Interleaved group convolutions for multitemporal multisensor crop classification

Abstract The multitemporal observation capability of remote sensing satellites is of increasing importance in crop monitoring. Nowadays we can obtain dense multitemporal data more easily because more and more Earth observation satellites are put into use. Multiband images derived from different satellite sensors usually have different spatial resolutions, and how to utilize different spatial resolution information of multisensors in multitemporal sequence is of great significance. In this paper, we propose a crop classification network based on interleaved group convolution for multitemporal multisensor data. Firstly, we selected the Central Valley of California as research area with rich crop categories. And multitemporal sequence of length 65 in this region was obtained from satellites Sentinel-2A/B and Landsat-8. Then, to coordinate the sensors with different spatial-spectral acquisition parameters, we convolve different bands of sensors to perform multiresolution fusion, and then further align the spectral dimensions. The interleaved group convolution, which has comparable model performance and fewer parameters and lower computational complexity than regular convolution, is introduced to reduce network parameters and reduce the computational complexity of multitemporal classification. Experimental results on our proposed multitemporal multisensor dataset have proved that our method has better performance in crop classification compared with traditional methods. And compared with regular convolution, our method can achieve roughly equivalent classification accuracy.