Forecasting provincial agricultural output value in China via multiple nighttime light indices and neural networks

It is crucial for comprehending the developmental trend of the agricultural economy, refining the agricultural industrial structure, and amending agricultural policies to accurately forecast and timely obtain the total output value of agriculture, forestry, animal husbandry, and fishery (TOVAFAF). This paper attempts to utilize eight nighttime light (NTL) indices originally constructed based on NPP-VIIRS NTL remote sensing data serving as multiple input variables to establish a more accurate and effective forecasting model for the TOVAFAF in various provinces of China. For the major challenge of characterizing the complex nonlinear relationship between NTL data and the TOVAFAF under the condition of limited samples, this paper employed single-hidden-layer back propagation (BP) neural network and extreme learning machine (ELM) as two basic modeling methods, proposed a novel ensemble particle swarm optimization (EPSO) algorithm as optimization mechanism for neural network models to overcome the limitation of traditional particle swarm optimization (PSO) algorithm, and used logarithmic transformation to enhance the correlation between NTL data and the TOVAFAF. The experimental results further substantiate that the neural network algorithms can effectively characterize the potential nonlinear relationship between NTL data and the TOVAFAF. And the neural network models optimized by the EPSO mechanism, those under logarithmic transformation, and the BP neural network series models exhibit superior forecasting performance than those optimized by the PSO mechanism, those under normalization, and the ELM series models, respectively. Furthermore, the EPSO-BP model under logarithmic transformation provides the best forecasting performance on the TOVAFAF for China's various provinces in 2023, with the mean relative error (MRE) of 20.65% and the determination coefficient (R²) of 0.8749 for the linear relationship between the actual and forecasting values, which presents a decrease of 11.55 percentage points in MRE and an increase of 35.38% in R² compared to the PSO-ELM model in our previous research.