Retrieving the National Main Commodity Maps in Indonesia Based on High-Resolution Remotely Sensed Data Using Cloud Computing Platform

: Indonesia has the most favorable climates for agriculture because of its location in the tropical climatic zones. The country has several commodities to support economics growth that are driven by key export commodities—e.g., oil palm, rubber, paddy, cacao, and co ﬀ ee. Thus, identifying the main commodities in Indonesia using spatially-explicit tools is essential to understand the precise productivity derived from the agricultural sectors. Many previous studies have used predictions developed using binary maps of general crop cover. Here, we present national commodity maps for Indonesia based on remote sensing data using Google Earth Engine. We evaluated a machine learning algorithm—i.e., Random Forest to parameterize how the area in commodity varied in Indonesia. We used various predictors to estimate the productivity of various commodities based on multispectral satellite imageries (36 predictors) at 30-meters spatial resolution. The national commodity map has a relatively high accuracy, with an overall accuracy of about 95% and Kappa coe ﬃ cient of about 0.90. The results suggest that the oil palm plantation was the highest commodity product that occupied the largest land of Indonesia. However, this study also showed that the land area in rubber, rice paddies, and cacao commodities was underestimated due to its lack of training samples. Improvement in training data collection for each commodity should be done to increase the accuracy of the commodity maps. The commodity data can be viewed online (website can be found in the end of conclusions). This data can further provide signiﬁcant information related to the agricultural sectors to investigate food provisioning, particularly in Indonesia.


Introduction
Agricultural industries are an integral part of the Earth's ecology and play a significant role in human livelihoods [1,2]. Indonesia has the most favorable climates in the world for agriculture [2]. Indonesia's wealth of natural resources and a range of goods is a significant and indispensable Land 2020, 9,377 3 of 15 imageries have been used to distinguish these crops [23]. Several recent studies have developed advanced classification algorithms and imagery data from high-resolution passive sensors to enhance the classification of small-landholder coffee areas [24][25][26]. Utilization of machine learning algorithms to identify shade-grown coffee areas has been evaluated in Nicaragua [27]. However, it has long proven challenging to detect shade-grown coffee precisely using remotely sensed data. Many studies across the tropics have conducted mapping of oil palm land cover using remote satellite sensing data [6,[28][29][30]. The methods to distinguish oil palm cover from other vegetative community types can generally be divided into two categories: image and phenology-based approaches [28]. Shade-grown coffee also retrieved using Landsat 8 imageries, topographical data from Shuttle Radar Topography Mission (SRTM), and precipitation data (CHIRPS) through Random Forest algorithm [27]. Therefore, multispectral imageries associated with the digital elevation model and physiographic or geophysical data can be used to retrieve commodity covers. Here, we present the commodity crops (i.e., oil palm, rubber, rice, cacao, and coffee) in one classification rule-set based on very-high satellite imageries (30-m) using the cloud computing platform.
In this decade, the cloud computing platform has been widely used to retrieve land cover maps in the planetary scale-i.e., Google Earth Engine (GEE). Google Earth Engine is a cloud-based geo-spatial analysis platform that delivers massive computing capabilities to address a variety of high-impact societal issues including deforestation, drought, disaster, disease, food security, water management, climate monitoring, and environmental protection [31]. In this study, we used the Google Earth Engine Platform to generate a good quality of satellite imageries and perform the machine learning classification to predict the commodity maps across Indonesia.

