High demand for caisim in Indonesia's main export commodity must be accompanied by a good planting process. The obstacle faced is that farmers are currently able to apply pesticides when the caisim plants have holes due to being eaten by pests. This control can be a good step to maximize the yield of caisim farming. The high demand for caisim in Indonesia's main export commodity must be carried out with a good planting process. However, many farmers have not implemented proper control of pests, one of which is farmers in Kebon Raya Dempo. The obstacles faced such as not being able to
detect pests correctly and provide pesticides with precision. This control can be a good step to maximize the yield of caisim farming. Motivated by CNN's success in image
classification, a learning-based approach has been carried out in this study to detect the presence of pests in caisim. The experimental results show differences in accuracy in each experiment with a dataset of 1000, consisting of 500 with pests and 500 without pests. The accuracy of the experiment A – CNN from Scratch is 48.33%, precision 1, recall 0.48, F1-score 0.65 results in an underfitting model, experiment B – CNN from Scratch is 73.00% precision 1, recall 0.64, F1- score 0.78 results in an overfitting model, the C– CNN from Scratch experiment is 92.00% precision 0.88, recall 0.96, F1-score 0.92 results in an overfitting model, the D – CNN add model VGG16 experiment is 95.00%, precision 0.91 , recall 0.98, F1-score 0.95 results in a usable model, experiment E – CNN add Xception model 97.00%, precision 0.96, recall 0.98, F1-score 0.97 results produce a usable model and experiment F – CNN add NASNetMobile model of 93.00%, precision 0.91, recall 0.93, F1-score 0.92 results in a model that can be used. Of the 6 trials, experiment A – CNN from Scratch experienced underfitting, experiment B – CNN from Scratch and experiment C – CNN from Scratch experienced overfitting. So that the models that can be used for the detection process are the D – CNN add model VGG16 experiment,xiii and the E – CNN add Xception experiment for large-scale architecture. As for the application of the model on mobile devices, a model has been proposed in the F – CNN experiment add the NASNetMobile model to adjust the capabilities of the device.

Tingginya permintaan sawi hijau pada komoditas ekspor utama Indonesia harus diringi dengan proses penanaman yang baik. Namun banyak petani yang belum menerapkan pengendalian yang tepat terhadap hama salah satunya pada petani di kebon
raya dempo. Adapun kendala yang dihadapi yaitu petani saat ini dapat melakukan pemberian pestisida ketika tanaman sawi hijau telah berlubang akibat dimakan hama. Pengendalian ini dapat menjadi langkah yang baik untuk memaksimalkan hasil pertanian sawi hijau. Dimotivasi oleh keberhasilan CNN dalam klasifikasi citra, pendekatan berbasis pembelajaran mendalam telah dilakukan pada penelitian ini untuk mendeteksi
keberadaan hama di sawi hijau. Hasil eksperimen menunjukkan perbedaan akurasi disetiap percobaan dengan dataset sebanyak 1000 citra sawi hijau, yang terdiri dari 500 citra dengan hama, dan 500 citra tanpa hama. Adapun akurasi dari percobaan A – CNN from Scratch sebesar 48,33%, precision 1, recall 0.48, F1-score 0.65 hasil mengahasilkan model yang underfitting, percobaan B – CNN from Scratch sebesar 73,00% precision 1, recall 0.64, F1-score 0.78 hasil mengahasilkan model yang overfitting, percobaan C– CNN from Scratch sebesar 92,00% precision 0.88, recall 0.96, F1-score 0.92 hasil mengahasilkan model yang overfitting, percobaan D – CNN add model VGG16 sebesar 95,00%, precision 0.91, recall 0.98, F1-score 0.95 hasil mengahasilkan model yang dapat
digunakan, percobaan E – CNN add model Xception sebesar 97,00%, precision 0.96, recall 0.98, F1-score 0.97 hasil mengahasilkan model yang dapat digunakan dan percobaan F – CNN add model NASNetMobile sebesar 93,00%, precision 0.91, recall 0.93, F1-score 0.92 hasil mengahasilkan model yang dapat digunakan. Dari 6 percobaan percobaan A – CNN from Scratch mengalami underfitting, percobaan B – CNN from
Scratch dan percobaan C – CNN from Scratch mengalami overfitting. Sehingga modelxi yang dapat digunakan untuk proses deteksi adalah percobaan D – CNN add model
VGG16, dan percobaan E – CNN add model Xception untuk arsitektur skala besar. Sedangkan untuk penerapan model di perangkat seluler maka telah diusulkan model pada percobaan F – CNN add model NASNetMobile untuk menyesuaikan kapabiltas dari perangkat