??Palm Oil Land Use Mapping in Malaysia??

When I was doing my Bachelor's in GIS, I spent 408 hours learning Land Use Planning and Monitoring. This is what my supplement to my diploma says.

For the project below, I spent much less time, but I must say it was like a flashback to my university years. This was my first time applying deep learning to create a land use map, and, however, I still see a few things to be improved, I believe the result for this work is "passed."

Background

Carey Island, located in Selangor district, near Kuala Lumpur, Malaysia ????, is renowned for its extensive oil palm plantations.

The largest owner of palm oil plantations on Carey Island is Sime Darby Plantation Berhad.

Sime Darby is one of the world's largest producers of palm oil and manages vast estates on the island, making them a significant player in the region's palm oil industry.

Considering my other big love - agritech, I enjoyed to watch this futuristic video about the latest plant oil tech innovations, created by this giant of the industry. And it also features Carey Island.

Challenge

The primary technical challenge in this project was processing and analyzing a massive dataset of 500,000 images to accurately classify the land use on Carey Island. The foundational satellite image had an impressive resolution of 30 centimeters per pixel, divided over a 20x20 meter grid.

The borders of the land use mask were delineated manually, using natural boundaries such as the Langat River and the Strait of Malacca. The region mask covered an area of 180 square kilometers.

This task required a sophisticated approach to handle the enormous volume of data while ensuring the precision of land-use classification.

Why?

Understanding the land use of Carey Island is crucial for several reasons. The island is a significant hub for oil palm plantations, a major economic driver in Malaysia. Accurate land-use mapping helps monitor the growth and health of these plantations, assess environmental impacts, and track changes over time.

Additionally, it aids environmentalists in evaluating afforestation and deforestation rates (as one of the climate change mitigation aspects).

Hence, understanding of land use is essential for making a balance between economic interests and environmental preservation.

Journey

The journey of creating the land-use map of Carey Island began with leveraging self-supervised learning techniques to process the 500,000 images. This initial step involved identifying four primary clusters, namely:

1. Palm Oil Tree
2. Road / infrastructure object
3. Other vegetation
4. Water body.

The next phase utilized supervised learning to refine these clusters and produce a more precise visual representation.

Constructed using PyTorch, the customized model demonstrated proper adaptability, capable of managing various imagery datasets even without prior self-supervised learning. However, fine-tuning was necessary for updated imagery to maintain pinpoint accuracy.

This AI-powered land-use map serves as a powerful tool for segmentation tasks:

1. Environmentalists can leverage it to assess specific patterns such as the growth trends of palm oil trees or the rates of afforestation and deforestation.
2. Additionally, it offers significant commercial value by enabling more efficient land management and planning for agricultural businesses and urban developers.

This project demonstrates the transformative role AI can play in understanding and managing the impacts of human interventions and climate change on the environment.

By automating the labor-intensive process of categorizing extensive datasets, this project demonstrates how technology can provide precise, actionable insights that drive positive environmental and economic outcomes.

Public Demonstrations

This poster was presented at the AI + Environment Summit at ETH Zurich in November 2023. It was a full-day summit, organised by ETH AI Center BioDivX, to inspire, ignite and innovate work in AI for the environment.

The project above was demonstrated under the title "Transforming Satellite Imagery into Environmental Insights: AI-Powered Land-Use Mapping of Carey Island in Malaysia."

Discussion

While working on this project and reflecting on it, I discovered some interesting research papers related to land use as well as land use and land cover (LULC). Hopefully, you'll find them interesting as well ??

Comparison of Three Machine Learning Algorithms Using Google Earth Engine for Land Use Land Cover Classification

The article examines the effectiveness of three machine learning models—CART, SVM, and RF—in classifying land use and land cover (LULC) in Mardan, Pakistan, using Sentinel-2 data. The study finds that the Random Forest (RF) model outperforms the other models with an overall accuracy of 98.68% and a Kappa coefficient of 0.97. This demonstrates RF's superior capability in processing high-resolution satellite imagery for reliable LULC mapping, highlighting its potential for environmental monitoring and land management applications

Characterizing land use/land cover change dynamics by an enhanced random forest machine learning model: a Google Earth Engine implementation

The other paper also explores the use of advanced machine learning techniques to improve the accuracy of LULC mapping in the Shrirampur area of Maharashtra, India.

The research team employs the Random Forest algorithm with 50 and 100 tree models and successfully generated LULC maps for 2014 and 2020 using multitemporal Landsat-8 satellite imagery.

The results showed that the RF-100 model provided higher accuracy, highlighting significant changes in agricultural land, built-up areas, wastelands, and water bodies.

This research emphasizes the importance of selecting optimal models for precise LULC mapping and offers practical recommendations for future studies in environmental monitoring and land management.

A novel geospatial machine learning approach to quantify non-linear effects of land use/land cover change (LULCC) on carbon dynamics

Finally, I really enjoyed this research . which is focused on specific topic of monitoring the carbon dynamics. The team developed a hybrid model combining linear regression and machine learning to analyze the impact of LULCC on CO2 emissions in mainland China from 2016 to 2019.

Using high-resolution Sentinel-2 imagery and top-down CO2 data, the study identifies significant changes in carbon sinks, including a 12% decrease in snow cover and a 7% decrease in water bodies, while urban areas expanded by 5%.

The study also notes changes in mixed forest dynamics, which significantly contribute to CO2 absorption, but experienced a 4% decrease during the same period, further impacting the overall carbon balance.

The findings highlight the importance of effective land management for carbon sequestration and climate mitigation.

I hope you enjoyed two editions of the newsletter "Above Spatial". Please, share your feedback on this edition and I'll do my best to bring you something interesting next time ??

Sincerely,

Andrii Seleznov, MSc (Hons) GIS

PS: if you need any advice regarding your GIS project - please let me know, I am open for collaborations ???????? (strong handshake)

References