A High-Resolution Cropland Map for the West African Sahel Based on High-Density Training Data, Google Earth Engine, and Locally Optimized Machine Learning

The West African Sahel Cropland map (WASC30) is a new 30-m cropland extent product for the nominal year of 2015. We used the computing resources provided by Google Earth Engine (GEE) to fit and apply Random Forest models for cropland detection in each of 189 grid cells (composed of 100 km2, hence a total of ~1.9 × 106 km2) across five countries of the West African Sahel (Burkina Faso, Mauritania, Mali, Niger, and Senegal). Landsat-8 surface reflectance (Bands 2–7) and vegetation indices (NDVI, EVI, SAVI, and MSAVI), organized to include dry-season and growing-season band reflectances and vegetation indices for the years 2013–2015, were used as predictors. Training data were derived from an independent, high-resolution, visually interpreted sample dataset that classifies sample points across West Africa using a 2-km grid (~380,000 points were used in this study, with 50% used for model training and 50% used for model validation). Analysis of the new cropland dataset indicates a summed cropland area of ~316 × 103 km2 across the 5 countries, primarily in rainfed cropland (309 × 103 km2), with irrigated cropland area (7 × 103 km2) representing 2% of the total cropland area. At regional scale, the cropland dataset has an overall accuracy of 90.1% and a cropland class (rainfed and irrigated) user’s accuracy of 79%. At bioclimatic zones scale, results show that land proportion occupied by rainfed agriculture increases with annual precipitation up to 1000 mm. The Sudanian zone (600–1200 mm) has the highest proportion of land in agriculture (24%), followed by the Sahelian (200–600 mm) and the Guinean (1200 +) zones for 15% and 4%, respectively. The new West African Sahel dataset is made freely available for applications requiring improved cropland area information for agricultural monitoring and food security applications.


Introduction
Timely and accurate information on cultivated areas is of paramount importance for food security planning [1,2]. This is particularly true in developing regions, like the West African Sahel, where most cropland is rainfed and agricultural production is susceptible to fluctuations in precipitation [3]. Earth Observation (EO) satellites can contribute significantly to providing information to the agricultural sector, as they allow for consistent land surface imaging over broad spatial extents (regionally or globally) with high revisit frequency [4]. That makes these technologies suitable for monitoring vegetation [5], cropland area [6][7][8], and agricultural production [9][10][11][12]. Optical remote sensing in 3 of 22 multiple non-agricultural classes, and agricultural classes including rainfed and irrigated cropland. The approach, based on expert visual interpretation, with specific local knowledge of the environments being classified, is expected to show better results than semi-or fully automated classifiers, particularly for the cropland class across West Africa [32].
Quality control for the reference data was carried out using multiple sources of ancillary data, including thousands of aerial photographs taken by the USGS team, high-resolution verification using Google Earth satellite imagery, and field validation in each country, facilitating systematic verification of land cover assessments [30]. In addition, image interpretation and land cover assessments carried out by national experts were reviewed and revised during regular collaborative workshops in West Africa, to ensure consistent practice between country teams and USGS partners. Further details are provided by Samasse et al. [15].
In this study, we regrouped the 25 land cover classes into 3 classes (rainfed and irrigated agriculture and non-agricultural) and used 50% of the 2 km by 2 km data points for year 2013 as reference information for training the classification algorithm and the other 50% for assessing the classified product. Reduced data-density in some areas (e.g., on the coastal and desert margins) resulted in a total of 383,464 reference data points (non-crop, rain-fed, and irrigated classes) across our West Africa study domain.

Google Earth Engine (GEE)
Google Earth Engine is a cloud-based platform for regional and planetary scale earth observation data retrieval and processing. Its advantage is to store the petabytes of freely available data (e.g., Landsat imagery) in the cloud, avoiding the need for data download, while providing high-performance parallel computing resources to process large datasets [26]. GEE thus facilitates computationally cumbersome geospatial analysis with minimal local computing and storage resources. GEE makes use of an application programming interface (API in JavaScript or Python), allowing for data processing and visualization at different scales. The GEE platform also implements several Machine Learning algorithms (Support Vector Machine, Random Forest) known to be effective for land cover and land use classification in general and cropland mapping in particular [8,[33][34][35]].

Landsat 8 Surface Reflectance (SR)
The Landsat mission is a joint initiative of the USGS and NASA providing consistent earth observation data at sub-100 m spatial resolution since the 1970s. Surface reflectance data from the Landsat-8 OLI (Operational Land Imager) and TIRS (Thermal Infrared Sensor) is available in GEE from April 2013 to present. Table 1 contains information on the Landsat-8 SR Tier 1 collection spectral bands used in this study.

Vegetation Indices
In addition to the individual band reflectances, remote sensing derived vegetation indices (VI) have been extensively applied to detect vegetation and monitor vegetation condition over large areas. These indices are generally based on the capability of vegetation to strongly reflect incident electromagnetic signal in the near-infrared (NIR) band compared to the optical bands. In this study we calculated four vegetation indices as candidate predictor variables for the RF classification algorithm to help separate crop and non-crop zones.
The NDVI can effectively detect growing vegetation [39] but gets quickly saturated in high biomass surfaces. In such conditions, other vegetation indices like EVI (Enhanced Vegetation Index) have been proposed to replace or supplement the NDVI.

EVI
The Enhanced Vegetation Index (EVI), described by Equation (2) provides improved sensitivity to vegetation condition and changes in high biomass areas as compared to the NDVI and also reduces the background effect of soil on vegetation index calculation [40,41]. In addition to the red and near-infrared bands, EVI includes in the calculation the blue band (B2) to correct atmospheric effects of aerosol.
where G is a gain factor; C1 and C2 are the coefficients of the aerosol resistance term, which uses blue band B2 to correct for aerosol influences in the red band B4, and L is the soil-adjustment factor as in SAVI. In this we used the coefficients adopted in the MODIS EVI algorithm, which are L = 1, C1 = 6, C2 = 7.5, and G = 2.5 [42].

