Abstract: The FY-3B satellite has the characteristics of a high frequency of observation and wide imaging range, which can provide long-term observation data for maize leaf area index (LAI) inversion research. Long short-term memory (LSTM) algorithm has the ability to extract temporal features from multi-period data and solve complex nonlinear problems between spectral data and LAI. The study was conducted based on   LAI and reflectance spectrum data of spring maize in Jinzhou City,Liaoning Province  measured  near the ground. To construct the LAI inversion model, the spectral response functions were used to simulate the FY-3B multi-spectral band data combined with 28 vegetation indices highly correlated with spring maize LAI. The inversion models were conducted using LSTM of different hidden layers, and the accuracies of the LSTM models were compared with the accuracy of the partial least-squares regression (PLSR) model. The results showed that the number of hidden layers greatly influences the fitting ability of the LSTM model. The three-layer LSTM model increased the LAI estimation accuracy R2 from 0.818 3 (single-layer LSTM), 0.780 0 (PLSR) to 0.869 2; correspondingly reducing the RMSE from 0.509 1,0.490 6 to 0.372 6. In short, the accuracy of the model was significantly improved.

Key words: spring maize, leaf area index, FY-3B, LSTM