Ground Object Recognition for Sentinel-1 Imagery Based on Multi-Temporal and Multi-Feature Data Fusion

The change of rice growth on cultivated land makes it difficult to recognize other ground objects, temporal SAR image is important data for ground object recognition. Aiming at the problem of false recognition of ground objects in SAR images at different time phases, this paper proposes a multi-feature data fusion method based on multi-temporal SAR images. Firstly, by selecting, judging, and fusing multi-features of multi-temporal SAR image, spatial features expression of multi-temporal SAR image is realized, then, judging and selecting the features within and between groups of different time-series images, the representative temporal features that can distinguish the growing ground objects are obtained statistically. Finally, the determining the weight methods are used to compare the recognition of ground objects under multi-temporal and multi-feature combination, the combination of feature weights that can reflect the separation of features in different time series is obtained. The SVM, KNN classifier, and LSTM network were compared for ground object recognition under multi-temporal and multi-feature weight confirmation data fusion (DF) methods and ground object recognition under different features. The experimental results show that the overall accuracy of ground objects is improved by 11.99% on average under the three classifiers, indicating that the data fusion method has better robustness. Compared with single feature (SM) and combined feature (CM) methods, data fusion (DF) method has remarkable performance in ground object recognition, especially for rice classification accuracy. In addition, combining the data fusion (DF) method with LSTM (DF-LSTM) can significantly improve the accuracy of ground object recognition in multi-temporal and multi-feature SAR data.