SAVI
The soil-adjusted vegetation index (SAVI; Huete, 1988) was developed to compensate for the effects of the soil background in sparsely vegetated areas. Equation (3) is the commonly used expression of SAVI with a soil adjustment factor L. This factor is found to reduce soil noise using the value L = 0.5 for a wide range of vegetation classes [43].

MSAVI
The Modified Soil-Adjusted Vegetation Index (MSAVI; [44]; Equation (4)) was proposed as an improved version of SAVI that minimizes the effect of bare soil [44].

Random Forest (RF)
We used the random forest (RF) technique as the main classification algorithm in this study. The RF model is an ensemble learning algorithm that can be used to predict both continuous (regression) and categorical (classification) responses. For a classification problem, the response variable is a class which links certain independent values to one of the categories present in the dependent variable [45]. An RF Remote Sens. 2020, 12, 1436 5 of 22 model comprises an ensemble of decision trees, where each tree constitutes a classifier, which can predict the response-variable using a random sub-sample of the independent variables and observations. Each tree uses a sub-ensemble of training values chosen randomly with replacement (i.e., bootstrap sample). The optimum number of predictors used to split data at each tree's node is log(m+1), where m is the total number of predictors involved. An ensemble of diverse trees minimizes the effect of bias from individual trees considerably improving the overall predictive accuracy of the model. The final class prediction is chosen by a maximum vote (classification). It has been shown that by increasing the number of trees in the model, the errors of prediction (also known as out-of-bag errors or OOB errors) converge, reducing problems with overfitting [45]. In this study, we used OOB error estimation during the training process to finetune RF model parameters and provide internal cross-validation before independent accuracy assessment.

Gridding and Accuracy Metrics
A grid of 100 km by 100 km squares was created using ArcMap based on the extent of the available training data. In total, 267 squares (labelled from S1 to S267 in Figure 1) were generated covering the study area. For simplicity, Figure 1 shows the positions and labels of the first and last grid-squares. Satellite image (Landsat) data for each square was classified independently using the RLCM reference data to train and evaluate local RF models.
Remote Sens. 2020, 12, x FOR PEER REVIEW 5 of 24 and categorical (classification) responses. For a classification problem, the response variable is a class which links certain independent values to one of the categories present in the dependent variable [45]. An RF model comprises an ensemble of decision trees, where each tree constitutes a classifier, which can predict the response-variable using a random sub-sample of the independent variables and observations. Each tree uses a sub-ensemble of training values chosen randomly with replacement (i.e., bootstrap sample). The optimum number of predictors used to split data at each tree's node is log(m+1), where m is the total number of predictors involved. An ensemble of diverse trees minimizes the effect of bias from individual trees considerably improving the overall predictive accuracy of the model. The final class prediction is chosen by a maximum vote (classification). It has been shown that by increasing the number of trees in the model, the errors of prediction (also known as out-of-bag errors or OOB errors) converge, reducing problems with overfitting [45]. In this study, we used OOB error estimation during the training process to finetune RF model parameters and provide internal cross-validation before independent accuracy assessment.

Gridding and Accuracy Metrics
A grid of 100 km by 100 km squares was created using ArcMap based on the extent of the available training data. In total, 267 squares (labelled from S1 to S267 in Figure 1) were generated covering the study area. For simplicity, Figure 1 shows the positions and labels of the first and last grid-squares. Satellite image (Landsat) data for each square was classified independently using the RLCM reference data to train and evaluate local RF models. Classification accuracy in this study was measured using the following metrics: Quantity disagreement (Q), Allocation disagreement (A), Overall Accuracy (OA), and class-specific measures such as User's Accuracy (UA) for Crop class, as suggested by Pontius and Millones [46]. Appendix A4 gives more details on disagreements analysis. The new dataset was also validated using detailed local field surveys conducted on agricultural activities at IPR/IFRA (Institut Polytechnique Rural de Formation et de Recherche Appliquée), a higher education institution in Mali. Figure 2 illustrates the steps employed in developing the cropland extent map, using the GEE platform and Random Forest machine learning approach. Landsat-8 images for each 16-day period were processed for each ~1 o grid cell. The Tier 1 Landsat-8 image collection was filtered spatially (Sahel grid level) and temporally (years 2013, 2014, and 2015, to match training data) before being filtered for clouds using the pixel_qa information (Table 1). Cloud-free images were used to compute vegetation indices in GEE using custom functions in JavaScript.

Workflow
In total, twenty bands were exported from GEE as candidate model predictors. Predictors included Landsat 8 surface reflectance bands (B2-B7) and VI averages for growing season (e.g., B2) and dry season (e.g., B2_1), with growing season defined from July-October of each year and November-June considered the dry season. In total, we have twelve (12) surface reflectance and eight (8) vegetation indices (Table 2). Classification accuracy in this study was measured using the following metrics: Quantity disagreement (Q), Allocation disagreement (A), Overall Accuracy (OA), and class-specific measures such as User's Accuracy (UA) for Crop class, as suggested by Pontius and Millones [46]. Appendix A.4 gives more details on disagreements analysis. The new dataset was also validated using detailed local field surveys conducted on agricultural activities at IPR/IFRA (Institut Polytechnique Rural de Formation et de Recherche Appliquée), a higher education institution in Mali. Figure 2 illustrates the steps employed in developing the cropland extent map, using the GEE platform and Random Forest machine learning approach. Landsat-8 images for each 16-day period were processed for each~1 o grid cell. The Tier 1 Landsat-8 image collection was filtered spatially (Sahel grid level) and temporally (years 2013, 2014, and 2015, to match training data) before being filtered for clouds using the pixel_qa information (Table 1). Cloud-free images were used to compute vegetation indices in GEE using custom functions in JavaScript.

