Spatial mapping of topsoil total nitrogen in mountainous and hilly areas of southern China using a continuous convolution neural network

Rapid and accurate acquisition of soil nutrient information using remote sensing technology is a prerequisite for precision agriculture. However, soil information is weak in remote sensing images of mountainous and hilly areas with high vegetation coverage. Therefore, a continuous convolutional neural network (CNN)-based method was developed for spectral inversion of soil total nitrogen (STN) content in mountainous and hilly areas of southern China. First, the correlation between laboratory-measured soil spectra, remote-sensing image spectra, and STN content under different land use types was analyzed. Image spectra were then transformed to laboratorymeasured spectra by two deep learning methods: convolutional neural network (CNNT) and multilayer perceptron (MLPT). The transformed spectra were used for STN modeling based on deep learning models (CNNS and MLPS), and the modeling results were corrected for residuals using ordinary kriging interpolation. The results showed that in vegetated areas, vegetation spectral features provided indirect evidence for soil nutrient status. Consequently, image spectra were successfully transformed into laboratory-measured spectra, and CNNT outperformed MLPT for spectral transformation. All bands of transformed spectra had high correlations with STN content and could therefore be used for STN inversion. Despite the rough estimation of STN content based on transformed spectra, the modeling results of CNNS and MLPS improved after residual correction. Overall, the corrected CNNS model achieved the highest inversion accuracy in terms of the coefficient of determination (0.702), root mean squared error (0.489 g center dot kg(-1)), and ratio of performance to interquartile range (3.009), which was similar to 20% higher than before correction. The optimal model estimated STN content accurately and achieved spatial mapping of regional surface-layer STN content. The study develops a powerful tool for real-time quantitative monitoring of STN changes in mountainous and hilly areas, and provides new possibilities for precision agriculture.