A deep learning framework under attention mechanism for wheat yield estimation using remotely sensed indices in the Guanzhong Plain, PR China

• Published in: International journal of applied earth observation and geoinformation Vol. 102; p. 102375
• Main Authors: Tian, Huiren, Wang, Pengxin, Tansey, Kevin, Han, Dong, Zhang, Jingqi, Zhang, Shuyu, Li, Hongmei
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2021; Elsevier
• Subjects: China; crop management; filling period; food security; irrigation; leaf area index; Leaf Area Index (LAI); Long short-term memory with an attention mechanism (ALSTM); meteorological data; neural networks; remote sensing; spatial data; temperature; vegetation; Vegetation Temperature Condition Index (VTCI); Winter wheat; Yield estimation; China; Vegetation Temperature Condition Index (VTCI); Leaf Area Index (LAI); Long short-term memory with an attention mechanism (ALSTM); Yield estimation; Winter wheat
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2021.102375

•Attention mechanism with LSTM (ALSTM) has been applied to improve wheat yield estimation.•ALSTM model produced more accurate crop yield estimates compared with LSTM model.•Attention mechanism facilitatesd interpretability of the LSTM network’s inner working.•ALSTM model provided stable performance at different sampling sites and years. The rapid and effective acquisition of crop yield information is critical to the stability of food markets and development and implementation of related policies. It is an important baseline observation that is used for ensuring regional and global food security. In this study, a novel deep learning framework was developed for winter wheat yield estimation using meteorological data and two remotely sensed indices, Vegetation Temperature Condition Index (VTCI) and Leaf Area Index (LAI) at the main growth stages of winter wheat in the Guanzhong Plain. The proposed deep learning model was based on Long Short-Term Memory (LSTM) neural network with an attention mechanism (ALSTM), which the main idea is to assign attention to the key parts of the input sequence that affect the target vectors so that the specific features can be accurately extracted. The ALSTM model provided an improved estimation accuracy (R2 = 0.63, MAPE = 8.20%, RMSE = 502.71 kg/ha, NRMSE = 11.15%) as compared with the LSTM (R2 = 0.55, MAPE = 13.46%, RMSE = 699.92 kg/ha, NRMSE = 15.52%). A validation based on leave-one-year-out-validation further substantiated the robustness of ALSTM with smaller values of NRMSE and MAPE (13.63% and 11.54%). We demonstrated that the ALSTM model provided good generalization ability for sampling sites under different farming systems, including irrigation and rain-fed sampling sites. In addition, we evaluated the relative importance of each input variable in determining yields based on stepwise sensitivity analysis. It was found that LAI at the heading-filling stage and the milk stage as well as VTCI at the jointing stage contributed more than other input feature variables towards the corresponding yield. In conclusion, our findings highlighted that the attention mechanism helped to improve the interpretability of neural networks and the ALSTM model along with remotely sensed biophysical indices can provide a reliable and robust estimation of crop yield. An accurate estimation of wheat yield is not only helping towards informed crop management decisions but it will improve efficiency and sustainability of farming operations.