In Ecuador, bananas are one of the main export products and key to the national economy. Ensuring the phytosanitary status of plantations is essential for both large and small producers. This study proposes a classification model for detecting diseases in banana leaves early, focusing on two high-impact pathologies: Black Sigatoka and Banana Streak Virus (BSV). To address the challenge of computational processing of high-resolution digital images, a compression technique using singular value decomposition (SVD) was implemented, preserving 90% of the original information and significantly reducing computational costs. The model uses a ResNet-34 architecture, trained in the Google Collaboratory environment to ensure accessibility and reproducibility. During training, four epochs were employed, achieving a 20.56% reduction in processing time and an accuracy of 98.78% in detecting BSV-infected leaves. The model was evaluated using a confusion matrix, demonstrating its effectiveness in binary classification (healthy vs. diseased leaves) and multiclass classification (differentiation between diseases). The results show that the proposal optimizes computational performance and diagnostic accuracy, facilitating field implementation without requiring advanced agronomic knowledge. Finally, the training time of the proposed model was optimized, as was the correct classification of diseased banana leaves.