Precise identification of levee relief well locations (or other similar infrastructure components) from satellite imagery can be challenging and time-consuming. In this research, we propose a novel approach utilizing a U-Net architecture to identify relief wells along flood control levees, as a proof of concept of combining freely-available imagery with machine learning techniques. The goal of this research is it develop techniques to make infrastructure asset management more efficient by automating the process, which could be extended to identify, inventory, or track other surficially identifiable objects relevant for different infrastructure systems. Our study highlights the crucial role of the convolution kernel size in the U-Net architecture, which significantly influences the accuracy of the results. Larger convolution kernel sizes, such as 11x11, excel in capturing extensive contextual information from the input image, potentially leading to superior outcomes. However, training with larger kernels is computationally intensive. Conversely, smaller convolution kernel sizes, like 3x3, excel in capturing local features. To strike a balance these considerations, we introduce a Dual Kernel Filter U-Net, which combines two U-Nets with distinct convolution kernel sizes, 3x3 and 11x11. This innovative approach aims to harness the strengths of both convolution kernel sizes to improve accuracy and overall performance. The proposed Dual kernel filter U-Net was trained and evaluated on a real dataset of relief wells from Google Earth imagery. Our evaluation results demonstrate that this model achieves an accuracy rate of 99.76% with training dataset 2626 satellite images. Notably, this significant accuracy enhancement is achieved without substantially increasing computational time, making it a promising advancement in satellite image analysis for object location identification and asset and disaster management.