利用遥感影像时间序列进行自动化智能解译农作物精细类型在农业资源调查、监管和规划等领域有着重要的作用。目前已有的深度学习方法通过卷积或循环网络获取遥感时序中局部的时序、空间信息,缺乏对遥感影像时间序列中时空信息的充分利用,导致分类精度不高。自注意力机制是一种能够通过获取全局特征来充分挖掘数据信息的方法,本文提出了一种多尺度时空全局注意力模型MSSTGAM(Multi-Scale Spatial-Temporal Global-Attention Model),采用空间自注意力机制和时序自注意力机制相结合以构建多尺度的时空全局注意力充分挖掘遥感影像时间序列中的信息用于农作物精细分类。在公开数据集PASTIS和自制Mississippi数据集上进行了检验和评估,实验结果表明本文提出的MSSTGAM能够有效地进行遥感影像时间序列的农作物分类,与其他方法相比定量分类精度最优,地块内的可视化结果在空间一致性上更好。本文的研究表明多尺度时空全局注意力对遥感影像时间序列的农作物精细分类更有效,具有重要的理论和应用价值。

Automatic intelligent interpretation of the fine types of crops by utilizing remote sensing image time series plays an important role in the fields of agricultural resource investigation, supervision and planning. The existing deep learning methods extract local spatial or local temporal information by convolutional or recurrent neural networks which have inadequate utilization of spatial-temporal information, resulting in low classification accuracy. The self-attention mechanism is able to fully exploit data information by obtaining global attention. Thus, we propose a multi-scale spatial-temporal global attention model (MSSTGAM), which combines spatial self-attention mechanism and temporal self-attention mechanism to construct multi-scale spatial-temporal global attention, and fully mine the information of remote sensing image time series for the fine classification of crop types. The proposed method is evaluated and tested on the publicly available dataset PASTIS and custom Mississippi dataset. The results demonstrate that MSSTGAM is capable of identifying crop types from remote sensing image time series, and achieves the best quantitative result compared with other methods. Moreover, the visualization result of MSSTGAM has better inner-parcel spatial consistency. This paper’s findings show that multi-scale spatial-temporal global attention has significant theoretical and practical significance and is more effective for the fine classification of crop type from remote sensing image time series.