Abstract: Accurate and rapid estimation of soil total nitrogen (TN) content in reclaimed farmland is the guarantee of land reclamation quality evaluation. Soil samples from reclaimed farmland of Yongcheng mining area were chemically treated and the hyperspectral data were measured indoors. Soil hyperspectral data were transformed by three mathematical methods and then correlated with TN contents, and then the sensitive bands were determined. Subsequently, the partial least squares regression (PLSR) models with BP neural networks (BPNN) and random forest (RF) were combined to establish PLSR-BPNN and PLSR-RF models of inversion of soil TN content based on hyperspectral data. The newly established models were compared with the traditional PLSR, BPNN and RF models. The results showed that the accuracy of the synthetic models (PLSR-BPNN and PLSR-RF) was significantly improved compared to the single model algorithm. In particular, the accuracy of the spectral data processed by the first-order differential method PLSR-BPNN was the highest, with a decision coefficient (R²) of 0.92 and a relative analysis error (RPD) of 4.01. Therefore, the PLS-BPNN model based on the first-order differential spectrum was the best one among the estimation models of soil TN content. The results provide reference for the retrieval of soil TN based on hyperspectral data in reclaimed farmland.

Key words: poplar plantation, furrow irrigation, productivity, drip irrigation, soil moisture sensor, soil moisture content