Electronic Theses and Dissertation

Penilaian manual kematangan Tandan Buah Segar (TBS) kelapa sawit masih bersifat subjektif, bergantung pada pengalaman pengamat, dan kurang efisien untuk perkebunan skala besar. Ketidaktepatan penentuan tingkat kematangan TBS dapat berdampak pada waktu panen dan kualitas Crude Palm Oil (CPO). Penelitian ini mengembangkan sistem klasifikasi kematangan TBS kelapa sawit empat kelas, yaitu Mentah, Mengkal, Matang, dan Busuk, berbasis Edge AI pada perangkat NVIDIA Jetson Orin Nano. Penelitian ini mengintegrasikan Random Search Hyperparameter Tuning, konversi model ke format Open Neural Network Exchange (ONNX), Post-Training Quantization (PTQ), evaluasi stabilitas akurasi antara Kaggle dan Jetson, pengukuran emisi karbon serta Software Carbon Intensity (SCI), dan pemilihan model optimal menggunakan metode Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS). Tiga arsitektur Convolutional Neural Network (CNN) yang dibandingkan adalah EfficientNetV2-S, MobileNetV3-Large, dan Inception-v4 dengan format FP32, FP16, INT8 Dynamic, dan INT8 Static MinMax. Dataset yang digunakan terdiri atas 4.506 citra TBS kelapa sawit. Hasil penelitian menunjukkan bahwa EfficientNetV2-S menghasilkan performa klasifikasi terbaik setelah tuning dengan akurasi 97,12%, precision 0,9713, recall 0,9718, dan F1-score 0,9710. Perbandingan akurasi model ONNX pada Kaggle dan Jetson menunjukkan bahwa proses deployment pada perangkat edge tidak menyebabkan degradasi akurasi yang signifikan, dengan perubahan akurasi berada pada rentang -0,33 hingga 0,89 poin. Dari sisi deployment, MobileNetV3-Large FP16 menjadi konfigurasi paling optimal berdasarkan evaluasi multi-kriteria karena memperoleh akurasi 96,45%, latensi terendah 15,15 ms, FPS tertinggi 6,19, konsumsi daya 6,14 W, emisi CO₂ terendah 0,176 g, dan SCI terendah 0,3599. Hasil ini menunjukkan bahwa model dengan akurasi tertinggi belum tentu menjadi pilihan terbaik untuk perangkat edge. Pemilihan model Edge AI perlu mempertimbangkan akurasi, stabilitas antarplatform, latensi, FPS, konsumsi daya, emisi karbon, dan SCI secara bersamaan agar sistem yang dihasilkan tidak hanya akurat, tetapi juga efisien dan berkelanjutan.

Manual assessment of oil palm Fresh Fruit Bunch (FFB) ripeness remains subjective, depends heavily on the observer’s experience, and is less efficient for large-scale plantations. Inaccurate determination of FFB ripeness levels can affect harvesting time and the quality of Crude Palm Oil (CPO). This study develops a four-class oil palm FFB ripeness classification system, consisting of Unripe, Under-Ripe, Ripe, and Overripe classes, based on Edge AI using the NVIDIA Jetson Orin Nano device. This study integrates Random Search Hyperparameter Tuning, model conversion to the Open Neural Network Exchange (ONNX) format, Post-Training Quantization (PTQ), accuracy stability evaluation between Kaggle and Jetson, carbon emission and Software Carbon Intensity (SCI) measurement, and optimal model selection using Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) method. The three Convolutional Neural Network (CNN) architectures compared in this study are EfficientNetV2-S, MobileNetV3-Large, and Inception-v4, with FP32, FP16, INT8 Dynamic, and INT8 Static MinMax formats. The dataset used consists of 4,506 oil palm FFB images. The results show that EfficientNetV2-S achieved the best classification performance after tuning, with an accuracy of 97.12%, precision of 0.9713, recall of 0.9718, and F1-score of 0.9710. The comparison of ONNX model accuracy on Kaggle and Jetson indicates that deployment on the edge device did not cause significant accuracy degradation, with accuracy changes ranging from -0.33 to 0.89 points. From the deployment perspective, MobileNetV3-Large FP16 was selected as the most optimal configuration based on multi-criteria evaluation, achieving an accuracy of 96.45%, the lowest latency of 15.15 ms, the highest FPS of 6.19, power consumption of 6.14 W, the lowest CO₂ emission of 0.176 g, and the lowest SCI of 0.3599. These findings indicate that the model with the highest accuracy is not necessarily the best choice for edge devices. Edge AI model selection should consider accuracy, cross-platform stability, latency, FPS, power consumption, carbon emissions, and SCI simultaneously to produce a system that is not only accurate but also efficient and sustainable.