Winter wheat is one of the important food crops in China, and its planting area and output are the second only to rice. In order to improve the accuracy of remote sensing estimation of winter wheat leaf area index (LAI) in arid regions, taking the LAI of winter wheat at the jointing stage as research object, based on the first derivative (FD) and second derivative (SD) differential preprocessing of the canopy hyperspectral data, the two-dimensional vegetation index (2DVI) and three-dimensional vegetation index (3DVI) of any band combination was calculated, and the correlation with LAI was carried out. To find the vegetation index of the best band combination;the artificial neural network (ANN), K-nearest neighbors (KNN) and support vector regression (SVR) were used to establish LAI estimation respectively model and verify the accuracy. The results showed that the correlation between vegetation index and LAI in any combination of wavelength bands was significantly improved, especially the correlation coefficient of FD-3DVI-4(714nm, 400nm, 1001nm) based on the FD preprocessing spectrum reached 0.93 (P<0.01), and the optimal combination band was mainly located in the red edge position. The estimation model based on the optimal FD-3DVI index and K-nearest neighbors algorithm performed outstanding, its R2=0.89, the root mean square error (RMSE) was the lowest (0.31), and the relative analysis error (RPD) was 2.41. It was conclused that the K-nearest neighbor algorithm was more suitable for solving the nonlinear problem and improve the estimation accuracy, and it can provide a theoretical basis for the later crop growth evaluation and reasonable fertilization.