Desert steppe with features of sparse vegetation and fragmented bare soil distribution, required for high spatial resolution and spectral resolution of remote sensing data. There were some problems with over calculation and time-consuming according to present situation of deep learning use for remote sensing. Firstly, multiple hidden layers with complex structure were common in remote sensing scenes application. Secondly, inherent characteristics of remote sensing data were lack of consideration when some classical models were applied directly. A low altitude unmanned aerial vehicle (UAV) platform was established with a hyperspectral remote sensing sensor on it, which gave full play to the strengths of spatial and spectral resolutions. A simplified learning classification model were proposed by using three-dimensional convolutional network (3D-CNN) in desert steppe with hyper parameters of learning rate, batch size, number and size of convolutional kernels optimized for the classification of vegetation, bared ground and indicators. The highest overall accuracy (OA) of the model was evaluated to be 99.746% after optimized. The results suggested that the optimization of simplified learning classification model should build on constantly adjusting hyper parameters and sufficiently comparing with classification results of various combinations for higher precision, shorter time-consuming and more reliable stability. These results demonstrated that the simplified learning classification model based on UAV hyperspectral remote sensing had good performance in classifying ground target in desert steppe.