The use of ALSTM-FCN for tobacco planting extraction from time-series Sentinel-1A Sar images

Spatial information on tobacco planting is crucial to many agricultural applications regarding tobacco production and management. This paper presents a deep learning model, i.e., Attention Long Short-Term Memory Fully Convolutional Network (ALSTM-FCN), to extract tobacco planting areas using time-series Sentinel-1A (S1A) SAR images. Using the ALSTM-FCN model, high-level temporal and spatial image features are fused to characterize the growth of tobacco planting. We applied the ALSTM-FCN to extract tobacco in the Fujian area using time-series S1A SAR data acquired in 2020. We compared the proposed method with a conventional LSTM and a machine learning method (e.g., Light GBM). Our results show that the extracted results by the ALSTM-FCN model have a higher extraction accuracy of 0.93 than that of the LSTM of 0.92 and the Light GBM of 0.91. We conclude that the proposed ALSTM-FCN method can be used as a promising solution for extracting tobacco using time-series SAR data in cloudy and rainy areas.

I. Introduction

Tobacco is a widely planted economic crop. Spatial information on tobacco planting is vital to many agricultural tobacco production and management applications. For farmers, growing tobacco can obtain a relatively higher income than main crops. For policymakers, however, it is important to control the extent of tobacco planting areas to give more land to main crops to guarantee food security. Timely and accurate tobacco mapping can provide much spatial information.