Estimating monthly gross value of industrial outputs at the pixel level with nighttime lights image products and a spatial-analysis-based CNN model

Understanding high spatiotemporal resolution monthly industrial production is crucial for detailed economic analysis and effective policy intervention. However, the lack of high-resolution spatial data and validation samples at the regional scales makes this estimation challenging. To address this, we propose a Spatial Analysis-based Convolutional Neural Network (SA-CNN) model that estimates the grid-level Gross Value of Industrial Output (GVIO) in Shanghai's monthly time series by accounting for heterogeneous nighttime light (NTL) patterns across different land uses. Firstly, we locate each factory based on the corresponding category in Point of Interest (POI) data and propose a new sampling strategy using buffer zones around each factory to capture local NTL patterns. Secondly, the SA-CNN extracts feature from monthly NTL patterns, incorporating urban and rural NTL statistical characteristics to address the few-shot problem. Finally, we map the monthly estimated GVIO grid data from 2014 to 2021 by establishing a linear correlation between NTL in factory areas and the estimated results. Experimental results indicate that SA-CNN outperforms the random forest and baseline models, with R² values of 0.94 for estimation results and 0.90 for the mapping grid. The spatial distribution of monthly GVIO in northern Shanghai was more balanced and developed faster than in the south and the overall GVIO increased oscillation. The SA-CNN method proposed a new sampling strategy to overcome the problem of the need for large samples in estimating regional industrial activities and offers a fast update and labor-free methodology for monthly GVIO estimation.