter using anan MNFMNF transform.transform. (d) Precisions obtained on the OLI dataset stacked with a DEM after performing an MNF transform.

Remote Sens. 2020, 12, 3123 10 of 16

4. Discussion To detect gossans in the Canadian Arctic, a CNN was trained multiple times with four different datasets (OLI, OLI+DEM, MNF, MNF+DEM) with varying hyperparameters values (number of epochs, learning rates, dropout probabilities). Figure7 demonstrates that performing an MNF before training the model, significantly improves the results. Indeed, the best precisions obtained for the OLI and OLI+DEM datasets are respectively 68% and 70%. Concerning the MNF and MNF+DEM datasets, the precisions are 77% and 75% respectively. These results suggest that redundant spatial information may confuse the network. Consequently, further testing could be done with less than five MNF bands used herein. This figure also demonstrates that, contrary as expected, the DEM did not bring relevant information to detect gossans with a CNN. The increase of 2% for the highest precisions in the cases OLI and OLI+DEM should not be considered as significant. Moreover, considering only the mean results for these datasets (56% and 56.8%), the difference is negligible. This could be explained by the resolution of the DEM used, the spatial resolution (30 m), and the vertical resolution (2–16 m) not being enough to detect bumps associated with gossan presence. Still, the experiment should be conducted again with a DEM having a higher resolution. There is no obvious trend that could be extracted from the tests with different configurations (Figure8). It seems that no significant change occurs when applying or not a no data threshold. Most of our tiles contain lakes that were masked. We wanted to verify that a model would not associate gossans with no data on a general trend. With no threshold, the number of tiles used in our dataset was 2276. With the lowest threshold (25%), we still could use 1984 tiles. It appears that this difference, 292 tiles representing 13% of the original dataset, was not critical regarding other aspects discussed below. Hence, no influence of the no data has been demonstrated. The use of a lithological filtering seems to decrease by a few percent the precision of our results. It is not what was expected as the idea was to decrease the risk of false negative in our dataset by avoiding tiles on lithologies associated with gossans. Two reasons to explain this result could be proposed. It is possible that avoiding large areas led to a set of tiles that was not representative of our whole data. Hence, the model could have been confused by a biased negative dataset. It is also possible that this unexpected result came from the way we chose the lithology to avoid. Indeed, we avoided tiles on lithologies where we found the highest count of gossans. We were aware that a bias could arise from this method: the larger the area, the higher the count. This way, we avoided lithologies not necessarily associated with gossans. So, we also tried to normalize these counts by the lithologies areas. Instead of having a gossan count “per lithology”, we obtained a gossan count “per lithology per surface unit”. But a huge artefact arose from this method. Some lithologies were present in small areas in our six clusters. Hence, even with a few gossans on these lithologies, their counts per unit of area were drastically exaggerated. So, this normalization was discarded. In the literature, it is common to see precisions higher than 90% [31,44] with several classes. What could explain that our best model reached 77% with two classes? Our first approach to detect the presence of gossans in the Canadian Arctic was using known bands ratios to highlight them. The band ratio the 660/490 nm (B4/B2 with Landsat 8) from [3] and the ratio 1610/865 nm (B6/B5) from [21] were tested. The use of the IFD index from [22,23] was also explored. This index highlights the absorption feature of iron in the near infrared (NIR) band (positive IFD values indicate an absorption). It is calculated by:

IFD = r (NIR) r(NIR) (1) int − where IFD denotes the iron feature depth index and r the reﬂectance. rint is the interpolated reﬂectance calculated by:

λ(900nm) λ(750nm) rint(900nm) = r(750nm) + r(1250nm) r(750nm) − (2) − ∗ λ(1250nm) λ(750nm) − Remote Sens. 2020, 12, 3123 11 of 16 Remote Sens. 2020, 12, x FOR PEER REVIEW 11 of 16

