Segmentasi Gulma Tanaman Padi Pada Citra UAV Multispektral Menggunakan SegFormer

Auditya, Marsyanda (2026) Segmentasi Gulma Tanaman Padi Pada Citra UAV Multispektral Menggunakan SegFormer. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221055-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (24MB) | Request a copy

Abstract

Gulma tanaman padi merupakan salah satu ancaman utama produktivitas lahan padi di Asia Tenggara, dengan potensi penurunan hasil panen yang signifikan bahkan melebihi 50% pada lahan yang terinfestasi gulma berat. Identifikasi dan pemetaan gulma secara manual pada skala lahan luas membutuhkan waktu dan tenaga yang sangat besar. Pendekatan segmentasi berbasis citra RGB terkendala oleh kemiripan morfologi antara gulma dan padi budidaya yang tidak dapat dibedakan hanya dari informasi warna. Kanal Near Infrared (NIR) dan Red Edge (RE) pada citra multispektral merekam perbedaan respons spektral kedua tanaman yang tidak tertangkap oleh RGB. Penelitian ini menerapkan model SegFormer untuk segmentasi gulma tanaman padi pada citra UAV multispektral dengan membandingkan dua strategi fusi, yaitu Early Fusion dan Late Fusion, menggunakan dataset WeedyRice-RGBMS-DB. SegFormer dipilih karena keunggulannya dibandingkan arsitektur Transformer lain untuk segmentasi gulma pada citra pertanian. Pra-pemrosesan meliputi augmentasi dengan enam variasi transformasi geometrik per citra, resize ke 512x512 piksel, konversi tensor, dan normalisasi statistik per kanal. Model dilatih selama 100 epoch menggunakan optimizer AdamW dengan learning rate 2x10^-6. Hasil pengujian menunjukkan seluruh skenario Late Fusion melampaui skenario RGB unimodal, sedangkan seluruh skenario Early Fusion berada di bawahnya. Skenario terbaik Late Fusion RGB+NIR mencapai mIoU 0.8848, IoU dan DSC kelas gulma tanaman padi masing-masing 0.8487 dan 0.9181. Hasil ini menunjukkan bahwa integrasi kanal multispektral melalui encoder terpisah meningkatkan akurasi segmentasi gulma tanaman padi dan berpotensi diterapkan dalam sistem pemetaan gulma berbasis UAV untuk pertanian presisi.
========================================================================================================================================
Weedy rice is one of the primary threats to paddy field productivity in Southeast Asia, with potential yield losses exceeding 50% under heavy infestation. Manual identification and mapping of weedy rice at field scale requires significant time and labor. RGB-based segmentation approaches are limited by the morphological similarity between weedy rice and cultivated rice, which cannot be distinguished by color information alone. The Near Infrared (NIR) and Red Edge (RE) channels in multispectral imagery record spectral response differences between the two plants that are not captured by RGB. This study applies the SegFormer model for weedy rice segmentation on multispectral UAV imagery by comparing two fusion strategies, namely Early Fusion and Late Fusion, using the WeedyRice-RGBMS-DB dataset. SegFormer was selected for its demonstrated superiority over other Transformer architectures for weed segmentation in agricultural imagery. Preprocessing includes augmentation of six geometric transformation variants per image, resizing to 512x512 pixels, tensor conversion, and per-channel statistical normalization. The model was trained for 100 epochs using the AdamW optimizer with a learning rate of 2x10^-6. Test results show that all Late Fusion scenarios surpassed the unimodal RGB baseline, while all Early Fusion scenarios performed below it. The best scenario, Late Fusion RGB+NIR, achieved an mIoU of 0.8848, with weedy rice class IoU and DSC of 0.8487 and 0.9181 respectively. These results indicate that integrating multispectral channels through separate encoders improves weedy rice segmentation accuracy and shows potential for application in UAV-based weed mapping systems for precision agriculture.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Segmentasi, Gulma Tanaman Padi, Citra UAV Multispektral, SegFormer, Pertanian Presisi, Segmentation, Weedy Rice, Multispectral UAV Imagery, SegFormer, Precision Agriculture
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Marsyanda Auditya
Date Deposited:	01 Aug 2026 02:18
Last Modified:	01 Aug 2026 02:18
URI:	http://repository.its.ac.id/id/eprint/140909

Actions (login required)

View Item