In recent years, biofuels synthesized from crops have provided an opportunity to reduce the dependence on fossil fuels. Biodiesel, for example, is produced using organic fats and vegetable oils, such as palm oil and can be used with petrol diesel. Biofuels may help strengthen the energy security in countries that do not have direct availability of fossil fuel deposits while reducing of greenhouse gas (GHG) emissions (Sorda et al., 2010). However, the production of biofuel crops can also have a negative impact on the environment, such as deforestation (Fargione et al., 2008). Moreover, biofuels also put more stress on water and land resources that could otherwise be used for the production of food (Cai et al., 2010). Therefore, the competing needs for land and water resources by food and biofuel production has been a very important issue of the food-water-energy debate (Tilman et al., 2009; Lambin and Meyfroidt, 2011).

A prime example of a very strong interplay between food, water and energy is the unsustainable palm oil production in Indonesia and other countries for biofuels and human consumption (Mukherjee and Sovacool, 2014). Global palm oil production is dominated by Indonesia and Malaysia, which together account for over 90% of total global palm oil (WorldAtlas, 2018). Indonesia is currently the largest producer and exporter of palm oil worldwide. Furthermore, due to the increase global population, demand for palm oil globally shows an increasing trend. This rapid growth of the palm oil industry has happened at the expense of severe damage to tropical forests which play a very critical role in the carbon cycle of the earth. Tropical forests cover only 7–10% of the Earth's surface (Malhi and Grace, 2000; Nightingale et al., 2004), but they are globally important, containing 40–50% of all carbon stored in terrestrial vegetation. The role of these forests in the global carbon cycle is important because it is estimated that tropical deforestation is responsible for 20% of global anthropogenic carbon emissions (Parry et al., 2007). Tropical rainforests on Southeast Asia are unique as they have exceptionally high biodiversity and large extent of tropical peatlands. Tropical peatlands are very rich sinks of carbon containing globally a peat carbon pool of 88.6 Gt (equal to 15–19% of the global peat carbon pool), of which 57.4 Gt carbon is in Indonesian peatlands alone (Page et al., 2011). Forest clearance, peatland drainage and fires (both natural and caused by humans) lead to significant emissions of CO₂ to the atmosphere. For example, 1997–98 Indonesian peatland fire released more than 0.87 Gt of carbon, equivalent to 14% of the average annual global fossil fuel emissions released during the 1990s (Page et al., 2002).

Apart from biodiversity loss and high carbon emissions, the large scale at which palm oil is cultivated can also lead to poor soil and water quality and availability problems. For example, palm oil mill effluent (POME) has been disposed of as untreated waste into natural water sources and has caused severe degradation of water quality. Furthermore, the general oil extraction process is water intensive, as it demands large quantities of water (Sheil et al., 2009; Rupani et al., 2010).

The unsustainable nature of plantation industry in Southeast Asia and other parts of the world has been acknowledged globally. Various companies and governments are trying to ensure that the palm oil plantations meet rigorous sustainability standards (Scarlat and Dallemand, 2011). For example, under a new European Union biofuel policy, any palm oil biodiesel imported to the region must, over its full life cycle, demonstrate a 35% savings in greenhouse gas emissions compared to fossil fuel diesel, and the feedstock cannot be grown in areas with high biodiversity value or a high stock of carbon. Major corporate groups in Indonesia and other Southeast Asian countries have also started moving toward sustainable production of palm oil through certification under the Roundtable on Sustainable Palm Oil (RSPO) (Schouten and Glasbergen, 2011) and other certification programs, such as International Sustainability and Carbon Certification (ISCC) (Moser et al., 2014) and ISPO certification (Paoli et al., 2013). However, evaluating the effectiveness of these diverse policies depends on the ability to monitor land cover change due to plantation expansion (Carlson et al., 2018). Hence, scalable and timely monitoring of these land uses is essential for understanding whether programs and policies are meeting their stated goals (WCS, 2010; Wakker and Asia, 2014).