OneOne can can see see on Figureon Figure9 that 9 that these these methods methods are highlyare highly e ffective effective in the in twothe two examples examples taken taken from from thethe literature. literature. Gossans Gossans are are clearly clearly visible visible on on each each case, case, particularly particularly with with the the IFDIFD index.index. However, However, the theresults on our sixsix clustersclusters areare lessless convincingconvincing (Figure(Figure 10 10).). KnownKnown gossansgossans inin our our database database seem seem to to showshow no no correlation correlation with with the the band band ratio ratio or or the the IFD. IFD. To To highlight highlight this this observation, observation, we dispatchedwe dispatched randomrandom points points on theon the images images and and then then extracted extracted their their ratio ratio and and IFD IFD values. values. A scatter A scatter plot plot showing showing B4/B2B4/B2 versus versus IFD IFD values values allows allows to compare to compare the repartitionthe repartition of the of gossansthe gossans and and the randomthe random points points in a in a 2D space.2D space. On On five five of ourof our six clusters,six clusters, the the gossans gossan ares undistinguishableare undistinguishable from from the the randomly randomly placed placed points.points. Only Only the clusterthe cluster six off erssix aoffers possible a possible separability separability (Figure 11 (Figure). Two explanations 11). Two explanations can be assumed. can be First,assumed. most of ourFirst, gossans most of are our covered gossans by are some covered vegetation. by some Second, vegetation. the occurrences Second, the of theoccurrences gossans are of the smallgossans compared are small to the compared spatial resolution, to the spatial making resolution their spectral, making contribution their spectral in a mixelcontribution (not spectrally in a mixel pure(not pixel) spectrally too small pure to pixel) be detected too small by to these be detected methods. by Indeed,these methods. spectra Indeed, of some spectra gossans of randomlysome gossans chosenrandomly show thatchosen some show of themthat some had aof vegetation them had a spectrum vegetation morphology spectrum morphology (Figure 11). (Figure Low spatial 11). Low resolutionspatial andresolution vegetation and vegetation cover could cover also could explain also why explain the band why ratioing the band and ratioing the IFD and index the IFD failed index to highlightfailed to highlight gossans in gossans our region in our of region study. of Hence, study. the Hence, CNN the must CNN classify must classi tiles despitefy tiles despite showing showing no apparentno apparent spectral spectral common common features. features. Moreover, Moreover, the non-augmented the non-augmented data data represent represent 2276 2276 tiles tiles only. only. ButBut training training a CNN a CNN requires requires a big a amountbig amount of labeled of labeled data. data. For For a comparison, a comparison, the the SAT-4 SAT-4 dataset dataset has has 400,000400.000 images images [44]. [44]. Also, Also, the the tiles tiles from from SAT-4 SAT-4 have have a high a high spatial spatial resolution resolution of 1 of m 1 per m per pixel pixel showing showing detaileddetailed features features that that could could be associated be associated with with the labels.the labels.

a b

1 km

c d

800 m

FigureFigure 9. B4 9./ B2B4/B2band band ratio ratio (a) and(a) and IFD IFD (b) applied(b) applied on theon the Khunayqiyah Khunayqiyah gossans gossans in Saudi in Saudi Arabia Arabia (long: (long: 45.0794653;45.0794653; lat: lat: 24.2805029). 24.2805029).B4 /B4/B2B2 band band ratio ratio (c) and(c) and IFD IFD (d) applied(d) applied on theon Highthe High Lake Lake gossan gossan in in CanadaCanada (bottom; (bottom; long: long:110.8473047; −110.8473047; lat: lat: 67.3826313). 67.3826313). The The gossans gossans are are highlighted highlighted by theby brightestthe brightest − pixelspixels in both in both band band ratio ratio and and IFD. IFD.

Remote Sens. 2020, 12, 3123 12 of 16

Remote Sens. 2020, 12, x FOR PEER REVIEW 12 of 16

a b

4 km

c d

2 km

Figure 10. : B4/B2 band ratio (a) and IFD (b) applied on a part of our cluster 5 (top; long: - Figure 10. B4/B2 band ratio (a) and IFD (b) applied on a part of our cluster 5 (top; long: 112.2914268; 112,2914268; lat: 66,59684054). B4/B2 band ratio (c) and IFD (d) applied on a part of our cluster− 6 lat: 66.59684054).(Bottom; long:B4/ B2-93,5884band ; lat: ratio 79,3667). (c) and The IFD red dots (d) appliedrepresent the on known a part gossans of our locations. cluster 6For (Bottom; the long: 93.5884; lat:band 79.3667). ratio and the The IFD, red the dots brightest represent pixels should the known have highlighted gossans locations.the gossans. For the band ratio and the − Remote Sens. 2020, 12, x FOR PEER REVIEW 13 of 16 IFD, the brightest pixels should have highlighted the gossans.

a b

c d

FigureFigure 11. Mean 11. Mean spectrum spectrum of all of gossans all gossans with with 6 randomly 6 randomly chosen chosen spectra spectra for for the the clusters clusters 55 ((a) and 6 6 (b) (Oﬀset( forb) (Offset clarity). for Vegetationclarity). Vegetation spectrum spectrum added added for comparison for comparison on the on subﬁgurethe subfigure (a). (a Scatter). Scatter plot plot of of IFD valuesIFD versus values band versus ratio band values ratio for values the gossans for the gossans and randomly and randomly placed placed points points for the for clustersthe clusters 5 ( c5) ( andc) 6 (d). A slightand 6 ( separationd). A slight separation is possible is on possible the latter. on the latter.

