Oil Palm Deep Learning (OPDL)

This is a repository for oil palm semantic segmentation from Sentinel-1 SAR and/or Sentinel-2 MS using fully convolutional neural network U-Net. The project was part of my MSc thesis supervised by Iris van Duren and Raian V. Maretto. The model was trained on University of Twente's CRIB.

The oil palm label was obtained from high-resolution oil palm map by Descals et al. 2021.

Background

Methods

Dataset preparation

Sentinel-1 median composite 2019

Sentinel-2 median composite 2019

Oil palm closed-canopy (Descals et al. 2021)

Oil palm mixed-canopy (Gaveau et al. 2021)

Split to train/val/test

Deep learning model training

Single input

Stack input

Dual input (with spatial attention)

Inference

Results

Model performances

Map accuracy

Changes on riparian zone

Interactive map for year 2021

An interactive view of the oil palm map along rivers produced from this study can be accessed at:
-Leaflet map
-https://bit.ly/oilpalmrivers (depcrecated)