Classification of remotely sensed images into different land cover classes has been a widely used approach in various earth science applications. Remote sensing data from various earth observation satellites, such as the Advanced Very High Resolution Radiometer (NOAA AVHRR), Satellite Pour l'Observation de la Terre (SPOT) VEGETATION, Moderate Resolution Imaging Spectroradiometer (MODIS), and LANDSAT, have been used to produce forest and land cover maps for large-scale land cover monitoring (Achard and Estreguil, 1995; Mayaux et al., 1998; Hansen et al., 2000).

In particular, remote sensing data acquired through various earth observation satellites provide immense opportunity to monitor land use/land cover (LULC) changes caused by plantation cultivation. However, current state of the art methods based on remote sensing data are limited in their temporal frequency and scalability due to various reasons, such as need for human intervention, use of very simple machine learning methods and are applicable only for small regions (Hansen et al., 2008, 2013; Hoscilo et al., 2011; Dong et al., 2012; Li and Fox, 2012; Margono et al., 2012; Miettinen et al., 2012a,b; Ziegler et al., 2012; Gutiérrez-Vélez and DeFries, 2013). To this date, there is no existing framework that can provide plantation extent maps in an automated fashion at yearly scales for large regions.

Even though yearly plantation maps are not available, a few organizations have developed plantation maps for a single or a few years. Two such datasets are Tree Plantation (TP) dataset (Petersen et al., 2016) and RSPO dataset (Gunarso et al., 2013). These datasets have reasonable accuracy as they have been prepared using visual interpretation by human experts. The TP dataset provides the location of tree plantations in selected tropical countries circa 2013–2014. According to the visual inspection conducted by Petersen et al. (2016), this dataset has very good recall (i.e., producer's accuracy, which is the fraction of true plantations that have been detected over the total amount of true plantations) but has poor precision (i.e., user's accuracy, which is the fraction of true plantations among the detected plantations). RSPO dataset is available for three different years namely 2001, 2005, and 2009. This product provides a very detailed map with 19 classes for each of these years. This dataset has higher precision but poor recall. Even though these datasets are imperfect, they can serve as different sources of noisy labels that can be used for training machine learning models.

Although we have a few manually created plantation maps as sources of labels, the automated detection of land cover changes to/from plantations using remote sensing datasets is still a challenging task due to various reasons:

High multi-modality: There exists a wide variety of land cover types on the earth's surface. Different taxonomies have been defined to categorize locations on the earth's surface at different level of granularity. For example, Table 1 shows three different taxonomies. Similarly, more detailed or coarse taxonomies can be defined depending on the application. Ideally, we wish to use the most detailed taxonomy to categorize the locations to obtain the maximum information about land cover change. But in practice, learning classification models that can distinguish all these classes is difficult. There exists no high accuracy map that can provide sufficient high quality training samples for all these classes. In this work, we aim to simplify the problem by using a coarse taxonomy. Specifically, we aggregate all land cover types into three classes namely, Forests, Plantations and all the remaining land cover types are labeled as Other land cover class. Forest class has been kept separate from the group of rest of the classes because we are also interested in estimating how many of the plantations were established by removing forests. Now each of these three classes have subclasses within them which makes these classes highly multi-modal in nature. Hence, we need training samples from all these modes in order to achieve better classification accuracy.

Noisy ground truth: In traditional classification settings, it is assumed that high quality ground truth labels are available for training the classification model. However, in this scenario high quality ground truth is not available. Instead, noisy labels from different sources are available. Hence, traditional classification techniques might have limited performance in this scenario.

High dimensionality: Most land cover classes have annual growth cycle and hence show a seasonal pattern. Classification cannot be done by using just a single time step as the separability between classes vary across the year. For example, a crop field just after harvest would look very similar to a barren land and hence would not be distinguishable. In order to achieve better classification performance, we need to incorporate both spectral properties at individual timesteps and the temporal pattern of growth of different land cover types.

Spatial and temporal heterogeneity: Due to atmospheric disturbances and natural variability in land cover types, classification models learned in a specific region for a specific time may not perform well when applied to other regions and time. Hence, there is a need to incorporate variations in feature space for better performance.