As most of the gossans presented in this work do not show recognizable patterns of iron oxides in their reflectance spectra and are undistinguishable from randomly placed points, we think that achieving 77% of precision is a substantial improvement compared to PCA and band ratioing.

5. Conclusions The remote detection of gossans in the Canadian Artic region to prioritize targets is the utmost importance for different fields of study such as mineralogy, mining exploration, microbiology, or exobiology. In this endeavor, we used satellite imagery from the Landsat 8 sensor OLI and 809 known gossans occurrences to construct labeled tiles (presence or absence of gossan). These data were used to train a convolutional neural network to identify possible presence of a gossan in a tile. Our work demonstrated that performing a minimum noise fraction before training the model was useful and increased the precision up to 77% on the test dataset (228 tiles). We showed that thresholding the no data proportion on tiles did not improve the precision. Results obtained from the literature were compared with our data. We showed that gossans, assumed as occurrences of yellow and red hues, was too restrictive to apply in our case. For example, a scatter plot of the values of different band ratios for gossans versus random points showed how challenging the detection of gossans could be in our region of study. For these reasons, achieving 77% precision is a success and the method deserves further attention. The results obtained herein could potentially be refined by modifying the input datasets or the method. For example, higher spatial resolution imagery and digital elevation models could be used as inputs. Such heavy datasets could be processed using tools such as “Google Colab” that offer

Remote Sens. 2020, 12, 3123 13 of 16

As most of the gossans presented in this work do not show recognizable patterns of iron oxides in their reﬂectance spectra and are undistinguishable from randomly placed points, we think that achieving 77% of precision is a substantial improvement compared to PCA and band ratioing.

5. Conclusions The remote detection of gossans in the Canadian Artic region to prioritize targets is the utmost importance for different fields of study such as mineralogy, mining exploration, microbiology, or exobiology. In this endeavor, we used satellite imagery from the Landsat 8 sensor OLI and 809 known gossans occurrences to construct labeled tiles (presence or absence of gossan). These data were used to train a convolutional neural network to identify possible presence of a gossan in a tile. Our work demonstrated that performing a minimum noise fraction before training the model was useful and increased the precision up to 77% on the test dataset (228 tiles). We showed that thresholding the no data proportion on tiles did not improve the precision. Results obtained from the literature were compared with our data. We showed that gossans, assumed as occurrences of yellow and red hues, was too restrictive to apply in our case. For example, a scatter plot of the values of different band ratios for gossans versus random points showed how challenging the detection of gossans could be in our region of study. For these reasons, achieving 77% precision is a success and the method deserves further attention. The results obtained herein could potentially be refined by modifying the input datasets or the method. For example, higher spatial resolution imagery and digital elevation models could be used as inputs. Such heavy datasets could be processed using tools such as “Google Colab” that offer online GPUs computing capacity. Moreover, hyperspectral imagery is now available in certain regions since the launch of the Precursore Iperspettrale Della Missione Applicativa (PRISMA) mission in 2019 by the Italian Space Agency. This mission provides free hyperspectral imagery, in 237 spectral channels between 400 and 2505 nm, that could provide less redundant information than the OLI dataset for the minimum noise fraction transform used herein and could thus potentially increase the precision of the model. However, the spatial coverage of this dataset is currently limited. In a future study, we propose to use convolutional neural networks and evidential fusion on X data to increase the precision of the model.