Workflow
In total, twenty bands were exported from GEE as candidate model predictors. Predictors included Landsat 8 surface reflectance bands (B2-B7) and VI averages for growing season (e.g., B2) and dry season (e.g., B2_1), with growing season defined from July-October of each year and November-June considered the dry season. In total, we have twelve (12)    The land-cover training data were reclassified to produce three classes corresponding to rain-fed crops (level "1"), irrigated crops (level "2") and the non-crop class (level "0"), with levels 1 and 2 combined as needed to make up the "Crop" class. The training samples were derived by sampling 50% of "Crop" and 50% of "Non-crop" classes selected randomly within each grid cell, representing a stratified random sampling approach. R software was used to fit the RF classifiers external to GEE to benefit from greater model-fitting flexibility in R. Optimal fitted RF models were then used for regional predictions. Predictions of crop (rainfed and irrigated) and non-crop classes were based on the best-fit models, optimal parameters, and non-correlated predictors from the tuning process (see Appendix A.1 for more details on optimal RF parameters, and Appendix A.3 for removing correlated variables). Classification outputs were initially assessed at grid-cell level using independent reference samples (i.e., samples not used for training and/or OOB error estimation), then grouped at country and regional levels.

Predictors
The Landsat-8 Tier 1 image collection available on GEE for the study area comprised 6803 image scenes as filtered for the years 2013, 2014, and 2015. Depending on the location and the time, the number of available images changes. This is due, for example, to the degree of cloud coverage in different years and locations. Figure 3 shows that the availability of Landsat images increases in 2014 and 2015, relative to 2013, with the increase related to launch and partial Landsat-8 collection in 2013.

Predictors
The Landsat-8 Tier 1 image collection available on GEE for the study area comprised 6803 image scenes as filtered for the years 2013, 2014, and 2015. Depending on the location and the time, the number of available images changes. This is due, for example, to the degree of cloud coverage in different years and locations. Figure 3 shows that the availability of Landsat images increases in 2014 and 2015, relative to 2013, with the increase related to launch and partial Landsat-8 collection in 2013. The candidate predictor variables are shown in Figure 4. Values of surface reflectance (SR) in Landsat bands are between 0 and 1, while vegetation index (VI) values range between -1 and 1, as expected. On average, the shortwave infrared 1 band (B6) has the highest reflectance value in both wet and dry periods, probably due to minimal atmospheric attenuation in this part of the electromagnetic spectrum and low surface vegetation moisture on average in the savanna areas, which would otherwise lower SWIR reflectance. The second highest value occurs at the near infrared band (B5). This band also shows the most pronounced difference between wet and dry means, showing its sensitivity to green vegetation that is mostly present in the wet seasons. All the vegetation indices ( Figure 4) show net distinctions between the wet and dry periods, particularly in the range of electromagnetic spectrum and low surface vegetation moisture on average in the savanna areas, which would otherwise lower SWIR reflectance. The second highest value occurs at the near infrared band (B5). This band also shows the most pronounced difference between wet and dry means, showing its sensitivity to green vegetation that is mostly present in the wet seasons. All the vegetation indices ( Figure 4) show net distinctions between the wet and dry periods, particularly in the range of wet season values. The candidate predictor variables are shown in Figure 4. Values of surface reflectance (SR) in Landsat bands are between 0 and 1, while vegetation index (VI) values range between -1 and 1, as expected. On average, the shortwave infrared 1 band (B6) has the highest reflectance value in both wet and dry periods, probably due to minimal atmospheric attenuation in this part of the electromagnetic spectrum and low surface vegetation moisture on average in the savanna areas, which would otherwise lower SWIR reflectance. The second highest value occurs at the near infrared band (B5). This band also shows the most pronounced difference between wet and dry means, showing its sensitivity to green vegetation that is mostly present in the wet seasons. All the vegetation indices ( Figure 4) show net distinctions between the wet and dry periods, particularly in the range of wet season values.

Reclassified Training Data
Absence of cropland in the training data examined for some grid squares prevented fitting meaningful local models in these regions. These grid squares are therefore assumed to have little or no agriculture ( Figure 5). Some cells, particularly in the northern drylands lacked any training data (RF algorithm requires >1 class in the training data). In total, 189 cells (~71% of the study domain) include some amount of cropland. The other 78 cells (white cells in Figure 5) are mainly located in the Northern Sahel and Sahara, where agricultural activities are absent (or occur only intermittently). On average, 2028 reference data on presence of rainfed and irrigated cropland and non-cropland were available in each of the 189 retained grid cells (~1014 for model training and ~1014 for error assessment).

Accuracy at Grid Level
Results show an average overall accuracy (OA) above 80%, with most 100 km squares having an OA in the range of 75% to 100% (Table 3). Despite the relatively high OA, the reliability of classified product is best measured in terms of the users' accuracy, which quantifies accuracy from the perspective of the user of the classified product. In total, 11% of assessed cells had a user's accuracy of less than 50%; 58% were between 50% and 75%, and the remaining 31% had a user's accuracy above 75% (Table 3). On average, accuracy at grid level was 78.8% and 56.6% for OA and UA, respectively. Please refer to Appendix A2 of this document for further details on grid level assessment.

