In order to achieve accurate, efficient, and low-cost online monitoring of the temperature throughout the urban underground utility tunnels, and to address the technical bottlenecks faced by distributed optical fiber temperature sensing (RDTS), such as low signal-to-noise ratio, limited spatial resolution, insensitivity to small-scale anomalies, high data storage costs and inability to conduct two-dimensional monitoring, a set of methods for improving RDTS performance is proposed, including a signal denoising method using GraphSAGE graph neural network, a spatial resolution improvement method combining total variation deconvolution and fully connected neural networks, small-scale anomaly detection method utilizing attention mechanism and K-Means clustering detection, data compression method through implicit neural representation, and finally, based on the processed one-dimensional temperature signal, a two-dimensional temperature field construction method through the special wiring strategy. This set of methods can effectively improve the performance of RDTS in measurement accuracy, anomaly detection sensitivity and coverage dimension, providing a high-precision, small-scale and low-cost solution for temperature monitoring throughout utility tunnel.