Author Contributions: Conceptualization, É.C. and M.L.; methodology, É.C.; M.L. and M.G.; validation, É.C.; M.L.; M.G. and M.-C.W.; formal analysis, É.C.; investigation, É.C.; resources, É.C., M.L.; M.G. and M.-C.W.; data curation, É.C., É.B.; writing—original draft preparation, É.C.; writing—review and editing, É.C.; M.L.; M.G. and M.-C.W.; visualization, É.C., É.B.; supervision, M.L.; project administration, M.L.; funding acquisition, M.L. All authors have read and agreed to the published version of the manuscript. Funding: This research was undertaken thanks to funding from the Université de Sherbrooke and from the Canada Research Chairs Program. Professor Lemelin is the Canada Research in Northern and Planetary Geological Remote Sensing (950-232175). Acknowledgments: The authors are grateful to Mathieu Turgeon-Pelchat and the four anonymous reviewers for comments that improved the original manuscript. NRCan contribution no. 20200423. Conﬂicts of Interest: The authors declare no conﬂict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

References

1. Velasco,F.; Herrero, J.M.; Suárez, S.; Yusta,I.; Alvaro, A.; Tornos,F.Supergene features and evolution of gossans capping massive sulphide deposits in the Iberian Pyrite Belt. Ore Geol. Rev. 2013, 53, 181–203. [CrossRef] 2. West, L.; McGown, D.J.; Onstott, T.C.; Morris, R.V.; Suchecki, P.; Pratt, L.M. High Lake gossan deposit: An Arctic analogue for ancient Martian surﬁcial processes? Planet. Space Sci. 2009, 57, 1302–1311. [CrossRef] 3. Harris, J.R.; Williamson, M.-C.; Percival, J.B.; Behnia, P.; Macleod, R. Detecting and Mapping Gossans Using Remotely-Sensed Data. In Environmental and Economic Signiﬁcance of Gossans; Geological Survey of Canada: Ottawa, ON, Canada, 2015; p. 7718. Remote Sens. 2020, 12, 3123 14 of 16