Reclassified Training Data
Absence of cropland in the training data examined for some grid squares prevented fitting meaningful local models in these regions. These grid squares are therefore assumed to have little or no agriculture ( Figure 5). Some cells, particularly in the northern drylands lacked any training data (RF algorithm requires >1 class in the training data). In total, 189 cells (~71% of the study domain) include some amount of cropland. The other 78 cells (white cells in Figure 5) are mainly located in the Northern Sahel and Sahara, where agricultural activities are absent (or occur only intermittently). On average, 2028 reference data on presence of rainfed and irrigated cropland and non-cropland were available in each of the 189 retained grid cells (~1014 for model training and~1014 for error assessment).

Accuracy at Grid Level
Results show an average overall accuracy (OA) above 80%, with most 100 km squares having an OA in the range of 75% to 100% (Table 3). Despite the relatively high OA, the reliability of classified product is best measured in terms of the users' accuracy, which quantifies accuracy from the perspective of the user of the classified product. In total, 11% of assessed cells had a user's accuracy of less than 50%; 58% were between 50% and 75%, and the remaining 31% had a user's accuracy above 75% (Table 3). On average, accuracy at grid level was 78.8% and 56.6% for OA and UA, respectively. Please refer to Appendix A.2 of this document for further details on grid level assessment.

Accuracy at Country Level
Assessment at country-scales indicates that the overall accuracy is around 90% for all the countries, except for Burkina Faso where it is slightly lower at~77% ( Figure 6). The country of Mauritania has the highest overall accuracy of 99%, but the accuracy to reliably identify crop class from the user's perspective in Mauritania is only about 71%, which is the lowest among the five countries.

Accuracy at Country Level
Assessment at country-scales indicates that the overall accuracy is around 90% for all the countries, except for Burkina Faso where it is slightly lower at ~77% (Figure 6). The country of Mauritania has the highest overall accuracy of 99%, but the accuracy to reliably identify crop class from the user's perspective in Mauritania is only about 71%, which is the lowest among the five countries. Assuming 75% as targeted value for crop user's accuracy, Mauritania is the only country where the classification performance fails to meet expectations. Highest accuracies occurred in Niger, with 85.6% accuracy, followed by Senegal, with 84.5%. Crop user's accuracy in Mali and Burkina Faso was between 75% and 80%.
In terms of crop area estimation, results show that rainfed agriculture is far more common than irrigated agriculture in the 5 countries, with irrigated cropland occupying only ~2% of the total (Figure 7). Cropland area is greatest in Niger (with 119 × 10 3 km 2 of cropland, representing 37.6% of the total agricultural area in the five Sahelian countries), followed by Burkina Faso (91 × 10 3 km 2 ; 28.8%), Mali (67 × 10 3 km 2 ; 21.3%), Senegal (38 × 10 3 km 2 ; 12.1%), and finally Mauritania (0.6 × 10 3 km 2 ; Assuming 75% as targeted value for crop user's accuracy, Mauritania is the only country where the classification performance fails to meet expectations. Highest accuracies occurred in Niger, with 85.6% accuracy, followed by Senegal, with 84.5%. Crop user's accuracy in Mali and Burkina Faso was between 75% and 80%. In terms of crop area estimation, results show that rainfed agriculture is far more common than irrigated agriculture in the 5 countries, with irrigated cropland occupying only~2% of the total (Figure 7). Cropland area is greatest in Niger (with 119 × 10 3 km 2 of cropland, representing 37.6% of the total agricultural area in the five Sahelian countries), followed by Burkina Faso (91 × 10 3 km 2 ; 28.8%), Mali (67 × 10 3 km 2 ; 21.3%), Senegal (38 × 10 3 km 2 ; 12.1%), and finally Mauritania (0.6 × 10 3 km 2 ; 0.21%) where agriculture is confined to the south of the country and the Senegal River Valley (Table 4 and Figure 7).

Discussion
The 30 m West African Sahel Cropland map (WASC30) covers five Sahelian countries of West Africa and shows in much more detail than previously available the agricultural zones of West Africa, including the "breadbasket" regions of Niger, Mali, Burkina Faso, and Senegal that are critical to the food security and economies at national and regional scales. We leveraged a distributed and dense sample dataset on actual land cover [29], with Landsat-8 data, to train locally optimized machinelearning predictors for rainfed and irrigated agriculture using the Google Earth Engine (GEE) platform. Earlier cropland products, covering the Sahelian region, generally combine irrigated and rainfed agriculture into a single cropland class, with accuracies generally less than 70% [15]. The average user's accuracy of the new crop extent map, considering the five countries of Burkina Faso, Mali, Mauritania, Niger, and Senegal ( Figure 6), is 79%, which is a considerable improvement relative to the best performing earlier products (GlobeLand30, 69%, and GFSAD30, 64%; [15]). Our accuracy statistic is also influenced by low accuracy in Mauritania, representing less than 1% of cropland area in the region (Figure 7). The user's accuracy for Senegal, Mali, Burkina Faso, and Niger (excluding Mauritania) is 81% for the new WASC30 cropland area map. The low accuracy reported for Mauritania is consistent with our previous findings in a study comparing accuracy of cropland classes in 12 pre-existing landcover products (Samasse et al., 2018). Explanations for this include the particularly small size of farms and the low intensity of agricultural activities in this country. However, it must be noted that, despite having lower accuracy compared to other countries, the cropland estimates for Mauritania in our new WASC30 map are an improvement on the pre-existing products.  Expressed as fraction of the total irrigated area in the five countries, more than the half of the total irrigated areas are in Mali (69.0%), particularly in the "Office du Niger" region, which is one of the oldest and largest irrigation schemes in West Africa [47]. The country of Mauritania, with less than 1% in rainfed agriculture area, has a larger share (4.4%) of the irrigated cropland in the region, more for example than in Burkina Faso with only 3.4% of irrigated cropland (Table 4 and Figure 7).

