Klasifikasi citra satelit kebakaran hutan dan lahan di Pulau Sumatra menggunakan U-Net = Satellite images wildfire classification in Sumatra using U-Net

Witarsah, Marcellinus Aditya (2022) Klasifikasi citra satelit kebakaran hutan dan lahan di Pulau Sumatra menggunakan U-Net = Satellite images wildfire classification in Sumatra using U-Net. Bachelor thesis, Universitas Pelita Harapan.

Preview	Text (Title) Title.pdf Available under License Creative Commons Attribution Non-commercial Share Alike. Download (65kB) | Preview
Preview	Text (Abstract) Abstract.pdf Available under License Creative Commons Attribution Non-commercial Share Alike. Download (293kB) | Preview
Preview	Text (ToC) ToC.pdf Available under License Creative Commons Attribution Non-commercial Share Alike. Download (1MB) | Preview
	Text (Chapter1) Chapter1.pdf Restricted to Registered users only Available under License Creative Commons Attribution Non-commercial Share Alike. Download (722kB)
	Text (Chapter2) Chapter2.pdf Restricted to Registered users only Available under License Creative Commons Attribution Non-commercial Share Alike. Download (4MB)
	Text (Chapter3) Chapter3.pdf Restricted to Registered users only Available under License Creative Commons Attribution Non-commercial Share Alike. Download (1MB)
	Text (Chapter4) Chapter4.pdf Restricted to Registered users only Available under License Creative Commons Attribution Non-commercial Share Alike. Download (16MB)
	Text (Chapter5) Chapter5.pdf Restricted to Registered users only Available under License Creative Commons Attribution Non-commercial Share Alike. Download (352kB)
Preview	Text (Bibliography) Bibliography.pdf Available under License Creative Commons Attribution Non-commercial Share Alike. Download (629kB) | Preview
	Text (Appendices) Appendices.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial Share Alike. Download (9MB)

Abstract

Indonesia termasuk dalam peringkat sepuluh dengan area hutan terluas di bumi. Namun, Gangguan terhadap sumber daya hutan dan keberlangsunganya tidak dapat dihindari, misalnya kebakaran hutan dan lahan. Kebakaran hutan dan lahan telah menjadi masalah serius karena berdampak negatif bagi makhluk hidup dan lingkungan sekitar. Dampak negatif ini dapat diminimalisir dengan mengidentifikasi kebakaran hutan dan lahan secara cepat. Sayangnya, manusia mengalami kesulitan dalam mengidentifikasi lokasi kebakaran hutan dan lahan apabila terjadi di daerah terpencil akibat keterbatasan pandang mata manusia. Keterbatasan ini dapat diatasi dengan cara mengidentifikasi kebakaran hutan dan lahan melalui citra satelit dengan menggunakan model deep learning, yakni U-Net. Penelitian ini melibatkan penggunaan composite bands dan perhitungan nilai delta normalized burned ratio dari citra satelit Landsat 8 dan Sentinel 2. Langkah ini akan menghasilkan pasangan gambar natural images dengan delta normalized burned ratio (mask), dan false-color images dengan delta normalized burned ratio (mask). Kemudian, kedua pasangan gambar akan digunakan untuk melatih model U-Net secara satu per satu. Hasil pelatihan menggunakan berbagai pasangan gambar akan dibandingkan demi tercapainya model yang andal dalam mengklasifikasi wilayah yang mengalami kebakaran hutan. Melalui serangkaian hyperparameter tuning, model U-Net terbaik ditemukan dengan accuracy 0.94, binary intersection over union 0.78, dice coefficient 0.69, dan binary cross-entropy loss 0.14. Berdasarkan hasil tersebut, dapat disimpulkan bahwa model ini secara baik dapat menandai area pixel kebakaran hutan dan lahan melalui citra satelit. Model telah berhasil diluncurkan dalam website yang dibuat dengan menggunakan Streamlit sehingga bisa diakses dan digunakan oleh publik. / Indonesia has one of the world's largest forest areas, ranking tenth. However, disturbances to forest resources and their sustainability cannot be avoided, e.g. wildfires. Wildfires have become a serious problem because they have an adverse effect on living things and the environment. This negative impact can be minimized by quickly identifying wildfires. Unfortunately, humans have difficulty in identifying the location of wildfires, especially in remote areas due to the limited visibility of the human eye. This limitation can be overcome by identifying wildfires through satellite imagery using a deep learning model, namely U-Net. This research involves using composite bands and calculating the delta normalized burned ratio from Landsat 8 and Sentinel 2 satellite images. This step will produce a pair of natural images with a delta normalized burned ratio (mask) and false-color images with a delta normalized burned ratio (masks). Then, the two pairs of images will be used to train the U-Net model one by one. The results of the training using various pairs of images will be compared with each other in order to achieve a reliable model in classifying wildfire areas. Through a series of hyperparameter tuning, the best U-Net model was found with an accuracy of 0.94, a binary intersection over union of 0.78, a dice coefficient of 0.69, and a binary cross-entropy loss of 0.14. Based on these results, it can be concluded that this model can properly mark the pixel area of wildfires through satellite imagery. The model has been successfully launched on a website created using Streamlit so that it can be accessed and used by the public.

Actions (login required)

View Item