Electronic Theses and Dissertation

Penentuan tingkat roasting biji kopi secara konvensional masih bergantung pada
pengamatan visual oleh roasting master sehingga bersifat subjektif dan rentan terhadap
kesalahan manusia. Di sisi lain, sebagian besar penelitian berbasis Deep learning masih
menggunakan pendekatan klasifikasi yang hanya menghasilkan kategori diskrit tanpa
mampu merepresentasikan perubahan tingkat roasting yang bersifat kontinu maupun
ambiguitas pada batas antar kelas. Penelitian ini bertujuan mengembangkan sistem estimasi
tingkat roasting berbasis citra melalui integrasi EfficientNetB0-Based Regression dan
Logika Fuzzy menggunakan Trapezoidal Membership Function. Tahap awal penelitian
dilakukan dengan membentuk Pseudo Agtron sebagai ground truth melalui ekstraksi
Feature vector menggunakan EfficientNetB0, reduksi dimensi menggunakan Principal
Component Analysis (PCA), pembentukan PCA Rank, serta pemetaan linier ke rentang
nilai Agtron 30–95 yang disesuaikan dengan karakteristik dataset. Selanjutnya
dikembangkan model EfficientNetB0-Based Regression menggunakan strategi fine-tuning
bertahap dan dievaluasi menggunakan skema 5-Fold cross validation serta data testing
independen. Nilai Pseudo Agtron hasil prediksi kemudian diintegrasikan dengan Logika
Fuzzy untuk menghitung derajat keanggotaan, menentukan tingkat roasting, dan
mengidentifikasi ambiguitas prediksi. Hasil penelitian menunjukkan bahwa model
menghasilkan rata-rata MAE sebesar 1,5503, RMSE sebesar 2,1442, MAPE sebesar
2,91%, dan R² sebesar 0,9887 pada 5-fold cross validation. Pada data testing, model
memperoleh MAE sebesar 1,1898, RMSE sebesar 1,4982, MAPE sebesar 2,24%, dan R²
sebesar 0,9945. Integrasi Logika Fuzzy berhasil merepresentasikan derajat keanggotaan
setiap tingkat roasting serta memberikan informasi Ambiguity Index dan tingkat
ambiguitas pada daerah transisi antar kelas. Hasil penelitian menunjukkan bahwa sistem
yang dikembangkan mampu mengestimasi nilai Pseudo Agtron secara akurat serta
memberikan interpretasi tingkat roasting yang lebih informatif, objektif, dan non-destruktif
sehingga berpotensi mendukung proses quality control pada industri kopi.

Conventional determination of coffee bean roast levels still relies on visual observation by the roasting master, making it subjective and prone to human error. On the other hand, most deep learning-based research still uses a classification approach that only produces discrete categories without being able to represent continuous changes in roast levels or ambiguities at class boundaries. This study aims to develop an image-based roasting level estimation system through the integration of EfficientNetB0-based regression and fuzzy logic using the trapezoidal membership function. The initial stage of the study involved defining Pseudo Agtron as the ground truth through feature vector extraction using EfficientNetB0, dimension reduction using Principal Component Analysis (PCA), the calculation of PCA ranks, and linear mapping to the Agtron value range of 30–95, adjusted to the dataset’s characteristics. Subsequently, an EfficientNetB0-Based Regression model was developed using a stepwise fine-tuning strategy and evaluated using a 5-fold cross-validation scheme and independent test data. The predicted Pseudo Agtron values were then integrated with Fuzzy Logic to calculate membership degrees, determine the roasting level, and identify prediction ambiguities. The results show that the model yielded an average MAE of 1.5503, an RMSE of 2.1442, a MAPE of 2.91%, and an R² of 0.9887 in a 5-fold cross-validation. On the test data, the model achieved an MAE of 1.1898, an RMSE of 1.4982, a MAPE of 2.24%, and an R² of 0.9945. The integration of Fuzzy Logic successfully represented the membership degrees of each roasting level and provided information on the Ambiguity Index and the level of ambiguity in the transition regions between classes. The results of the study show that the developed system is capable of accurately estimating Pseudo Agtron values and providing a more informative, objective, and non-destructive interpretation of roasting levels, thereby potentially supporting the quality control process in the coffee industry.