Study Area
This study was carried out in Indonesia ( Figure 1) -a tropical country with the potential to support agricultural and plantation production of commodity crops [32]. Indonesia is the fourth most populous country with the largest Muslim population in the world with a total of 267,026,366 people [33]. High demand of food and economics due to population density is connected to agricultural land [32].
Land 2020, 9, x FOR PEER REVIEW 3 of 16 imageries have been used to distinguish these crops [23]. Several recent studies have developed advanced classification algorithms and imagery data from high-resolution passive sensors to enhance the classification of small-landholder coffee areas [24][25][26]. Utilization of machine learning algorithms to identify shade-grown coffee areas has been evaluated in Nicaragua [27]. However, it has long proven challenging to detect shade-grown coffee precisely using remotely sensed data. Many studies across the tropics have conducted mapping of oil palm land cover using remote satellite sensing data [6,[28][29][30]. The methods to distinguish oil palm cover from other vegetative community types can generally be divided into two categories: image and phenology-based approaches [28]. Shade-grown coffee also retrieved using Landsat 8 imageries, topographical data from Shuttle Radar Topography Mission (SRTM), and precipitation data (CHIRPS) through Random Forest algorithm [27]. Therefore, multispectral imageries associated with the digital elevation model and physiographic or geophysical data can be used to retrieve commodity covers. Here, we present the commodity crops (i.e., oil palm, rubber, rice, cacao, and coffee) in one classification rule-set based on very-high satellite imageries (30m) using the cloud computing platform. In this decade, the cloud computing platform has been widely used to retrieve land cover maps in the planetary scale-i.e., Google Earth Engine (GEE). Google Earth Engine is a cloud-based geospatial analysis platform that delivers massive computing capabilities to address a variety of highimpact societal issues including deforestation, drought, disaster, disease, food security, water management, climate monitoring, and environmental protection [31]. In this study, we used the Google Earth Engine Platform to generate a good quality of satellite imageries and perform the machine learning classification to predict the commodity maps across Indonesia.

Study Area
This study was carried out in Indonesia ( Figure 1) -a tropical country with the potential to support agricultural and plantation production of commodity crops [32]. Indonesia is the fourth most populous country with the largest Muslim population in the world with a total of 267,026,366 people [33]. High demand of food and economics due to population density is connected to agricultural land [32].

Training Data
We used a variety of databases to retrieve commodity identification training for the model input. A database of five main commodities-i.e., oil palm, rubber, coffee, cacao, and paddy was compiled from the following sources: (1) Biodiversity records from the Global Biodiversity Information Facility database (s; www.gbif.org); keywords "Hevea brasiliensis", "Elaeis guineensis", "Coffea arabica", "Coffea canephora", "Theobroma cacao", and "Oryza sativa"; (2) the database for rice paddy cover in Indonesia created by The Ministry of Agriculture; (3) Industrial oil palm extent of Indonesia; digitized from Austin et al., (2017); (4) Field survey in four pilot locations-i.e., Aceh Province, South Tapanuli Regency, Pelalawan Regency, and Sintang Regency were also used for training data. Point acquisition to obtain commodity information by survey was conducted in 2018-2019 during the dry season; we also conducted interviews with local people to identify dominant land cover types in the areas. Geographic coordinates were recorded in decimal degrees, based on the WGS 1984 datum. We also used 7621 points from other land cover types data (22 classes), retrieved from the Ministry of Environment and Forestry of Indonesia.
Data filtering was essential for further analyses due to the records from the global database, GBIF in particular may contain potential errors or duplicated data. Records from all of the databases were integrated or merged into a training dataset. Records of commodity with the dataset without any hard facts, including relevant or detailed descriptions, were excluded from analysis. For this study, this resulted in 236 data points for coffee (both arabica and robusta), 29 data points for cacao, 52 for rubber, 294 rice paddies, 504 points of oil palm, and 7621 points of other land cover types. Specific information related to the other land cover types for the training data consists of: 1683 data points for dryland primary forest, 1651 data points for dryland secondary forest, 66 for primary mangrove forest, 231 for primary peat-swamp forest, 192 data points for plantation forest, 590 points of shrub, 215 points of plantation, 94 points of settlement, 186 points of bare land, 111 points of savanna, 149 points of waterbody, 65 points of secondary mangrove forest, 262 points of secondary peat-swamp forest, 309 points of mixture of shrub and swamp, 398 points of dryland agriculture, 1088 points of mixture dryland agriculture, 188 points of wetland cultivation, 37 points of cultivated pond, 10 points of built-up area, 9 points of transmigration area, 27 points of mining, and 60 points of swamp.