4. Essalhi, M.; Sizaret, S.; Barbanson, L.; Chen, Y.; Lagroix, F.; Demory, F.; Nieto, J.M.; Sáez, R.; Capitán, M.Á. A case study of the internal structures of gossans and weathering processes in the Iberian Pyrite Belt using magnetic fabrics and paleomagnetic dating. Min. Depos. 2011, 46, 981–999. [CrossRef] 5. Hunt, J.; Lottermoser, B.G.; Parbhakar-Fox, A.; Van Veen, E.; Goemann, K. Precious metals in gossanous waste rocks from the Iberian Pyrite Belt. Miner. Eng. 2016, 87, 45–53. [CrossRef] 6. Peterson, R.C.; Williamson, M.-C.; Rainbird, R.H. Gossan Hill, Victoria Island, Northwest Territories: An analogue for mine waste reactions within permafrost and implication for the subsurface mineralogy of Mars. Earth Planet. Sci. Lett. 2014, 400, 88–93. [CrossRef] 7. Williamson, M.-C. Environmental and Economic Significance of Gossans; Comission Géologique du Canada: Ottawa, ON, Canada, 2015; p. 100. [CrossRef] 8. Cruz, C.; Noronha, F.; Santos, P.; Mortensen, J.K.; Lima, A. Supergene gold enrichment in the Castromil-Serra da Quinta gold deposit, NW Portugal. Mineral. Mag. 2018, 82, S307–S320. [CrossRef] 9. Valente, T.; Rivera, M.J.; Almeida, S.F.P.; Delgado, C.; Gomes, P.; Grande, J.A.; de la Torre, M.L.; Santisteban, M. Characterization of water reservoirs affected by acid mine drainage: Geochemical, mineralogical, and biological (diatoms) properties of the water. Environ. Sci. Pollut. Res. 2016, 23, 6002–6011. [CrossRef][PubMed] 10. Santos, E.S.; Abreu, M.M.; Macías, F.; de Varennes, A. Chemical quality of leachates and enzymatic activities in Technosols with gossan and sulfide wastes from the São Domingos mine. J. Soils Sediments 2016, 16, 1366–1382. [CrossRef] 11. Santos, E.S.; Abreu, M.M.; Macías, F.; Magalhães, M.C.F. Potential environmental impact of technosols composed of gossan and sulfide-rich wastes from São Domingos mine: Assay of simulated leaching. J. Soils Sediments 2017, 17, 1369–1383. [CrossRef] 12. Santos, E.S.; Abreu, M.M.; Macías, F. Rehabilitation of mining areas through integrated biotechnological approach: Technosols derived from organic/inorganic wastes and autochthonous plant development. Chemosphere 2019, 224, 765–775. [CrossRef] 13. Shuster, J.; Reith, F.; Izawa, M.; Flemming, R.; Banerjee, N.; Southam, G. Biogeochemical Cycling of Silver in Acidic, Weathering Environments. Minerals 2017, 7, 218. [CrossRef] 14. Hedrich, S.; Schippers, A. Distribution of Acidophilic Microorganisms in Natural and Man-made Acidic Environments. Curr. Issues Mol. Biol. 2020, 25–48. [CrossRef] 15. Floyd, M.A.M.; Williams, A.J.; Grubisic, A.; Emerson, D. Metabolic Processes Preserved as Biosignatures in Iron-Oxidizing Microorganisms: Implications for Biosignature Detection on Mars. Astrobiology 2019, 19, 40–52. [CrossRef][PubMed] 16. Clark, R.N. Spectroscopy of Rocks and Minerals, and Principles of Spectroscopy. In Manual of Remote Sensing; Remote Sensing for the Earth Sciences; Andrew, N.R., Ed.; John Wiley and Sons: New York, NY, USA, 1999; Volume 3, pp. 3–58. 17. Laakso, K.; Rivard, B.; Rogge, D. Enhanced detection of gossans using hyperspectral data: Example from the Cape Smith Belt of northern Quebec, Canada. ISPRS J. Photogramm. Remote Sens. 2016, 114, 137–150. [CrossRef] 18. Beiranvand Pour, A.; S Park, T.Y.; Park, Y.; Hong, J.K.; M Muslim, A.; Läufer, A.; Crispini, L.; Pradhan, B.; Zoheir, B.; Rahmani, O.; et al. Landsat-8, Advanced Spaceborne Thermal Emission and Reflection Radiometer, and WorldView-3 Multispectral Satellite Imagery for Prospecting Copper-Gold Mineralization in the Northeastern Inglefield Mobile Belt (IMB), Northwest Greenland. Remote Sens. 2019, 11, 2430. [CrossRef] 19. Abrams, M.J.; Ashley, R.P.; Rowan, L.C.; Goetz, A.F.H.; Kahle, A.B. Mapping of hydrothermal alteration in the Cuprite mining district, Nevada, using aircraft scanner images for the spectral region 0.46 to 2.36 µm. Geology 1977, 5, 713–718. [CrossRef] 20. Abrams, M.J.; Brown, D.; Lepley, L.; Sadowski, R. Remote sensing for porphyry copper deposits in southern Arizona. Econ. Geol. 1983, 78, 591–604. [CrossRef] 21. Gahlan, H.; Ghrefat, H. Detection of Gossan Zones in Arid Regions Using Landsat 8 OLI Data: Implication for Mineral Exploration in the Eastern Arabian Shield, Saudi Arabia. Nat. Resour. Res. 2018, 27, 109–124. [CrossRef] 22. Mielke, C.; Muedi, T.; Papenfuss, A.; Boesche, N.K.; Rogass, C.; Gauert, C.D.K.; Altenberger, U.; de Wit, M.J. Multi- and hyperspectral spaceborne remote sensing of the Aggeneys base metal sulphide mineral deposit sites in the Lower Orange River region, South Africa. S. Afr. J. Geol. 2016, 119, 63–76. [CrossRef] Remote Sens. 2020, 12, 3123 15 of 16

