Melon productivity in Indonesia has experienced a significant decline due to leaf diseases, while manual detection performed by farmers remains subjective, time-consuming, and highly dependent on individual experience. To address this issue, this study aims to develop a mobile-based melon leaf disease detection system utilizing a Convolutional Neural Network (CNN) architecture integrated into the Tani Cerdas Android application via the TensorFlow framework. The dataset consists of 250 images of melon leaves categorized into five classes: healthy, aphids, fusarium wilt, leaf caterpillars, and unknown. Data were collected from two different melon farms employing distinct cultivation methods and processed through the machine learning life cycle, including data cleaning, manual labeling using one-hot encoding, splitting into 80% training and 20% validation sets, model training, and performance evaluation. The CNN model was trained for 11 epochs using ReLU and Softmax activation functions and a dropout rate of 0.2 to reduce the risk of overfitting. Training results achieved an accuracy of 91.5% with a loss value of 0.313, while model validation reached 71.9% accuracy. The ROC-AUC evaluation indicated excellent classification performance in most classes (AUC 0.99–1.00), although performance in the fusarium wilt class remained lower (AUC 0.87). Deployment of the model into the Tani Cerdas application achieved an average field accuracy of 86.33%. This study demonstrates the effectiveness of CNN and TensorFlow integration in supporting rapid and independent detection of melon leaf diseases via mobile devices, offering potential for the development of similar systems for other horticultural commodities.