Features Derives
This study used a total of 36 features from satellite imageries-i.e., Landsat 8 imageries, Shuttle Radar Topographic Mission (SRTM) data, and Joint Research Center (JRC) global surface water data at 30 m × 30 m spatial resolution to identify the commodities. Table 1 shows the details of the dataset used in this study.
We used reflectance of Landsat 8 top of atmosphere (TOA) with Tier 1 level data where the data was already geo-registered with root mean square error (RMSE) less than 12 m resolution which represented the best quality imagery available for the collected data [34]. Cloud-masking using Quality Assessment band (BQA) and filter statistics by pixels (i.e., median) over the period of Landsat 8 TOA reflectance imageries were performed before further analysis within The Google Earth Engine. The research was also carried out using moving average lag time of 1-year analysis of images collection to determine the recent period of Landsat 8 due to cloud cover issues in the humid tropical region [35]. Furthermore, we used images from January 1st 2017 to December 31st 2019 to represent the 2018 data with clear cloud cover. EVI represents an optimized vegetation index which has the ability to improve vegetation monitoring by reducing atmospheric disturbance and canopy background signal. EVI is calculated using the following equation [36]: where NIR or RED or BLUE are spectral reflectance (atmospherically corrected) of near-infrared, red, and blue wavelengths respectively, C1 and C2 are the coefficient of the aerosol resistance with the values of 6 and 7.5, G represents gain factor (2.5), and L refers to 1 [36]. SAVI represents a vegetation index with reduced effects of soil brightness in areas where vegetation density is sparse or low [37]. This index is calculated as the following equation: where NIR and RED are spectral reflectance (atmospherically corrected) of near-infrared and red wavelengths and L refers to soil brightness correction factor (0.5).
Interspecific covariate has been calculated to represents more statistics variation in the machine learning algorithm for commodity detection. Covariate between two-bands can be calculated using the following equation: where BAND i or BAND j represents spectral reflectance of Landsat 8 (atmospherically corrected) for blue, green, red, near-infrared, shortwave-infrared 1, and shortwave-infrared 2. Thus, this variable is the normalized difference of each multispectral band in Landsat 8.

Commodity Cover Prediction
To identify the national commodity cover in Indonesia-i.e., oil palm, rubber, coffee, cacao, and rice paddies, we used machine learning classification through the Random Forest (RF) algorithm within the multiclass classification (total of 27 classes) in the Google Earth Engine Platform. GEE has 10 classifiers with different algorithm to estimate a pixel-based classification-i.e., CART, RF, Minimum Distance, GMO MaxEnt, Naive Bayes, SVM, Perceptron, IKPamir, and Winnow [29]. Random Forest gives the best performance from other classifiers [41]; therefore, we choose this algorithm to predict commodity classes. A Random Forest classifier provides an ensemble model that effectively distinguishes spectrally similar agriculture land and forest cover by generating multiple trees from training data and its predictors [27,31,42]. Many studies have investigated the performance of the RF algorithm to identifying land cover from hyperspectral, multispectral imageries, and digital elevation model data as well [43][44][45].
Random Forest classifiers with high variance and low bias perform by growing the forest to a user-defined number-of-trees [42]. This study evaluated various RF models using different parameterization of the number-of-trees (N); i.e., N = 25, N = 50, N = 100, and N = 500 to obtain the best model from the various parameterization of RF classifiers. We use 36 number of features derived from Landsat 8 imageries and the digital elevation model as well. The number of variables per split has been defined as root square of the number of features; i.e., 6. We describe each of these methods in the following flowchart ( Figure 2).

Model Evaluation
Accuracy assessment for the overall model was measured by the error matrices which were used to estimate the user's accuracy (the accuracy of classification despite commission error) and the producer's accuracy (the accuracy of classification despite omission error). Then, we provided a Kappa coefficient to examine the goodness of the predictions relative to validated reference data [46]. However, the Kappa coefficient has come under recent question as a useful metric for accuracy assessments [47]. Quantity disagreement is the amount of difference between the reference categories and the classified categories that is due to the less than perfect match in the proportions of the categories. Allocation Land 2020, 9, 377 7 of 15 disagreement is defined as the amount of difference between the reference categories and the classified categories that is due to the less than maximum match in the spatial allocation of the categories, given the proportions of the categories in the reference and comparison maps. The total disagreement is the sum of quantity disagreement and allocation disagreement [47]. We also compare the statistics of total areas for each commodity with reference data from the Statistics Bureau of Indonesia by province [8]. We perform zonal statistics for each commodity by provincial administration to calculate total of commodity area.
The variable importance can be calculated generally using mean decrease in accuracy (MDA) or mean decrease in Gini (MDG) [48]. The mean decrease in Gini coefficient estimates each variable contribution to the homogeneity of the nodes [14]; besides MDA coefficient is the average change in accuracy across the forest [49]. Many studies used the MDG to calculate the variable importance [14,49,50]. The MDG provides the good robustness while the MDA tend to unstable to small perturbations of the dataset [49,51]. We estimated the relative importance of each predictor to the model using the percent contribution and variable importance for the best model using mean decrease in Gini [42]. We also investigated the response curves of the 4th most important variables to explore how the predictors affected the commodity prediction.