In this paper, we study the effectiveness of a machine learning framework to map plantation at annual scale. This framework learns a multi-class ensemble from noisy ground truth data obtained by manual labeling (i.e., TP and RSPO). The key contribution of this work is to study the effectiveness of several components in this framework by comparing the performance for a set of variants of this framework.

We compare the quality of annual plantation extents generated by the proposed framework with the existing datasets in the Kalimanthan region of Indonesia that were used for training our algorithms. Specifically, we have analyzed the portion of Kalimanthan that overlaps with MODIS tile h29v09. Through visual inspection of high resolution imagery and manually labeled set of points, we show that the proposed framework can overcome the imperfectness of the available products and thus has the potential to produce high-quality large-scale plantation maps with little manual effort.

A wide variety of methods have been proposed that use remote sensing at different spatial and temporal scale for monitoring changes in land cover. However, a vast majority of these methods focus on detection deforestation activity only (Hansen et al., 2008; Hoscilo et al., 2011; Margono et al., 2012). The widely used global deforestation product (Hansen et al., 2013) does not differentiate between forest and plantations. Similarly, land cover product from NASA also does not model the plantation class separately.

Methods focused on detection conversions of land cover types to/from plantations have several issues that make them unsuitable for monitoring plantation related activities at large scale. Some methods involve extensive human involvement as they use visual interpretation in the detection process (Miettinen et al., 2012a,b; Ziegler et al., 2012). Few automatic machine learning based methods have also been proposed in the literature but they use very simple techniques such thresholding (Dong et al., 2012; Gutiérrez-Vélez and DeFries, 2013), nearest neighbor method (Li and Fox, 2012). Some sophisticated machine learning methods have only shown success in selected small-scale test dataset (Jia et al., 2017a,b). Due to these reasons current state of the art methods have limited applicability.

In this section, we will first describe the proposed ensemble learning framework. We then introduce several baselines methods which are variants of our proposed method. By comparing against these baselines, we are able to demonstrate the effectiveness of each component in the proposed framework.

In this paper, we study several key components in a machine learning framework for automatically creating plantation maps. These components includes class aggregation, data sampling, learning model selection, and post-processing. The evaluation of multiple baselines derived from this framework confirms the effectiveness of each component. The visual verification of the proposed framework on a large region in Indonesia (MODIS tile h29v09) shows that the proposed method can generate high-quality annual plantation maps and our detection achieves a better balance of precision and recall than those datasets that were used for training our proposed framework.

The methods we presented here can be used to create plantation mapping products. In future, we will make use of the generated plantation maps to understand how plantation conversion impacts the environment and better monitor the policy compliance.

For example, the analysis using a combination of generated plantation maps and auxiliary datasets, such as the Plantation Concessions Dataset¹ has an implication on potential illegal plantation areas. We also plan to leverage the fire product (Mithal et al., 2018) to detect uncontrolled fires from plantation conversion and study their impact to deforestation. In addition, we will analyze the correlations between plantation dynamics and the carbon emission² to study how cultivating plantations leads to large amount of carbon emissions.

Our proposed method also remains limited in terms of validation and imagery inputs, which need to be addressed in future work. The first limitation lies in that the reference data used in the validation process (e.g., TP, RSPO, and DigitalGlobe) are mostly created through manual inspection. However, some plantations may not be easily identified visually due to their advanced age and associated high tree cover. Second, our analysis was also limited by the resolution of the MODIS data. While the high resolution of Landsat data (30 m) and Sentinel data (10 m) offer potential to map plantations more accurately, the low temporal frequency of Landsat (16 days) and Sentinel (5/10 days) makes it hard to find images with little noise (e.g., clouds). A joint multi-scale learning framework has potential to better delineate the boundary of target classes with a higher spatial resolution while also taking advantage of rich temporal knowledge from more frequently collected satellite data.

The datasets for this manuscript are not publicly available because the RSPO dataset is currently private and is still being updated. Requests to access the datasets should be directed to jiaxx221@umn.edu.

XJ was the primary author who conducted most of the experiments and analysis. AK conducted some parts of the experiments. KC, JG, and PW were our domain experts who greatly helped in analyzing the results and also helped with the text. VK was the one who managed the entire team and also helped to review the work.