Discussion
The 30 m West African Sahel Cropland map (WASC30) covers five Sahelian countries of West Africa and shows in much more detail than previously available the agricultural zones of West Africa, including the "breadbasket" regions of Niger, Mali, Burkina Faso, and Senegal that are critical to the food security and economies at national and regional scales. We leveraged a distributed and dense sample dataset on actual land cover [29], with Landsat-8 data, to train locally optimized machine-learning predictors for rainfed and irrigated agriculture using the Google Earth Engine (GEE) platform. Earlier cropland products, covering the Sahelian region, generally combine irrigated and rainfed agriculture into a single cropland class, with accuracies generally less than 70% [15]. The average user's accuracy of the new crop extent map, considering the five countries of Burkina Faso, Mali, Mauritania, Niger, and Senegal (Figure 6), is 79%, which is a considerable improvement relative to the best performing earlier products (GlobeLand30, 69%, and GFSAD30, 64%; [15]). Our accuracy statistic is also influenced by low accuracy in Mauritania, representing less than 1% of cropland area in the region (Figure 7). The user's accuracy for Senegal, Mali, Burkina Faso, and Niger (excluding Mauritania) is 81% for the new WASC30 cropland area map. The low accuracy reported for Mauritania is consistent with our previous findings in a study comparing accuracy of cropland classes in 12 pre-existing landcover products (Samasse et al., 2018). Explanations for this include the particularly small size of farms and the low intensity of agricultural activities in this country. However, it must be noted that, despite having lower accuracy compared to other countries, the cropland estimates for Mauritania in our new WASC30 map are an improvement on the pre-existing products.

Irrigated Cropland
Based on the estimated crop areas (Table 4), irrigated land represents just 2% of the total cropland area. Thus, a specific accuracy is not reported for this sub-class of "Crop". However, Figure 8 shows clearly the intensive irrigation activities in Senegal and Mauritania adjacent to the Senegal River, in Mali in the "Office du Niger" zone, and in Niger adjacent to the Niger River. Irrigated croplands in the region are generally supported by hydroelectric dams on the major rivers (e.g., Niger, Senegal), providing both electricity and increased agricultural production. For example, the Diama dam in Senegal and the Markala dam in Mali are two operational hydroelectric infrastructures promoting intensive irrigated crop production in the Senegal valley and the Office du Niger zone in Mali, respectively [48,49]. In Mauritania, 44% of the 664 km 2 mapped as cropland is irrigated.

Irrigated Cropland
Based on the estimated crop areas (Table 4), irrigated land represents just 2% of the total cropland area. Thus, a specific accuracy is not reported for this sub-class of "Crop". However, Figure  8 shows clearly the intensive irrigation activities in Senegal and Mauritania adjacent to the Senegal River, in Mali in the "Office du Niger" zone, and in Niger adjacent to the Niger River. Irrigated croplands in the region are generally supported by hydroelectric dams on the major rivers (e.g., Niger, Senegal), providing both electricity and increased agricultural production. For example, the Diama dam in Senegal and the Markala dam in Mali are two operational hydroelectric infrastructures promoting intensive irrigated crop production in the Senegal valley and the Office du Niger zone in Mali, respectively [48,49]. In Mauritania, 44% of the 664 km 2 mapped as cropland is irrigated.

Intensive Rainfed Cropland Zones
Analysis of the more dominant rainfed cultivated areas shows several "hot spots" of intensive agricultural activities (Figure 9). For example, the Seno Plain (red circle), east of the Dogon Plateau in

Intensive Rainfed Cropland Zones
Analysis of the more dominant rainfed cultivated areas shows several "hot spots" of intensive agricultural activities (Figure 9). For example, the Seno Plain (red circle), east of the Dogon Plateau in Mali, has been devoted to intensive agricultural activities since the 1930s [50]. Recent studies using Earth Observation data have reported cropland expansion in this region driven by the need to feed a rapidly increasing population with accelerated expansion between 2000 and 2013 facilitated by modern technology [32]. Rapid population growth and conducive soils, with development of processing infrastructure, have also contributed to the high density of rainfed cropland in south-eastern Niger (blue circle). This area, known as the Tarka Plain and Goulbi Agricultural Zone, in the Maradi-Zinder region of Niger, is considered the most important agricultural zone of Niger (CILSS, 2016). It is an area of enormous agricultural potential, mainly in rainfed cropland [51]. Cereal (Millet, Maize, Sorghum, and Rice) cultivation is practiced, with more advanced systems in the Tarka plain, where the rural population density is particularly high. Because of the anthropogenic pressure, it is common to see an integrated system where agriculture, livestock, and forests share the same space [52]. Similarly, the West-Central Agricultural Zone in Senegal (black circle), known as the Peanut Basin (Bassin Arachidier) for the suitability of dominant soils to grow peanuts, is also characterized by high rural populations, with rainfed agriculture focused on cultivation of peanut, millet, sorghum, and beans [53].
Remote Sens. 2020, 12, x FOR PEER REVIEW 13 of 24 common to see an integrated system where agriculture, livestock, and forests share the same space [52]. Similarly, the West-Central Agricultural Zone in Senegal (black circle), known as the Peanut Basin (Bassin Arachidier) for the suitability of dominant soils to grow peanuts, is also characterized by high rural populations, with rainfed agriculture focused on cultivation of peanut, millet, sorghum, and beans [53].

