University of Twente Student Theses


Detecting of natural forest to oil palm conversions in tropical wetlands based on sentinel imagery using deep learning

Muzakki, Fauzan (2021) Detecting of natural forest to oil palm conversions in tropical wetlands based on sentinel imagery using deep learning.

Link to full-text:
(only accessible for UT students and staff)
Abstract:Natural forest conversion to oil palm in tropical wetlands is dangerous for humans and biodiversity. Their capability to store carbon and flood prevention becomes the main reason that stakeholders should protect natural forest areas and tropical wetlands. The moratorium was enacted in 2018 to ban the expansion of oil palm plantations by companies or smallholder farmers through natural forest loss in tropical wetlands. Visual inspection of Google Earth imagery shows violations of the moratorium remains in Pelalawan Regency, Riau Province. The finding raises a question about the effectiveness of the moratorium to reduce oil palm expansion in tropical wetlands. Meanwhile, the Indonesian government legalized the conventional techniques (on-screen digitization) to map land cover in Indonesia. It is uncertain because prone to human error and time-consuming. Moreover, RPSO has been criticized by researchers because they have lack land trajectory information to track the data of previous land cover before becoming oil palm plantations. Remote sensing data has been used by academia to discriminate between land cover types and change detection. The synergies of Sentinel 1 and Sentinel 2 is the potential to overcome the limitation of Sentinel 1 and Sentinel 2 data alone for land cover classification and change detection. Moreover, the FCN model also has been used to automate semantic segmentation of remote sensing data in different resolutions and scales in the decoder-encoder style. This study examines which method (single imagery-based, synergy of Sentinel 1 and Sentinel 2 data) produces the reliable land cover classification. Then, using the best technique is to generate the map of natural forest to oil palm conversions in tropical wetlands, Pelalawan Regency, Riau Province, Indonesia. Also, this study wants to evaluate the moratorium's progress by comparing the change detection before (2016-2018) and after the moratorium period (2018-2020). This study figured out that the synergy of Sentinel 1 and Sentinel 2 data improved land cover classification compared to the single imagery-based technique. Using a GIS-practical approach, the post-classification composition technique achieved the highest accuracy over time with the averaged F1 score of 0.67. The natural forest to oil palm conversion map presents that the conversion patterns of the riparian zones and peatlands are different spatially. Conversions in riparian zones elongated follow the river, whereas conversions were irregularly distributed in peatlands. Regardless of misclassification in the conversion map, the moratorium looks like it did not work as it should be because this study observed the conversion area increased after the moratorium period, especially in riparian zones. Nevertheless, the method used in this study is more automated than conventional techniques used by the Indonesian government. Further investigation is necessary to improve the land cover classification and change detection accuracy using different sampling techniques and deep learning models. Moreover, the natural forest's trend, time, and magnitude to oil palm conversions based on Sentinel 1 and Sentinel 2 time-series data are also essential to further understand how effectively moratorium reduces oil palm expansion through natural forest loss.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Spatial Engineering MSc (60962)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page