Model Evaluation
Accuracy assessment for the overall model was measured by the error matrices which were used to estimate the user's accuracy (the accuracy of classification despite commission error) and the producer's accuracy (the accuracy of classification despite omission error). Then, we provided a Kappa coefficient to examine the goodness of the predictions relative to validated reference data [46]. However, the Kappa coefficient has come under recent question as a useful metric for accuracy assessments [47]. Quantity disagreement is the amount of difference between the reference categories and the classified categories that is due to the less than perfect match in the proportions of the categories. Allocation disagreement is defined as the amount of difference between the reference categories and the classified categories that is due to the less than maximum match in the spatial allocation of the categories, given the proportions of the categories in the reference and comparison

Accuracy Assessment
Generally, the RF algorithm with 100 number of trees gives the highest accuracy based on overall accuracy and Kappa coefficient (OA = 95.2% and κ = 0.90; Table 2). The best RF model (N = 100) has relatively strong agreement based on κ and OA [52]. Cacao has the lowest value of producer's accuracy (PA) ranging from 39.6% to 52.2%. Otherwise, oil palm has the highest PA ranging from 78.9% to 93.2%. Besides, oil palm also has the lowest user's accuracy (UA) values ranging from 58.4% to 85.8%. Coffee has the highest value of UA ranging from 83.3% to 97.9% (Table 3). Table 3. Producer's (%) and User's (%) accuracy of classified maps for commodity cover using different Number of Trees. PA refers to Producer's accuracy and UA refers to User's accuracy.  Nevertheless, the standard Kappa coefficient had some considerable conceptual problems [53]-thus we also calculate quantity and allocation disagreement [47]. Figure 3 summarizes the omission disagreement, agreement, and commission disagreement by category for each Random Forest 2019 commodity maps. Vertical axis shows the categories while the horizontal axis shows the number of validation observations. The Random Forest algorithm with number of trees 100 ( Figure 3C) has the least disagreement both commission and omission. Figure 3 shows that the omission disagreement is greater than the commission disagreement for the most of commodity (i.e., rubber, oil palm, coffee, cacao, and also paddy in Figure 3B,C), therefore those commodities were underestimated by all methods. In contrast, paddy in Figure 3A was overestimated.

Spectral Characteristics
We can see a relatively contrast characteristic of commodity from Near Infra-Red Band (NIR; B5) and Short-Wave Infra-Red Band (SWIR; B6). The result shows that coffee, cacao, and rubber have a similar spectral characteristic from Landsat 8 multi-spectral bands (in moderate values of near infra-red range). Moreover, oil palm has a highest spectral value in near infra-red range. Otherwise, paddy has the lowest spectral value of B5 (Figure 4).

Spectral Characteristics
We can see a relatively contrast characteristic of commodity from Near Infra-Red Band (NIR; B5) and Short-Wave Infra-Red Band (SWIR; B6). The result shows that coffee, cacao, and rubber have a similar spectral characteristic from Landsat 8 multi-spectral bands (in moderate values of near infrared range). Moreover, oil palm has a highest spectral value in near infra-red range. Otherwise, paddy has the lowest spectral value of B5 (Figure 4). The results show that elevation, land surface temperature (B11), covariate of B6 and B7 (ND_B6_B7), and EVI were the four most important variables by mean decrease in Gini coefficient values (325, 75, 52, and 38, respectively). EVI was the most importance variable in commodity maps

Spectral Characteristics
We can see a relatively contrast characteristic of commodity from Near Infra-Red Band (NIR; B5) and Short-Wave Infra-Red Band (SWIR; B6). The result shows that coffee, cacao, and rubber have a similar spectral characteristic from Landsat 8 multi-spectral bands (in moderate values of near infrared range). Moreover, oil palm has a highest spectral value in near infra-red range. Otherwise, paddy has the lowest spectral value of B5 (Figure 4). The results show that elevation, land surface temperature (B11), covariate of B6 and B7 (ND_B6_B7), and EVI were the four most important variables by mean decrease in Gini coefficient values (325, 75, 52, and 38, respectively). EVI was the most importance variable in commodity maps The results show that elevation, land surface temperature (B11), covariate of B6 and B7 (ND_B6_B7), and EVI were the four most important variables by mean decrease in Gini coefficient values (325, 75, 52, and 38, respectively). EVI was the most importance variable in commodity maps of Indonesia-therefore, we analyze spectral characteristic from EVI for various commodities ( Figure 5).
Coffee is found at higher altitude areas (1151 ± 389) m asl. Besides, cacao, oil palm, paddy, and rubber were found in relatively lower elevation (117 ± 161) m asl than coffee. In the land surface temperature perspectives, paddy was suitable to grow in high temperature areas (291.50 ± 1.27) Kelvin. Otherwise, coffee plantation has the lowest temperature (289.20 ± 1.75) Kelvin than the other commodities. The average of EVI value for oil palm, rubber, cacao, coffee, and paddy were 0.85, 0.81, 0.76, 0.70, and 0.59, respectively.

