A graph-based deep learning framework for field scale wheat yield estimation

Dong Han

doi:10.1016/j.jag.2024.103834

2024-05-01

DOI：10.1016/j.jag.2024.103834

原文链接

查看原文

摘要

En 中文

Accurate estimation of crop yield at the field scale plays a pivotal role in optimizing agricultural production and food security. Conventional studies have mainly focused on employing data-driven models for crop yield estimation at the regional scale, while large challenges may occur when attempting to apply these methods at the field scale. This is primarily due to the inherent complexity of obtaining reliable ground labels of yield for field validation, and the geographical independence and correlation that exists between fields. To effectively solve this problem, this study couples geographical, crop physiological knowledge and deep learning networks, and builds a graph-based deep learning framework by integrating high-medium spatial resolution active and passive remote sensing data (Sentinel-1, Sentinel-2 and Sentinel-3) and uses it to estimate field scale winter wheat yield. Firstly, a deep learning framework based on graph theory was constructed to achieve accurate estimation of field scale time series winter wheat growth parameter (Leaf Area Index, LAI), and then the growth mechanism of winter wheat and the specific factors affecting wheat yield formation were further considered, so as to improve the yield estimation accuracy of the traditional data-driven yield estimation model. Finally, the yield estimates of the proposed method were compared and analyzed for farmlands under different categories of agricultural disasters. The results showed that the graph-based two-branch network architecture (the Seq_Gra_Gd model) with the optimal meteorological data input strategy (meteorological data of the previous 15 d) had the optimal LAI estimation accuracy, and except for the jointing stage of winter wheat, the Seq_Gra_Gd model had a high and stable LAI estimation accuracy at the other main growth stages. The Seq_Gra_Gd model achieved good accuracy in estimating winter wheat yield (R-2 = 0.73, RMSE = 590.43 kg.ha(-1)), and the introduction of the graph convolution module enabled the model to take into account the spatial distribution characteristics of stripe rust and lodging disasters well, which improved the yield estimation accuracy of affected winter wheat.

Keyword:

Active-passive remote sensing

Deep learning

Crop physiological basis

Geographical knowledge

Field scale yield estimation