Cropland Distribution Relative to Climate and Climate Zones
For the purpose of this work, we divided the study area based on the annual precipitation, following a steep gradient of decreasing rainfall from south to north. Figures 10 and 11  Above 1000 mm MAP, rainfed cropland proportion declines (Figure 11), in part due to shift to forest production and in part since wetter forested zones may have soils unsuitable for agriculture. However, irrigated cropland area is largely decoupled from MAP, being clustered around the flood plains of the perennial rivers in West Africa.
The Saharan desert region (MAP < 200 mm/y; Figure 10) constitutes about 61% of the total study area ( Table 5). Significant part of northern Mauritania, Mali, and Niger fall in this region. It is generally characterized by an arid climate with high average temperatures, a very low relative humidity, and rare and highly irregular precipitation, making difficult for crops to grow. However, irrigated farming may be present in some areas using appropriate irrigation technologies [54,55], mainly for small scale production of vegetables. Figure 11 shows our results illustrating the very low to non-existent agricultural activities in the Sahara.

Cropland Distribution Relative to Climate and Climate Zones
For the purpose of this work, we divided the study area based on the annual precipitation, following a steep gradient of decreasing rainfall from south to north. Figures 10 and 11 show the distribution of both rainfed and irrigated cropland as located in the West African Sahel Cropland map (WASC30) under 100 mm rainfall bins. Mean Annual Precipitation (MAP) is derived from eleven years (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015) of CHIRPS (Climate Hazards Group InfraRed Precipitation with Station) data retrieved from Google Earth Engine. Results show that rainfed cropland area generally increases with MAP between 200 and 1000 mm MAP, reflecting water limitations to agricultural activities in the arid zones (200-400 mm) and more suitable conditions in the South Sahel and Soudan (400-1000 mm). Above 1000 mm MAP, rainfed cropland proportion declines (Figure 11), in part due to shift to forest production and in part since wetter forested zones may have soils unsuitable for agriculture. However, irrigated cropland area is largely decoupled from MAP, being clustered around the flood plains of the perennial rivers in West Africa. The second largest zone is the Sahelian (200-600 mm), occupying 23% of the total study area. Rainfed cropland intensity increases with annual rainfall (Figure 11). Compared to the other climatic zones, the Sahel has the highest proportion of irrigated cropland, as irrigation activities along both Niger and Senegal rivers occur mainly in this climatic zone ( Figure 10). This irrigation proportion is, however, less than 1% against about 15% for rainfed agriculture (Table 5). Figure 11. Rainfed and irrigated cropland as percentage of the total area in MAE interval (e.g., The total area of 800-900 mm zone is about 100 × 10 3 km 2 , and the fraction of this area occupied by rainfed cropland is about 30%). The second largest zone is the Sahelian (200-600 mm), occupying 23% of the total study area. Rainfed cropland intensity increases with annual rainfall (Figure 11). Compared to the other climatic zones, the Sahel has the highest proportion of irrigated cropland, as irrigation activities along both Niger and Senegal rivers occur mainly in this climatic zone ( Figure 10). This irrigation proportion is, however, less than 1% against about 15% for rainfed agriculture (Table 5). Figure 11. Rainfed and irrigated cropland as percentage of the total area in MAE interval (e.g., The total area of 800-900 mm zone is about 100 × 10 3 km 2 , and the fraction of this area occupied by rainfed cropland is about 30%). Figure 11. Rainfed and irrigated cropland as percentage of the total area in MAE interval (e.g., The total area of 800-900 mm zone is about 100 × 10 3 km 2 , and the fraction of this area occupied by rainfed cropland is about 30%).
The Saharan desert region (MAP < 200 mm/y; Figure 10) constitutes about 61% of the total study area ( Table 5). Significant part of northern Mauritania, Mali, and Niger fall in this region. It is generally characterized by an arid climate with high average temperatures, a very low relative humidity, and rare and highly irregular precipitation, making difficult for crops to grow. However, irrigated farming may be present in some areas using appropriate irrigation technologies [54,55], mainly for small scale production of vegetables. Figure 11 shows our results illustrating the very low to non-existent agricultural activities in the Sahara. The second largest zone is the Sahelian (200-600 mm), occupying 23% of the total study area. Rainfed cropland intensity increases with annual rainfall (Figure 11). Compared to the other climatic zones, the Sahel has the highest proportion of irrigated cropland, as irrigation activities along both Niger and Senegal rivers occur mainly in this climatic zone ( Figure 10). This irrigation proportion is, however, less than 1% against about 15% for rainfed agriculture (Table 5).
Among the four climatic zones (Figure 10), rainfed agriculture activities are most intensive in the Sudanian zone (600-1200 mm). Representing 15% of the study area, the Sudanian is the third largest climatic zone, after the Sahara and the Sahel. About 24% of this climatic zone is occupied by rainfed cropland (Table 5). It covers major cereal production zones in Mali and Burkina Faso and southern parts of the Peanut Basin in Senegal ( Figure 10). The precipitation range is also suitable for cash crops (e.g., cotton), root crops, and mixed cereal-root system (e.g., cassava, yam, sweet potato, particularly in Southern Mali). Irrigation is not common in the Sudanian zone, largely due to the low occurrence of main rivers in the region.
At more than 1200 mm MAP, the Guinean zone covers little of the total area of the study domain (less than 1%). In this region, some 4% is occupied by rainfed cropland, dominated by root crop cultivation (yams, sweet potatoes, cassava) [56].

Fallows in WASC30
The visual interpretation approach adopted for the 2 km RLCM dataset [29,31] and used as training or reference information in this study classified large and long-term fallows as savanna. However, in the more intensive rainfed cropland regions with reduced fallow periods, small areas of fallow were generally classified as active agricultural land use. Overall, therefore, we consider that the WASC30 represents active agriculture, inclusive of short-term fallow fields but exclusive of longer-term fallow (or abandoned) areas that have not been actively cropped in recent years. That makes the final cropland class a reliable reference for developing active cropland extent.

