AUTHOR=Hu Yimin , Meng Ao , Wu Yanjun , Zou Le , Jin Zhou , Xu Taosheng 

TITLE=Deep-agriNet: a lightweight attention-based encoder-decoder framework for crop identification using multispectral images

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 14 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2023.1124939

DOI=10.3389/fpls.2023.1124939

ISSN=1664-462X

ABSTRACT=Computer vision technology based on multispectral images has shown a broad prospect in the identification of large-scale crops. However, existing crop identification networks often struggle to balance accuracy with a lightweight design, and there is a lack of accurate recognition methods for non-large-scale crops. This paper proposed an improved encoder-decoder framework based on DeepLab v3+ to accurately identify crops with different planting patterns. The network used ShuffleNet v2 as the backbone to extract features at different levels. The decoder module incorporates a convolutional block attention module, which combines the channel attention mechanism and spatial attention mechanism to fuse attention features across the channel and spatial dimensions. In this paper, two datasets DS1 and DS2 were established, where DS1 was collected from the area with large-scale planting, and DS2 was collected from the area with scattered planting. On DS1, the improved network was 0.972, 0.981 and 0.980 in mean intersection over union(mIoU), overall accuracy(OA) and recall, respectively. This represented an improvement of 7.0, 5.0, and 5.7\% over the original DeepLab v3+. On DS2, the improved network enhanced mIoU, OA, and recall by 5.4, 3.9, and 4.4\%, respectively. Meanwhile, the parameters and giga floating-point operations(GFLOPs) are much smaller than DeepLab v3+ and other classic networks. The results show that the Deep-agriNet performs better for crops with different planting scales, and can be an effective tool for crop identification in various countries and regions.