Accurately knowing the spatiotemporal distribution of indoor environmental parameters is important for indoor thermal comfort adjusting and building energy saving. Due to the significant distributed parameter characteristics of the building thermal environment, this task is difficult, especially for large-space buildings with the features of multi-input and multi-output. For a real large-space building, the reduced-dimensional model of indoor temperature field is rebuilt based on proper orthogonal decomposition (POD) and an offline-online strategy. POD technology is used to extract environmental features from steady / dynamic experimental data. Based on the solved POD modes, the optimal layout of temperature sensors is obtained by a heuristic reasoning method. By combining linear stochastic estimation (LSE) method together with the selected sensors, an offline-online strategy is proposed for rapid reconstruction and prediction of the indoor temperature fields. Digital interpolation is applied for soft sensing outside the original sampling locations. An experimental study for a large-space cafeteria with sixteen indoor air conditioning units is conducted. A total of 72 sensors are placed to collect indoor 3D temperature field data offline. Results show that using only six sensors of them, the large-space indoor temperature field could be accurately estimated online with the steady-state error of 0.35 (average RMSE) and dynamic-state errors of 0.0525 / 0.0540 (heating / cooling, average RMSE).