Validation Using Local Scale Data
At local scale, our new cropland dataset has been assessed using recent (2012) GPS field surveys mapping land use and land cover at an Agricultural College (IPR/IFRA (Institut Polytechnique Rural de Formation et de Recherche Appliquée)) in the town of Koulikoro, just north of Bamako. Data was collected in collaboration with Laval University (Quebec, Canada) as part of the PACM research project (Des arbres et des champs contre la pauvreté au Mali). IPR/IFRA is a higher education institution in Mali managing an area of about 380 ha, including experimental farms and other lands for cereals and tree crop production. Figure 12 shows that the WASC30 captures the distribution of cultivated areas at IPR/IFRA with an area ratio of 195 ha/199 ha = 98% (i.e., the WASC30 product underestimates cropland area at this field station by 2%). This slight difference could be attributed to the small size of some sparse experimental plots making difficult their detection in the 30 m Landsat data. No irrigated pixels were detected at IPR/IFRA, which is consistent with the absence of irrigation trials at the site.

Conclusions
In this study, the Random Forest ensemble learning method has been applied to individual 100 km grid cells to develop a 30 m Landsat-derived active cropland dataset across five Sahelian countries with unprecedented details and higher accuracy as compared to existing land cover products. The developed dataset has an overall accuracy of 90.1% and a cropland class (rainfed and irrigated) user's accuracy of 79%.
Information derived from the new dataset reveals the total cropland area in West African Sahel to be 316 × 10 3 km 2 with 7 × 10 3 km 2 irrigated and 309 × 10 3 km 2 rainfed. This confirms that agriculture in Sahelian West Africa is almost entirely rainfed. The Sudanian zone (600-1200 mm) comprises most of the rainfed cultivated areas, while the Sahelian areas in proximity to main rivers present the highest proportion of irrigated land. Results also show that these irrigation activities in the region remain not well developed, comprising only about 2% of the total cropland area, despite the tremendous potential offered by, for example, the Senegal and Niger rivers. This may be due to the lack of well-developed infrastructure for irrigation and high investment costs to manage water and make it available where it is most needed. More efforts in developing irrigated land in Sahel region would expand farmers' production opportunities by reducing risks linked to climate fluctuations.
This study benefitted from the large and regularly distributed RLCM training dataset that allowed us to fit locally optimized random forest models in each of 189 grid cells (each 100 × 100 km) across the five-country study domain. This allowed us to minimize the effects of soil, topographic, and climatic differences that would increase errors in models fit at coarser regional and continental scales, thus improving overall accuracy of the final WASC30 product.
Geospatial data in general, and Landsat time series in particular, provide a critical source of information for the important task of producing accurate statistics on cultivated areas, particularly in developing countries where timely accurate georeferenced agricultural data are sometimes missing. The new cropland dataset will contribute to filling this void in West Africa Sahel.
Data Availability: The West African Sahel Cropland Dataset (WASC30) is currently available for visualization as a Google Earth Engine Application (https://savannalabnmsu.users.earthengine.app/view/wa-cropmap-30m). Further inquiries on data availability including download options should be addressed to the corresponding author. Acknowledgments: Our particular thanks to Dr. Gray Tappan and colleagues at USGS EROS Data Center who made available the West African Land Cover time series database (https://pubs.er.usgs.gov/publication/fs20173004) that was invaluable as training and validation for this analysis.

Conflicts of Interest:
The authors declare no conflict of interest.

Appendix A.
Appendix A. 1

. Tuning RF Major Parameters
Separate RF models were developed for each of the 189 cropland cells. Correlated predictors were removed following the example illustrated by Figure A4 before fitting models and tuned using the "tuneRF" function in R software. This tuning function helps determine the best number of variables available for splitting at each tree node (mTry) for a number of trees (nTree) based on the minimum values of the Out-Of-Bag (OOB) errors. The chosen values of nTree to run "tuneRF" included 500, 1000, 1500, and 2000. Occurrences of nTree corresponding to the optimal mTry are reported in Figure A1A. It appears that RF models (classifiers) show better performance in 44% of the cells for nTree = 500 and 15% of the cells for nTree = 2000. Between these two limits, 24% and 17% of the cells have shown minimal errors of OOB at nTree = 1000 and nTree = 1500, respectively. Frequency distribution of resulting optimum mTry values from the tuning process is shown on Figure A1B. Numbers 2, 1, 4, 8, 16, and 3 have been used as mTry values to fit the best models for the classification.
The Random Forest (RF) is widely accepted as an efficient ensemble approach for land cover classification using remotely sensed data. It handles well imbalanced data, missing values, and outliers [57]. However, tuning RF two major parameters (number of trees: nTree, number of variables available for splitting at each tree node: mTry) to get optimum values may be time and resource consuming, even in parallel processing environments like Google Earth Engine. In this study, we selected nTree in {500, 1000, 1500, 2000} for reduced computational time while ensuring sufficient trees for model convergence (Breiman, 2001). The best mTry for most of the grid cells has been achieved with nTree = 500, others for nTree = 2000, which are the limits of tuned nTree values ( Figure A1A). Using a larger range of nTree values (e.g., including values below 500 and above 2000) could probably result in better mTry, yielding higher classification performance for the final cropland product.