Spatial Distributions of Commodity Maps
The results show that generally, 6.5% (25.39 million ha) of the total geographic area was net commodity of 2019) with the total area of coffee, cacao, rubber, paddy, and oil palm were 1.96 million ha, 96.98 thousand ha, 279.59 thousand ha, 8.15 million ha, and 14.89 million ha, respectively. The commodities were rapidly growing in Sumatera, Kalimantan, and Java regions ( Figure 6). Elevation predictor shows the higher discrimination between the commodities ( Figure S1). Coffee is found at higher altitude areas (1151 ± 389) m asl. Besides, cacao, oil palm, paddy, and rubber were found in relatively lower elevation (117 ± 161) m asl than coffee. In the land surface temperature perspectives, paddy was suitable to grow in high temperature areas (291.50 ± 1.27) Kelvin. Otherwise, coffee plantation has the lowest temperature (289.20 ± 1.75) Kelvin than the other commodities. The average of EVI value for oil palm, rubber, cacao, coffee, and paddy were 0.85, 0.81, 0.76, 0.70, and 0.59, respectively.

Spatial Distributions of Commodity Maps
The results show that generally, 6.5% (25.39 million ha) of the total geographic area was net commodity of 2019) with the total area of coffee, cacao, rubber, paddy, and oil palm were 1.96 million ha, 96.98 thousand ha, 279.59 thousand ha, 8.15 million ha, and 14.89 million ha, respectively. The commodities were rapidly growing in Sumatera, Kalimantan, and Java regions ( Figure 6). Oil palm plantation was the most dominant commodity than the others, particularly in Sumatera and Kalimantan. Meanwhile, Java Island was one of the most suitable regions for paddy growth. North Sumatera Province has the highest total areas of coffee, rubber, and cacao plantations covering 748.43 thousand ha, 61.06 thousand ha, and 32.17 thousand ha, respectively. East Java Province has the highest total areas of paddy field (1.21 million ha). Riau Province has the highest total areas of oil palm plantation (2.51 million ha). However, research on the comparisons of coffee, rubber, and cacao with the spatial reference data had not been carried out due to the lack of data (Figure 7). Official data of oil palm plantation and paddy field were retrieved from Austin's et al. [6] dataset and the Ministry of Agriculture (MoA), respectively. However, oil palm extent from Austin's et al. [6] only represents the large-scale oil palm plantations in Indonesia's major producing regions of Sumatra, Oil palm plantation was the most dominant commodity than the others, particularly in Sumatera and Kalimantan. Meanwhile, Java Island was one of the most suitable regions for paddy growth. North Sumatera Province has the highest total areas of coffee, rubber, and cacao plantations covering 748.43 thousand ha, 61.06 thousand ha, and 32.17 thousand ha, respectively. East Java Province has the highest total areas of paddy field (1.21 million ha). Riau Province has the highest total areas of oil palm plantation (2.51 million ha). However, research on the comparisons of coffee, rubber, and cacao with the spatial reference data had not been carried out due to the lack of data (Figure 7). Oil palm plantation was the most dominant commodity than the others, particularly in Sumatera and Kalimantan. Meanwhile, Java Island was one of the most suitable regions for paddy growth. North Sumatera Province has the highest total areas of coffee, rubber, and cacao plantations covering 748.43 thousand ha, 61.06 thousand ha, and 32.17 thousand ha, respectively. East Java Province has the highest total areas of paddy field (1.21 million ha). Riau Province has the highest total areas of oil palm plantation (2.51 million ha). However, research on the comparisons of coffee, rubber, and cacao with the spatial reference data had not been carried out due to the lack of data (Figure 7). Official data of oil palm plantation and paddy field were retrieved from Austin's et al. [6] dataset and the Ministry of Agriculture (MoA), respectively. However, oil palm extent from Austin's et al. [6] only represents the large-scale oil palm plantations in Indonesia's major producing regions of Sumatra, Official data of oil palm plantation and paddy field were retrieved from Austin's et al. [6] dataset and the Ministry of Agriculture (MoA), respectively. However, oil palm extent from Austin's et al. [6] only represents the large-scale oil palm plantations in Indonesia's major producing regions of Sumatra, Kalimantan, and Papua. Besides, the paddy field extent from the MoA was not covering small-holders paddy field as well.