23. Mielke, C.; Boesche, N.; Rogass, C.; Kaufmann, H.; Gauert, C.; de Wit, M. Spaceborne Mine Waste Mineralogy Monitoring in South Africa, Applications for Modern Push-Broom Missions: Hyperion/OLI and EnMAP/Sentinel-2. Remote Sens. 2014, 6, 6790–6816. [CrossRef] 24. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet classification with deep convolutional neural networks. Commun. ACM 2017, 60, 84–90. [CrossRef] 25. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. 26. Lin, M.; Chen, Q.; Yan, S. Network in Network. Available online: https://arxiv.org/pdf/1312.4400.pdf (accessed on 4 March 2014). 27. Girshick, R.; Donahue, J.; Darrell, T.; Malik, J. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. In Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, USA, 23–28 June 2014; pp. 580–587. 28. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE Trans. Pattern Anal. Mach. Intell. 2017, 39, 1137–1149. [CrossRef][PubMed] 29. Dai, J.; He, K.; Sun, J. Instance-Aware Semantic Segmentation via Multi-task Network Cascades. In Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 3150–3158. 30. Cheng, G.; Yang, C.; Yao, X.; Guo, L.; Han, J. When Deep Learning Meets Metric Learning: Remote Sensing Image Scene Classification via Learning Discriminative CNNs. IEEE Trans. Geosci. Remote Sens. 2018, 56, 2811–2821. [CrossRef] 31. Liu, Q.; Basu, S.; Ganguly, S.; Mukhopadhyay, S.; DiBiano, R.; Karki, M.; Nemani, R. DeepSat V2: Feature augmented convolutional neural nets for satellite image classification. Remote Sens. Lett. 2020, 11, 156–165. [CrossRef] 32. Ducart, D.F.; Silva, A.M.; Toledo, C.L.B.; Assis, L.M. de Mapping iron oxides with Landsat-8/OLI and EO-1/Hyperion imagery from the Serra Norte iron deposits in the Carajás Mineral Province, Brazil. Braz. J. Geol. 2016, 46, 331–349. [CrossRef] 33. Li, S.; Dragicevic, S.; Castro, F.A.; Sester, M.; Winter, S.; Coltekin, A.; Pettit, C.; Jiang, B.; Haworth, J.; Stein, A.; et al. Geospatial big data handling theory and methods: A review and research challenges. ISPRS J. Photogramm. Remote Sens. 2016, 115, 119–133. [CrossRef] 34. GEM: Geo-mapping for Energy and Minerals. Available online: https://www.nrcan.gc.ca/earth-sciences/ resources/federal-programs/geomapping-energy-minerals/18215 (accessed on 22 April 2019). 35. Harrison, J.; St-Onge, M.; Petrov,O.; Strelnikov,S.; Lopatin, B.; Wilson, F.; Tella, S.; Paul, D.; Lynds, T.; Shokalsky,S.; et al. Geological Map of the Arctic; Geological Survey of Canada: Ottawa, ON, Canada, 2011. [CrossRef] 36. Zanter, K. Landsat 8 Surface Reflectance Code (LASRC) Product Guide; USGS: Sioux Falls, SD, USA, 2019; p. 39. 37. Canadian Digital Elevation Model, 1945–2011. Available online: https://open.canada.ca/data/en/dataset/ 7f245e4d-76c2-4caa-951a-45d1d2051333 (accessed on 29 January 2020). 38. Bottou, L. Large-Scale Machine Learning with Stochastic Gradient Descent. In Proceedings of the COMPSTAT’2010, Paris, France, 22–27 August 2010; Lechevallier, Y., Saporta, G., Eds.; Physica-Verlag HD: Heidelberg, Germany; pp. 177–186. 39. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. In Proceedings of the 3rd International Conference on Learning Representations, San Diego, CA, USA, 7–9 May 2015. 40. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J. Mach. Learn. Res. 2014, 15, 1929–1958. 41. Adivarekar, B. Simple Keras CNN with 95.3% Accuracy. Available online: https://www.kaggle.com/ bhumitadivarekar/simple-keras-cnn-with-95-13-accuracy (accessed on 15 January 2020). 42. Green, A.A.; Berman, M.; Switzer, P.; Craig, M.D. A transformation for ordering multispectral data in terms of image quality with implications for noise removal. IEEE Trans. Geosci. Remote Sens. 1988, 26, 65–74. [CrossRef] 43. Hutchison, D.; Kanade, T.; Kittler, J.; Kleinberg, J.M.; Mattern, F.; Mitchell, J.C.; Naor, M.; Nierstrasz, O.; Pandu Rangan, C.; Steffen, B.; et al. Learning to Detect Roads in High-Resolution Aerial Images. In Computer Vision – ECCV 2010; Daniilidis, K., Maragos, P., Paragios, N., Eds.; Lecture Notes in Computer Science; Springer: Berlin/Heidelberg, Germany, 2010; Volume 6316, pp. 210–223. ISBN 978-3-642-15566-6. Remote Sens. 2020, 12, 3123 16 of 16

44. Basu, S.; Ganguly, S.; Mukhopadhyay, S.; DiBiano, R.; Karki, M.; Nemani, R. DeepSat: A learning framework for satellite imagery. In Proceedings of the 23rd SIGSPATIAL International Conference on Advances in Geographic Information Systems - GIS ’15, Seattle, WA, USA, 3–6 November 2015; ACM Press: New York, NY, USA; pp. 1–10. 45. Perez, L.; Wang, J. The Eﬀectiveness of Data Augmentation in Image Classiﬁcation using Deep Learning. Available online: https://arxiv.org/pdf/1712.04621.pdf (accessed on 13 December 2017).

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).