In order to improve the accuracy of the land cover classification model based on near-infrared spectroscopy, the soil near-infrared spectrum data released by Eurostat was adopted as the research object in this study, and the land cover classification model based on short-time Fourier transform (STFT) preprocessing method and different convolution scale fusion are studied to achieve the rapid identification of cultivated land, forest land and grassland. In order to meet the requirements of two-dimensional convolution, 4200 wavelength points in the 400~2 500 nm band of the one-dimensional spectrum were transformed into two-dimensional images by SFTF, so the spectrum information about spectral data was extractedfor the later modeling. The samples were randomly divided into the training set, validation set and test set according to the ratio of 6∶2∶2. Two kinds of convolution neural network(CNN) models, including the single-size kernel CNN and themulti-size kernel fusion CNN, were established to classify the land cover. The CNN models adopted the ReLU activation function, batch normalization (BN), and dropout methods to prevent the model’s gradient disappearance. The Early Stopping method trains the network to prevent the model from overfitting. Firstly, this paper discussed the influence of different STFT window lengths (64, 100 and 128) and convolution kernel sizes (3×3, 5×5 and 7×7) on the classification effect of the different models. The experimental results show that the overall classification accuracy of all model shit a high point when the STFT window length was 100 and the window overlap length was 50%. The model’s classification accuracy decreases with the convolution kernel size. That is, the accuracy of the model with a smaller convolution kernel size is relatively higher. The overall classification accuracy of the CNN model with a convolution kernel size of 3×3 reaches 78.76%, which is higher than that of the CNN model with a convolution kernel size of 5×5 and 7×7. The CNN models with different convolution kernel sizes have good classification results for a certain land cover type.That is, the model with 3×3 convolution size has the best classification effect for cultivated land, 5×5 convolution size has the best classification effect for woodland, 7×7 convolution size has the best classification effect for grassland. Secondly, a fusion-CNN model based on the hybrid of multi-size convolution kernels is proposed. The model integrates the advantages of different size convolution kernels. The classification accuracy of the fusion-CNN for three land cover types has been improved to varying degrees, with an overall classification accuracy of 84.39%. The Fusion-CNN model overcomes the shortcomings of the single size convolution kernel CNN model, such as a long selection period for appropriate convolution kernel size and cumbersome parameter adjustment steps, and can simplify and speed up the modeling process. Fusion-CNN convolution fusion network can more effectively extract the internal characteristic information on soil near-infrared spectroscopy to obtain high and stable land cover classification accuracy.