Fine crop classification in high resolution remote sensing based on deep learning

Abstract

Mapping the crop type can provide a basis for extracting information on crop planting structure, and area and yield estimation. Obtaining large-scale crop-type mapping by field investigation is inefficient and expensive. Traditional classification methods have low classification accuracy due to the fragmentation and heterogeneity of crop planting. However, the deep learning algorithm has a strong feature extraction ability and can effectively identify and classify crop types. This study uses GF-1 high-resolution remote sensing images as the data source for the Shuangcheng district, Harbin city, Heilongjiang Province, China. Two spectral feature data sets are constructed through field sampling and employed for training and verification, combined with basic survey data of grain production functional areas at the plot scale. Traditional machine learning algorithms, such as random forest (RF) and support vector machine (SVM), and a popular deep learning algorithm, convolution neural network have been utilized. The results show that the fusion of multi-spectral information and vegetation index features helps improve classification accuracy. The deep learning algorithm is superior to the machine learning algorithm in both classification accuracy and classification effect. The highest classification accuracy of Crop Segmentation Network (CSNet) based on fine-tuning Resnet-50 is 91.2%, kappa coefficient is 0.882, and mean intersection over union is 0.834. The classification accuracy is 13.3% and 9.5% points higher than RF and SVM, respectively, and the best classification performance is obtained. The classification accuracy and execution efficiency of the model are suitable for a wide range of crop classification tasks and exhibit good transferability.