<![CDATA[ABSTRAK Produktivitas cabai merah di Matani Satu terhambat oleh serangan patogen, diagnosis manual yang subjektif, serta infrastruktur internet perkebunan yang tidak stabil. Penelitian ini bertujuan membangun model Deep Learning berarsitektur EfficientNet-B0 untuk deteksi dini penyakit daun cabai merah dan merancang sistem rekomendasi penanggulangan cerdas berbasis Progressive Web App (PWA). Metodologi penelitian menggunakan 2.141 dataset hibrida dengan tahapan akuisisi data, pra-pemrosesan, pemodelan transfer learning, penalaan parameter menggunakan bobot kelas (Class Weights) untuk memitigasi ketidakseimbangan data, perancangan basis pengetahuan, hingga implementasi sistem. Hasil evaluasi menunjukkan model mencapai akurasi validasi sebesar 94,39% dengan tingkat loss 0,21. Pendekatan Cost-Sensitive Learning terbukti efektif mengenali kelas penyakit minoritas secara sempurna. Fitur Explainable AI berupa Grad-CAM juga berhasil diimplementasikan guna memberikan transparansi visual pada area infeksi daun. Kesimpulannya, integrasi inferensi AI di sisi klien (client-side) melalui PWA menjadi solusi yang efektif, memungkinkan para petani mandiri untuk mendiagnosis penyakit dan memperoleh rekomendasi tindakan agronomis secara real-time meskipun dalam kondisi tanpa koneksi internet (offline). ABSTRACT Red chili productivity in Matani Satu is hindered by pathogen attacks, subjective manual diagnosis, and unstable plantation internet infrastructure. This study aims to build a Deep Learning model with an EfficientNet-B0 architecture for early detection of red chili leaf diseases and design an intelligent treatment recommendation system based on a Progressive Web App (PWA). The research methodology utilizes 2,141 hybrid datasets involving data acquisition, preprocessing, transfer learning modeling, parameter tuning using Class Weights to mitigate data imbalance, knowledge base design, and system implementation. Evaluation results indicate the model achieved a validation accuracy of 94.39% with a loss rate of 0.21. The Cost-Sensitive Learning approach proved effective in perfectly recognizing minority disease classes. The Explainable AI feature, Grad-CAM, was also successfully implemented to provide visual transparency of leaf infection areas. In conclusion, the integration of client-side AI inference through PWA provides an effective solution, enabling independent farmers to diagnose diseases and obtain agronomic action recommendations in real-time even without an internet connection (offline). ]]>
