Electronic Theses and Dissertation

Penyakit tanaman menjadi salah satu tantangan utama dalam pertanian modern karena dapat menyebabkan penurunan produktivitas dan kualitas hasil panen. Penelitian ini mengembangkan model Multitask Convolutional Neural Network (CNN) dengan self-attention mechanism untuk melakukan klasifikasi otomatis jenis tanaman dan penyakit daun secara bersamaan. Model dirancang menggunakan tiga convolutional blocks dengan penerapan Squeeze-and-Excitation (SE) attention blocks untuk meningkatkan kemampuan ekstraksi fitur yang relevan. Sistem ini mampu mengklasifikasikan lima jenis tanaman, yaitu cabai, jagung, kentang, padi, dan tomat, serta empat belas jenis penyakit daun dengan total 22 kombinasi tanaman–penyakit yang valid. Dataset yang digunakan terdiri dari 14.801 citra daun yang dibagi secara stratified ke dalam data pelatihan (72%), validasi (18%), dan pengujian (10%). Proses pelatihan dilakukan menggunakan optimizer Adam yang dikombinasikan dengan learning rate scheduler dan early stopping untuk meningkatkan stabilitas pelatihan serta mengurangi risiko overfitting. Evaluasi pada data uji menunjukkan performa yang sangat baik dengan akurasi klasifikasi tanaman-penyakit sebesar 98,31% (1.456 prediksi benar dari 1.481 data uji), akurasi klasifikasi tanaman sebesar 99,53%, dan akurasi klasifikasi penyakit sebesar 98,72%. Model juga menunjukkan efisiensi komputasi yang tinggi dengan jumlah parameter sebanyak 136.291, ukuran model sebesar 1,64 MB, serta waktu inferensi rata-rata 57,92 ms per citra atau sekitar 17 citra per detik. Model selanjutnya diintegrasikan ke dalam aplikasi mobile berbasis Android menggunakan Flutter dan TensorFlow Lite, sehingga memungkinkan proses klasifikasi penyakit tanaman dilakukan secara langsung di lapangan tanpa memerlukan koneksi internet.

Kata kunci: Klasifikasi penyakit tanaman, Multi-task learning, Self-attention, CNN, Implementasi Android, Tensorflow lite

Plant diseases represent a major challenge in modern agriculture, as they can significantly reduce crop productivity and quality. This study develops a multitask Convolutional Neural Network (CNN) with a self-attention mechanism to simultaneously perform automatic classification of plant species and leaf diseases. The proposed model is designed using three convolutional blocks with the integration of Squeeze-and-Excitation (SE) attention blocks to enhance the extraction of relevant and discriminative features. The system is capable of classifying five plant species, namely chili, maize, potato, rice, and tomato, as well as fourteen types of leaf diseases, resulting in a total of twenty-two valid plant–disease combinations. The dataset consists of 14,801 leaf images, which are stratified into training (72%), validation (18%), and testing (10%) sets. Model training is conducted using the Adam optimizer combined with a learning rate scheduler and early stopping to improve training stability and reduce overfitting. Experimental results on the test set demonstrate strong performance, achieving a plant-disease classification accuracy of 98.31% (1,456 correct predictions out of 1,481 samples), a plant classification accuracy of 99.53%, and a disease classification accuracy of 98.72%. Furthermore, the model exhibits high computational efficiency with only 136,291 parameters, a model size of 1.64 MB, and an average inference time of 57.92 ms per image, corresponding to approximately 17 images per second. The proposed model is successfully integrated into an Android-based mobile application using Flutter and TensorFlow Lite, enabling real-time plant disease classification in field conditions without requiring an internet connection. Keywords: Plant disease classification, Multi-task learning, Self-attention, CNN, Android implementation, Tensorflow lite