Road extraction plays a crucial role in various sectors, including transportation systems, disaster relief distribution, and urban planning, necessitating a more efficient method than the current labor-intensive approach. The conventional process, reliant on operator efforts, proves costly, time-consuming, and energy intensive. To address this, the employment of deep learning models such as DeepLab V3+ are done, leveraging encoders like ResNet 50, ResNet 101, and MobileNet V2, known for their effectiveness in deep learning applications. Although Deep Learning has demonstrated faster and more automated road extraction, the bottleneck persists in the manual creation of training data through road annotations by operators. Our research focuses on accelerating road extraction by incorporating enhanced annotations from Object-Based Image Analysis (OBIA) along with organic annotations into the training data to expedite dataset creation while ensuring extraction model accuracy. Specifically, we investigate the optimal ratio of synthetic to organic annotations that yields the highest road extraction accuracy. Moreover, we enhance OBIA-derived road annotations and regulate their integration into the training data. Our findings reveal an optimal composition of 25% for OBIA annotations and 50% for improved OBIA annotations, as exceeding these numbers results in diminished model performance. Significantly, the further improvement of OBIA annotations substantially boosts model performance metrics exemplified in the use of 100% composition in training data. On average, each model produces Pixel Accuracy of 0.942, IoUr of 0.012, mean IoU of 0.477, and Dice Score of 0.495 for every use of 100% OBIA annotations in the training data. Improvement in model performance evaluation metrics occurs when using 100% improved OBIA annotations in the training data where on average each model produces Pixel Accuracy values of 0.954, IoUr of 0.433, mean IoU of 0.692, and Dice Score of 0.771. The experimental results demonstrate the advantages of our proposed method, indicating a reduction in the time required to prepare Deep Learning datasets, reducing the number of organic annotations required to as little as 50% while maintaining model performance by leveraging OBIA road annotations as training data.