Multi-scale Residual Spatial-Spectral Attention Longan-Litchi Extraction Network for Cultivated Land Non-grain Monitoring

With the continuous development of remote sensing earth observation technology, the acquisition of remote sensing data has become more convenient, and it has been widely used in fields such as agricultural resource investigation, disaster monitoring, and agricultural information management. Hyperspectral imagery has the characteristics of “unity of maps and spectra”, wide wavelength range, large number of bands, complete spectral curves and rich information, which can provide more detailed spectral information for earth observation. It is also of great significance to the development of precision agriculture. Quickly and accurately sensing and predicting changes in land types within agricultural land, cultivated land planting conditions and changes, and improving the ability to dynamically monitor cultivated land protection are the basis for ensuring food security. Planting commercial crops on cultivated land is one of the cases of non-grain conversion of cultivated land, and it is also a problem that needs to be monitored. The extraction of economic crops by means of remote sensing is of great significance to the monitoring of non-grain conversion of cultivated land. At present, there are problems of misclassification between cultivated land and garden land. The identification accuracy of crops and economic crops is not high, which directly affects the effectiveness of refined management of cultivated land protection. As a common economic crop, litchi and longan are widely planted on cultivated land, resulting in the non-grain phenomenon of cultivated land. However, there are few related studies on the extraction of litchi and longan. This paper chooses litchi and longan as the extraction object to study the above problems. In this paper, we propose a multi-scale residual spatial-spectral attention longan-lychee extraction network for hyperspectral imagery. The network first uses the idea of feature pyramids to extract image features under different receptive fields by using convolution kernels of different sizes. Secondly, the fused multi-scale features are screened through a spatial-spectral enhanced attention module, highlighting the spectral features with strong representation and high discrimination that are effective for classification results. Finally, a residual structure is introduced to optimize model training and convergence. Guangdong is one of the most important production areas of litchi and longan in China, ranking first in the country in both area and output. In this paper, Maoming City, Guangdong Province is used as the experimental area for research, and the hyperspectral image of Zhuhai-1 satellite is used as the experimental data. The experimental results show that the method in this paper has certain advantages in the extraction of litchi and longan, which can improve the efficiency and accuracy of longan and litchi extraction, and provide important data and technical support for the monitoring of cultivated land non-grain.