Abstract: Abstract: Accurate land use classification is highly required using Unmanned Aerial Vehicle (UAV) images, especially the data selection. In this study, the UAV orthographic remote sensing images were acquired at different aerial heights in Tongji Village, Baishui County, Weibei dry land, China. The land use was then classified using a variety of deep learning and machine learning. The DJI Mavic 2Pro was used to obtain 80 and 160m aerial images in the study area. There were 96 routes, the total length of routes was 42.43 km, the heading overlap degree was 75%, the side overlap degree was 60%, and a total of 2 248 original aerial photos were taken at a flight height of 80 m. At 160 m flight height, there were 20 routes with a total length of 17.90 km, the heading overlap degree was 70%, the side overlap degree was 55%, and a total of 502 original aerial images were taken in this case. The geo-positioning of the photo control points was performed on the Zhuolin A8 handheld Beidou GPS locator. Agisoft PhotoScan 1.4.5 software was used to splice and process the original single-image data. A comparison was made on the visual interpretation of different aerial photography heights and the prediction of various deep learning and machine learning models. Labelme4.5.6 software was used for the visual interpretation. As such, the best performance was achieved during this time. The results show that the performance of deep learning was far better than that of traditional machine learning. The best-performing of deep learning (DeepLabv3+) presented a pixel accuracy of 90.06%, which was 24.65, and 21.32 percentage points higher than that of random forest (RF) and support vector machine (SVM), respectively. The improved DeepLabv3+_BA model performed the best overall classification. The improvement of deep learning was attributed to two aspects. Firstly, the BN layer was removed after the first two separate convolution layers in the Entry flow in the encoder Xception part of the original DeepLabv3+ model. The BN layer was removed in ASPP after the last three separate convolutional layers in the Exit flow. The BN layer was removed after each dilated convolutional layer. Secondly, the ASPP atrous rate combination design was re-optimized, according to the characteristics of the data set. The pixel accuracy of the improved model was 91.37%, which was 7.43, 10.12, 2.27, and 1.31 percentage points higher than those of FCN, SegNet, UNet, and DeepLabv3+, respectively. The number of iterations required for the best accuracy was reduced by about 50%, compared with the other four deep-learning models. Taking the extraction of apple orchard as an example, the F1 value of DeepLabv3+_BA was 89.10%, which was 19.94, 23.68, 2.04, 2.97, 2.4, and 0.78 percentage points higher than those of SVM, RF, FCN, SegNet, UNet, and DeepLabv3+, respectively. The accuracy of various algorithms was higher than 80 m on 160 m datasets. The performance of various deep learning on the test set demonstrated that the accuracy of DeepLabv3+_BA reached more than 90% for the apple orchard, bare field, stubble field, and road ground object classification. The improved model DeepLabv3+_BA presented higher accuracy and robustness of ground object classification. This finding can also provide a strong reference for the land use information census using UAV images and deep learning.