Appendix A.2. Accuracy at Grid Level
Results on the accuracy assessment do not show all the 189 trained and classified squares (100 km by 100 km grid unit). Twenty-one (21) of them have been entirely classified as Non-crop. They are considered as NoData for the assessment. The general trend appearing on Figure A2 is that in average classified squares have an overall accuracy above 80%. For most squares, overall accuracy falls in the range of 75% to 100%. The country of Burkina Faso has the maximum of units with overall accuracy within 50%-75%, and none in the grid has been classified with a correct proportion of less than 50%. This relatively high overall accuracy is contrasted by the crop class specific accuracy. Figure A3 gives insight into the "Crop" class user's accuracy at grid level. In total, 11% of assessed cells have a user's accuracy less than 50%, 58% of them have theirs between 50% and 75%, and the remaining 31% have a user's accuracy above 75% (Table 3).
Remote Sens. 2020, 12, x FOR PEER REVIEW 18 of 24 Figure A1. Occurrences of the number of tree (nTree) corresponding to minimum out-of-bag errors (A) and the obtained number of variables available for splitting at each tree node, mTry (B).
The Random Forest (RF) is widely accepted as an efficient ensemble approach for land cover classification using remotely sensed data. It handles well imbalanced data, missing values, and outliers [57]. However, tuning RF two major parameters (number of trees: nTree, number of variables available for splitting at each tree node: mTry) to get optimum values may be time and resource consuming, even in parallel processing environments like Google Earth Engine. In this study, we selected nTree in {500, 1000, 1500, 2000} for reduced computational time while ensuring sufficient trees for model convergence (Breiman, 2001). The best mTry for most of the grid cells has been achieved with nTree = 500, others for nTree = 2000, which are the limits of tuned nTree values ( Figure  A1,A). Using a larger range of nTree values (e.g., including values below 500 and above 2000) could probably result in better mTry, yielding higher classification performance for the final cropland product.

A2. Accuracy at Grid Level
Results on the accuracy assessment do not show all the 189 trained and classified squares (100 km by 100 km grid unit). Twenty-one (21) of them have been entirely classified as Non-crop. They are considered as NoData for the assessment. The general trend appearing on Figure A2 is that in average classified squares have an overall accuracy above 80%. For most squares, overall accuracy falls in the range of 75% to 100%. The country of Burkina Faso has the maximum of units with overall accuracy within 50%-75%, and none in the grid has been classified with a correct proportion of less than 50%. This relatively high overall accuracy is contrasted by the crop class specific accuracy. Figure  A3 gives insight into the "Crop" class user's accuracy at grid level. In total, 11% of assessed cells have a user's accuracy less than 50%, 58% of them have theirs between 50% and 75%, and the remaining 31% have a user's accuracy above 75% (Table 3). Figure A1. Occurrences of the number of tree (nTree) corresponding to minimum out-of-bag errors (A) and the obtained number of variables available for splitting at each tree node, mTry (B). Figure A1. Occurrences of the number of tree (nTree) corresponding to minimum out-of-bag errors (A) and the obtained number of variables available for splitting at each tree node, mTry (B).
The Random Forest (RF) is widely accepted as an efficient ensemble approach for land cover classification using remotely sensed data. It handles well imbalanced data, missing values, and outliers [57]. However, tuning RF two major parameters (number of trees: nTree, number of variables available for splitting at each tree node: mTry) to get optimum values may be time and resource consuming, even in parallel processing environments like Google Earth Engine. In this study, we selected nTree in {500, 1000, 1500, 2000} for reduced computational time while ensuring sufficient trees for model convergence (Breiman, 2001). The best mTry for most of the grid cells has been achieved with nTree = 500, others for nTree = 2000, which are the limits of tuned nTree values ( Figure  A1,A). Using a larger range of nTree values (e.g., including values below 500 and above 2000) could probably result in better mTry, yielding higher classification performance for the final cropland product.

A2. Accuracy at Grid Level
Results on the accuracy assessment do not show all the 189 trained and classified squares (100 km by 100 km grid unit). Twenty-one (21) of them have been entirely classified as Non-crop. They are considered as NoData for the assessment. The general trend appearing on Figure A2 is that in average classified squares have an overall accuracy above 80%. For most squares, overall accuracy falls in the range of 75% to 100%. The country of Burkina Faso has the maximum of units with overall accuracy within 50%-75%, and none in the grid has been classified with a correct proportion of less than 50%. This relatively high overall accuracy is contrasted by the crop class specific accuracy. Figure  A3 gives insight into the "Crop" class user's accuracy at grid level. In total, 11% of assessed cells have a user's accuracy less than 50%, 58% of them have theirs between 50% and 75%, and the remaining 31% have a user's accuracy above 75% (Table 3).

A3. Correlated Variables / Predictors
An example of removal of highly correlated variables, based on a correlation coefficient > 0.99 is shown in Figure A4. In these cases, we anticipate no additional useful explanatory information is available by including both variables. Predictor EVI is similar to SAVI and MSAVI, thus only MSAVI was maintained to develop the model by the algorithm. Similarly, predictors EVI_1 and SAVI_1 bring the same information as MSAVI_1; they can then be removed, reducing computational time in classifying this grid cell. Since RF is generally robust to correlation, correlations < 0.99 were permitted.

Appendix A.3. Correlated Variables/Predictors
An example of removal of highly correlated variables, based on a correlation coefficient > 0.99 is shown in Figure A4. In these cases, we anticipate no additional useful explanatory information is available by including both variables. Predictor EVI is similar to SAVI and MSAVI, thus only MSAVI was maintained to develop the model by the algorithm. Similarly, predictors EVI_1 and SAVI_1 bring the same information as MSAVI_1; they can then be removed, reducing computational time in classifying this grid cell. Since RF is generally robust to correlation, correlations < 0.99 were permitted. . Figure A4. Correlated variables.