Klasifikasi Tingkat Kematangan Buah Kelapa Sawit Menggunakan Convolution Neural Netwok Berbasis Perangkat Edge

Dynisyah, M Filza Abqory Dynisyah (2026) Klasifikasi Tingkat Kematangan Buah Kelapa Sawit Menggunakan Convolution Neural Netwok Berbasis Perangkat Edge. Other thesis, Intitut Teknologi Sepuluh Nopember.

Text 5024211055-Master_Thesis.pdf Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Penentuan tingkat kematangan buah kelapa sawit merupakan faktor krusial dalam men jaga kualitas dan produktivitas minyak kelapa sawit. Namun demikian, proses penilaian ke matangan yang masih dilakukan secara manual di lapangan cenderung bersifat subjektif dan tidak konsisten, sehingga berpotensi menurunkan efisiensi panen. Oleh karena itu, penelitian ini bertujuan untuk mengembangkan sistem klasifikasi tingkat kematangan visual buah kelapa sawit berbasis citra digital menggunakan metode Convolutional Neural Network (CNN) yang diimplementasikan pada perangkat edge computing. Pada penelitian ini, buah kelapa sawit dik lasifikasikan ke dalam tiga kelas tingkat kematangan visual, yaitu setengah matang, matang, dan terlalu matang. Empat arsitektur CNN digunakan dan dibandingkan, yaitu MobileNetV3, ResNet18, EfficientNet-B0, dan DenseNet121. Proses pelatihan model dilakukan pada perangkat PC high-end menggunakan GPU NVIDIA GeForce RTX 4090. Selanjutnya, evaluasi model pada perangkat laptop dilakukan menggunakan berkas model berformat .pth. Sementara itu, pada perangkat edge Jetson Nano dilakukan konversi model dari format .pth ke .onnx serta optimasi menggunakan TensorRT untuk mendukung proses inferensi pada perangkat dengan sumber daya komputasi terbatas. Hasil pengujian menunjukkan bahwa DenseNet121 dan Mo bileNetV3 memperoleh nilai akurasi tertinggi pada perangkat PC high-end, yaitu sebesar 95,28%. Di sisi lain, EfficientNet-B0 menunjukkan performa paling stabil pada perangkat laptop dengan nilai akurasi mencapai 99,98%. Pada implementasi di perangkat edge Jetson Nano, model DenseNet121 kembali menunjukkan performa terbaik dengan nilai akurasi sebesar 94,88%. Penurunan akurasi pada perangkat edge dipengaruhi oleh keterbatasan sumber daya komputasi serta proses konversi dan optimasi model. Berdasarkan hasil tersebut, dapat disimpulkan bahwa pendekatan CNN berbasis edge computing mampu mengklasifikasikan tingkat kematangan vi sual buah kelapa sawit dengan baik dan berpotensi diterapkan secara real-time di lingkungan perkebunan. Penelitian ini diharapkan dapat menjadi dasar pengembangan sistem cerdas un tuk mendukung proses panen dan sortasi buah kelapa sawit secara lebih objektif, efisien, dan konsisten.
=================================================================================================================================
Determining the ripeness level of oil palm fruit is a crucial factor in maintaining the quality and productivity of palm oil. However, ripeness assessment in the field is still predominantly conducted manually, which tends to be subjective and inconsistent, thereby potentially reduc ing harvesting efficiency. Therefore, this study aims to develop an image-based oil palm fruit ripeness classification system using the Convolutional Neural Network (CNN) method imple mented on edge computing devices. In this study, oil palm fruit is classified into three visual ripeness levels, namely semi-ripe, ripe, and overripe. Four CNN architectures are employed and compared, including MobileNetV3, ResNet18, EfficientNet-B0, and DenseNet121. Model training is conducted on a high-end PC using an NVIDIA GeForce RTX 4090 GPU. Subse quently, model evaluation on a laptop is performed using the trained model in .pth format. Meanwhile, for deployment on the Jetson Nano edge device, the model is converted from .pth to .onnx format and optimized using TensorRT to support inference on resource-constrained hardware. Experimental results show that DenseNet121 and MobileNetV3 achieve the highest accuracy on the high-end PC, reaching 95.28%. On the other hand, EfficientNet-B0 demon strates the most stable performance on the laptop with an accuracy of up to 99.98%. In the edge computing implementation, DenseNet121 again achieves the best performance with an accuracy of 94.88%. The decrease in accuracy on the edge device is influenced by limited computational resources as well as the model conversion and optimization processes. Based on these results, it can be concluded that a CNN-based edge computing approach is capable of effectively classify ing the visual ripeness level of oil palm fruit and has strong potential for real-time deployment in plantation environments. This research is expected to serve as a foundation for developing intelligent systems to support more objective, efficient, and consistent oil palm harvesting and sorting processes.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Klasifikasi tingkat kematangan, buah kelapa sawit, Convolutional Neural Network, edge computing, Jetson Nano ========================================================= Oil palm fruit ripeness classification, Convolutional Neural Network, Edge computing, Jetson Nano
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T58.6 Management information systems T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5105.546 Computer algorithms T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7888.3 Digital computers
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	M Filza Abqory Dynisyah
Date Deposited:	15 Jan 2026 09:46
Last Modified:	15 Jan 2026 09:46
URI:	http://repository.its.ac.id/id/eprint/129668

Actions (login required)

View Item