Daily DeepCropNet: A hierarchical deep learning approach with daily time series of vegetation indices and climatic variables for corn yield estimation

• Published in: ISPRS journal of photogrammetry and remote sensing 2024-03, Vol.209, p.249-264
• Main Authors: Xiong, Xingguo, Zhong, Renhai, Tian, Qiyu, Huang, Jingfeng, Zhu, Linchao, Yang, Yi, Lin, Tao
• Format: Article
• Language: English
• Subjects: Climate stress; Crop yield; Daily time series; Hierarchical deep learning; Remote sensing
• ISSN: 0924-2716; 1872-8235
• DOI: 10.1016/j.isprsjprs.2024.02.008

Accurate large-scale crop yield estimation under climate variability is essential to understanding the dynamics of global food security. The deep learning method has shown well performance for crop yield estimation because of its high capacity for temporal pattern recognition. However, most existing deep learning models were usually based on multi-source time series with weekly or coarse temporal resolutions, which simplified the dynamics of crop growth progress. How to effectively learn the long-term and short-term crop response from the remotely sensed and agro-meteorological time series at daily intervals remains challenging. In this study, a hierarchical deep learning model, named Daily DeepCropNet (DDCN), has been developed to process daily time series of satellite-derived vegetation indices and climatic variables for corn yield estimation. The DDCN model was a three-level pyramid structure that incorporated the corn phenology durations to extract crop growth patterns from the daily inputs. The Transformer and Long Short-Term Memory (LSTM) were implemented as a temporal filter in each level to capture the short-term crop growth patterns. The temporal patterns were hierarchically transferred from weekly to growing season aggregation to learn long-term dependency. The experiment was conducted at county level over the US Corn Belt from 2006 to 2020. LSTM, Transformer, and Random Forest (RF) models were built for comparison. The input analysis showed that the daily time series of climatic variables provided more temporal fluctuations than weekly aggregated inputs. The DDCN model provided the highest accuracy for end-of-season yield estimation under normal climate conditions in 2018–2020 (RMSE = 0.86 Mg ha−1) and extreme heat and drought conditions in 2012 (RMSE = 1.17 Mg ha−1). Further analyses indicated that the DDCN model had the potential for capturing the short-duration stresses. We also found that the DDCN model identified the silking phase as the key growth stage and achieved higher accuracy than other models for in-season yield estimation. The results showed that features extracted by the DDCN model had higher correlation coefficients with corn yield for the upper levels. By evaluating the contributions of different variables, we found that vegetation indices yield higher impact than climatic variables for yield estimation under both normal and stressful climate conditions. This study provided a promising approach to extracting the temporal pattern