Pache.co Information Technologies

Oil palm is widely cultivated owing to the increasing demand from various industries. Approximately 50% of supermarket products are estimated to contain palm oil ingredients (Guadalupe et al., 2019; Yue et al., 2020). According to the FAO (2020), vegetable oil production nearly doubled from 2000 to 2017, primarily due to increased palm oil production. Further, palm oil exceeded soybean oil production, ranking first in 2017 by contributing 35% of the total global vegetable oil. This recent domination will likely continue since oil palm can produce 6–10 times more vegetable oil per hectare than other oil crops (Murphy, 2015; Nomanbhay et al., 2017).

Indonesia contributes to almost 60% of global palm oil production, making it the largest palm oil producer on the planet (Bakhtary et al., 2021; USDA, 2021). In 2018, the palm oil trade created a surplus of USD 16.5 billion for crude palm oil (CPO) and USD 1.8 billion for palm kernel oil (PKO), making it the largest contributor to Indonesian foreign exchange (Dirjenbun, 2020). In addition, the revenue gain from the palm oil trade remained stable from 2015 to 2019 compared with other products (GAPKI, 2020), making it one of the most profitable and competitive Indonesian products.

However, the expansion of the oil palm industry is deemed responsible for deforestation, environmental degradation, social conflict, and community eviction. It has been estimated that 53.8% of oil palm expansion in Indonesia from 1989 to 2013 came from forested areas (Vijay et al., 2016). Similarly, Austin et al. (2019) found that 23% of forest loss in Indonesia between 2001 and 2016 was due to the expansion of oil palm plantations. Land clearing for oil palm plantations has triggered forest and peat fires (Cattau, Harrison, et al., 2016; Chisholm et al., 2016), releasing massive greenhouse gas emissions and smoke, detrimentally impacting human health and activities across Southeast Asian countries (Huijnen et al., 2016; Marlier et al., 2015). Both deforestation and forest fires have caused emerging biodiversity hotspots (Koh & Wilcove, 2008; Meijaard et al., 2018), including increased surface runoff and subsequent flooding risk (Dislich et al., 2017; Manoli et al., 2018; Merten et al., 2016). Numerous social conflicts have also arisen regarding tenurial issues (Dhiaulhaq et al., 2018; Krishna et al., 2017; Pasaribu et al., 2020; Prabowo et al., 2017); land and environmental degradation threatening local livelihood (Abram et al., 2017), labor rights (Murray Li, 2018; Pasaribu & Vanclay, 2021), unequal cost-benefit sharing (Pasaribu et al., 2020; Qaim et al., 2020), declining community well-being (especially in remote areas; Santika et al., 2019; 2021), and community displacement (Murray Li 2017, 2018).

Palm oil-based businesses have grown into a global chain, encompassing various industries, traders, retailers, and suppliers, including plantations managed by companies and smallholders (Pacheco et al., 2017), and they must be jointly responsible for maintaining business sustainability. Certain initiatives, such as the Roundtable on Sustainable Palm Oil (RSPO), have been introduced and, to some extent, have successfully reduced deforestation (Carlson et al., 2018; Cattau, Marlier, & DeFries, 2016; Noojipady et al., 2017).

Nevertheless, the continued forest and peat fires have prompted the Indonesian government to establish a moratorium policy to stop new oil palm plantation development via Presidential Instruction Number August 2018 and a moratorium policy to open primary forests and peatlands via Presidential Instruction Number May 2019. To date, these policies have succeeded in limiting the expansion of large-scale plantations by prohibiting the issue of new permits (Gatra, 2021; Koalisi moratorium sawit, 2021; Xin et al., 2021). However, smallholder oil palm development falls outside the jurisdiction of these regulations (Purnomo et al., 2020) as its management and expansion are barely monitored.

Smallholder oil palm plantations are receiving increased attention, as they cover >40% of the oil palm plantation area with lower productivity (contributing to 33.5% of the national production; Dirjenbun, 2020). In parallel, the recent growth of smallholder oil palms has accelerated (Bakhtary et al., 2021; Schoneveld et al., 2019). The price of oil palm, which is higher than other commodities, has driven farmers to convert their cultivated commodities or open new lands that often encroach on conservation zones and forested areas (Daemeter, 2015). It has been estimated that 68% or even bigger of illegal oil palm plantations in forest state areas (i.e., forest areas belonging to the state) are smallholder plantations (Nurfatriani et al., 2019).

Thus far, Indonesian smallholder oil palm data are only available in tabular form, without spatial references. This situation complicates the government’s efforts to control and monitor smallholder oil palm expansion or provide support to enhance good agricultural practices. Consequently, the extension of low productivity and unmaintained and poor agricultural practices across smallholder farming are causing significant economic and environmental losses (Dalheimer, Kubitza, & Brümmer, 2021).

