Electronic Theses and Dissertation

Singkong (Manihot esculenta Crantz) merupakan komoditas pangan strategis di Indonesia yang produksinya terancam oleh empat penyakit utama, yaitu Cassava Mosaic Disease (CMD), Cassava Bacterial Blight (CBB), Cassava Brown Streak Disease (CBSD), dan Cassava Green Mottle (CGM). Deteksi penyakit berbasis deep learning konvensional memerlukan data berlabel dalam jumlah besar yang sulit diperoleh. Penelitian ini menerapkan Few-Shot Learning (FSL) berbasis AI multimodal yang mengintegrasikan fitur visual CLIP ViT-L/14@336px dan EfficientNet-B0 dengan deskripsi teks minimal per kelas. Dataset yang digunakan adalah Cassava Leaf Disease dari Makerere AI Lab yang terdiri atas 21.397 citra dari lima kelas. Eksperimen menggunakan 16 sampel latih dan 4 sampel uji per kelas, membandingkan Zero-Shot CLIP (baseline), Tip-Adapter (non-parametrik), dan Linear Probe++ (LP++) dengan optimizer L-BFGS. Pada skenario 0-shot hingga 16-shot, Zero-Shot CLIP hanya mencapai akurasi 25,00%, sedikit di atas tebakan acak (20,00%), sementara Tip-Adapter meningkat hingga 75,00% pada 8-shot namun stagnan pada 16-shot. LP++ dengan L-BFGS mengungguli kedua metode lainnya, mencapai akurasi 95,00% mulai 8-shot dan bertahan pada 16-shot, melampaui metode fully supervised ITIMCA (78,00%). Confusion matrix pada 16-shot menunjukkan empat kelas (CBSD, CGM, Healthy, CMD) mencapai akurasi 100%, dengan satu kesalahan klasifikasi CBB ke CGM akibat kemiripan gejala klorosis. Ablation study menunjukkan deskripsi teks minimal (5-7 kata) menghasilkan performa setara dengan deskripsi panjang dan LP++ tidak sensitif terhadap panjang prompt. Kombinasi CLIP, EfficientNet-B0, dan L-BFGS dengan deskripsi teks minimal merupakan adaptasi few-shot yang efisien dan andal untuk klasifikasi penyakit daun singkong pada data sangat terbatas, dengan implikasi praktis bagi sistem pendukung keputusan pertanian singkong di Indonesia.
Kata Kunci: Klasifikasi Penyakit Singkong, Few-Shot Learning, AI Multimodal, CLIP, Tip-Adapter, Linear Probe++.

Cassava (Manihot esculenta Crantz) is a strategic Indonesian food commodity, with production threatened by four major leaf diseases: Cassava Mosaic Disease (CMD), Cassava Bacterial Blight (CBB), Cassava Brown Streak Disease (CBSD), and Cassava Green Mottle (CGM). Conventional deep learning-based detection requires large amounts of labeled data difficult to obtain. This study applies multimodal AI-based Few-Shot Learning (FSL) integrating visual features from CLIP ViT-L/14@336px and EfficientNet-B0 with minimal text descriptions per class. The dataset used is the Cassava Leaf Disease dataset from Makerere AI Lab, comprising 21,397 field images across five classes. The experiment used 16 training samples and 4 test samples per class, comparing Zero-Shot CLIP (baseline), Tip-Adapter (non-parametric), and Linear Probe++ (LP++) with the L-BFGS optimizer. From 0-shot to 16-shot, Zero-Shot CLIP achieved only 25.00% accuracy, slightly above random guessing (20.00%), while Tip-Adapter rose to 75.00% at 8-shot but stagnated at 16-shot. LP++ with L-BFGS consistently outperformed both, reaching 95.00% accuracy from 8-shot onward and maintaining it at 16-shot, surpassing the fully supervised ITIMCA (78.00%). The confusion matrix at 16-shot showed perfect 100% accuracy for four classes (CBSD, CGM, Healthy, CMD), with one CBB-to-CGM misclassification due to similar chlorosis symptoms. An ablation study showed minimal text descriptions (5-7 words) yield performance comparable to longer ones, and LP++ is insensitive to prompt length. The combination of CLIP, EfficientNet-B0, and L-BFGS with minimal text descriptions provides an efficient and reliable few-shot approach for cassava leaf disease classification under highly limited data, with practical implications for agricultural decision support in Indonesia. Keywords: Cassava Disease Classification, Few-Shot Learning, Multimodal AI, CLIP, Tip-Adapter, Linear Probe++.