Multielement Classification of a Small Fragmented Planting Farm Using Hyperspectral Unmanned Aerial Vehicle Image

Aiming at identifying cropland in the Yangtze River Delta, we used unmanned aerial vehicle (UAV) to obtain high spatial and spectral resolution (HSSR) remote sensing images of a small farm in the southern Jiangsu Province. After feature augmentation and compression, we used two 3-D convolutional neural network (3-D CNN) algorithms and the baseline neural network (Baseline-NN) algorithm to classify the UAV-HSSR images. The classification results showed that these three classification methods could achieve the fine-scale classification of all elements in the study area, with an overall accuracy of 86.560%, 85.416%, and 94.926%, and Kappa coefficients of 0.846, 0.833, and 0.936, respectively. The findings of this study indicate that hyperspectral UAV images have significant potential in the classification tasks of highly fragmented small farms, although the salt and pepper phenomenon was observed in the results of three classification methods.