Remote sensing images captured by UAV

The experiment used a DJI Phantom drone equipped with an aerial camera (DJI FC6310S) to capture images of farmland. The images have three bands of red, green, and blue, with a spatial resolution of 0.1 ms. Different types and different regions of farmland images can increase the generalization ability of the model to a certain extent. Therefore, the raw data for the experimental study was taken from an image of size 4893 \u00d7 1807 taken in Hechi City, Guangxi and four images of size 7360 \u00d7 4912 taken in Kaifeng City, Henan Province. These remote sensing images captured by UAV have the characteristics of high clarity, high resolution, large proportion of farmland, and rich spatial distribution infor\u0002mation of farmland, which is conducive to random cropping these large-scale images and then partition datasets. Use the open-source image annotation tool LabelMe to annotate the Pascal Voc dataset format, and then synchronize the annotated image with its original image for preprocess\u0002ing. First, randomly crop the original image and its anno\u0002tated image to a size of 512 \u00d7 512. To further expand the data sample, the randomly cropped images were subjected to data augmentation operations such as horizontal flipping, vertical flipping, and diagonal mirroring, resulting in a data\u0002set of 3080 images. Randomly divide 3080 images into a training set (2520 images), a validation set (280 images), and a testing set (280 images) in a 9:1:1 ratio.