There are presently two groups of smallholder oil palms—plasma and independent smallholders—where the former is cooperated with and managed by companies, while the latter is comprised of independently managed farms and covers ≥66% of the smallholder plantation area (Daemeter, 2015; Jelsma et al., 2017; Naylor et al., 2019). Collecting spatial data, especially for independent smallholders, is challenging, as they are smaller-scale, scattered, fragmented, remotely distributed (in some cases), irregular (i.e., maintain a variety of farmland shapes, planting distances, oil palm ages, vigor, varieties, etc.), mixed with other commodities and land uses, and rapidly changing.

Various efforts have been made to map smallholder oil palms. Kehati (2019) designed a smallholder oil palm mapping project based on a primary field survey assisted by unmanned aerial vehicle (UAV) and aerial photography data; yet, their findings are not fully applicable, as Indonesia contains scattered and vast areas of smallholder oil palms. Further research developed a smallholder oil palm mapping method using a finer scale satellite imagery (e.g., Sentinel, SPOT, etc.), combining optical and radar scenes, or using machine learning for land-use segmentation and classification (Descals et al., 2019; Okarda et al., 2018); although this method is limited in detecting smaller, fragmented, and scattered smallholder plantations. A study by the Indonesian National Institute of Aeronautics and Spaces (LAPAN) and World Resources Institute (WRI) using Sentinel imagery demonstrated its inability to detect oil palm plantations <25 ha (Pusfatja (Pusat Pemanfaatan Penginderaan Jauh)–LAPAN, 2021), whereas smallholder oil palms have an average size of ≤ 2–4 ha per farmer, (Aisyah et al., 2021; Nurliza, 2020; Watts et al., 2021).

Participatory mapping has also been introduced to accelerate data collection (Big Lottery Fund (Badan Informasi Geospasial), 2021). As expected, oil palm farmers are willing to register their farmlands in exchange for support and assistance from the government. However, this is not an easy approach, as many farmers maintain farmlands spread over vast areas. In addition, not all farmers cultivating on land have legal permission (Qaim et al., 2020; Schoneveld et al., 2019).

The increased availability of high-resolution imagery and deep-learning approaches have recently brought new opportunities. Dong et al. (2020) developed a finer segmentation method to map oil palm land-use in complex environments; however, it is only effective for capturing large, regular, and continuous palm oil plantations. Others have developed deep-learning methods to detect individual oil palm trees; however, they only focused on large-scale plantations aimed at counting trees (Li et al., 2017; Liu et al., 2020), detecting age (Mubin et al., 2019), identifying diseases (Yarak et al., 2021), and assessing plant health (Yarak et al., 2021; Zheng et al., 2021). These studies used UAVs to capture very high-resolution imagery of oil palms (Khan et al., 2021), but covering larger areas using UAVs is challenging.

Object identification models applied to high-resolution satellite imagery offer a better approach. Unfortunately, recent studies mainly focused on large-scale plantations where oil palm trees appear as single and easily identifiable objects (Li et al., 2017, 2019; Mubin et al., 2019). Freudenberg et al. (2019) considered oil palm and coconut trees as single-class objects, thus, cannot be used for detecting smallholder oil palms. These studies also worked on increasing the models’ accuracy and neglecting the speed factor.

Notably, the present study requires a more advanced approach, as small and scattered smallholder oil palms within mixed land-use areas inhibit efforts to achieve high accuracy of oil palm detection. In addition, high accuracies should be achieved at relatively high speeds to accelerate spatial data collection required for controlling and monitoring rapidly growing smallholder plantations. Therefore, accuracy and speed are equally essential, although building such a model is more complex.

Using high-resolution satellite imagery for object detection is also problematic due to the following issues: (1) objects in satellite images viewed from overhead (with some variations of visual angles) can inhibit object identification (Groener et al., 2019; Tian et al., 2020); (2) image quality can be affected by cloud cover, temperature, humidity, haze, sun angle, etc. (Groener et al., 2019); (3) object size decreases as the images cover more extensive areas (Tian et al., 2020); and (4) small object detection is challenging and requires considerable time and resources (Nguyen et al., 2020; Tong et al., 2020). The issues become increasingly complicated when smallholder oil palms are occasionally distributed within multiple commodities and complex land uses.

Against this backdrop, the present study aimed to (1) develop a new model for mapping smallholder oil palms, (2) describe the spatial distribution of smallholder oil palms, and (3) identify their potential impact on the environment. A deep-learning approach was applied to high-resolution satellite imagery for detecting oil palm trees with high speed and accuracy as a base for building maps of smallholder oil palm plantations. Model automation via software applications was also constructed to ease the model implementation. The generated map of smallholder oil palms can serve as a basis for analyzing how their spatial distribution might violate spatial planning policies and extend environmental risk.