Farmland Recognition of High Resolution Multispectral Remote Sensing Imagery using Deep Learning Semantic Segmentation Method

Farmland mapping is an important step for estimating grain yields. However extraction of farmland from multispectral remote sensing images (RSIs) is still a challenging work, as farmland is located on not only plains but also mountains, which displays divergent and confusing characteristics in RSIs. To solve the problem of lacking the multispectral remote sensing image dataset for pretraining, we extend Deep Feature Aggregation Net (DFANet) with fewer network parameters, a semantic segmentation network, to automatically map farmland from 3-band to multispectral images in a pixel-wise strategy. In this network, we first utilize more information aggregation. The fully-connected attention module is then replaced by a proposed convolution attention module. Finally, a new proposed decoder is used to recover the details of the feature map. Experimental results show that the model with multispectral RSIs outperforms the baselines.