Discussion
Google Earth Engine is a cloud computing platform with the ability to carry out the machine learning classifications and perform pre-processing big-data computing of multispectral satellite imageries [18,29]. This study shows that various types of commodity, such as coffee, cacao, oil palm, rubber, and paddy plantations can be mapped at a high spatial resolution of 30 m. Furthermore, the results retrieved with the Random Forest algorithm, which is one of the machine learning approaches, outperformed prior studies. This research realized that the model has relatively high accuracy (OA = 95% and κ = 0.90) for the Random Forest algorithm with N value of 100. Moreover, it also provides a good result to discriminate various types of the commodity using multispectral satellite imageries. The results of this study were also compared with the official reference data from Statistical Bureau of Indonesia [8] for the oil palm and paddy commodities and also from the previous studies used to identify the commodity [29,54]. The difference between the study's oil palm plantation map (14,896,964 ha) with the areas from BPS (14,677,561 ha) was approximately 1%; with the highest value found in Riau Province as the center of oil palm commodity province both for smallholder farmers and industrial plantation [55]. Furthermore, the difference between our paddy rice field map (8,152,435 ha) with paddy rice field areas from BPS (10,677,887 ha) was 24%. This result was still underestimated towards reference data from BPS. However, this result was still underestimated towards reference data from BPS, therefore the similar spatial characteristics between maps and other reference data shows East Java as the highest total area of paddy rice field in Indonesia as shown in Figure S2 [8,[56][57][58].
This study shows a lower omission error of the oil palm plantation than the prior research carried out by Lee et al. [29] (PA = 88% vs. PA = 93% in our estimation). However, the study of [29] has a slightly lower commission error of oil palm plantation than this study's estimation (UA = 88% vs. UA = 86% in our estimation). According to the [54], this study has higher performances, both producer and user accuracies (PA = 82% vs. PA = 90% in our estimation and UA = 65% vs. UA = 94% in our estimation). This study also shows the higher producer's and user's accuracies of coffee plantation than the previous research (PA = 86% compared to 82% in Kelley's et al. [27]. data and UA = 98% compared to 80% in Kelley's et al. [27] dataset). Nevertheless, this study's results for cacao plantation still has a poor result based on comparison from the previous study (PA = 51% compared to 96% in Numbisi's et al. [17] dataset). However, this research has a higher user's accuracy of cacao plantation (UA = 95% compared to 86% in Numbisi's et al. [17] dataset). The study also shows higher omission and commission errors of rubber plantation compared to Kou et al. [21]; PA = 95% vs. PA = 68% in our estimation and UA = 91% vs. UA = 87% in our estimation. The high omission error in cacao and rubber plantations suggests that there was still a lack of training data for those commodities.
To improve the accuracy of the commodity maps, further field works for collecting more training data of commodity needs to be carried out, particularly for cacao, rubber, and coffee. The previous study has suggested that a large number of representative samples will increase the performances of the model [59]. Compiling secondary data from the governments (e.g., the Ministry of Agriculture or the Ministry of Environment and Forestry) and defining the consensus of its data within our data (i.e., commodity maps) also should be carried out-therefore, a general agreement of the commodity data can be revealed.

Conclusions
The study successfully produced the high spatial resolution of 30 m for commodity maps in Indonesia region. The results show that the total geographic area of the net commodity in 2019 was 6.5% (25.39 million ha) with coffee, cacao, rubber, paddy, and oil palm covering 1.96 million ha, 96.98 thousand ha, 279.59 thousand ha, 8.15 million ha, and 14.89 million ha, respectively. The product has relatively high accuracy, with an overall accuracy of 95% and a Kappa coefficient of 0.90. However, several commodities were still further underestimated due to lack of training samples. Therefore, improvement in training data collection for each commodity needs to be carried out