DeepCropNet: a deep spatial-temporal learning framework for county-level corn yield estimation

Abstract

Large-scale crop yield estimation is critical for understanding the dynamics of global food security. Understanding and quantifying the temporal cumulative effect of crop growth and spatial variances across different regions remains challenging for large-scale crop yield estimation. In this study, a deep spatial-temporal learning framework, named DeepCropNet (DCN), has been developed to hierarchically capture the features for county-level corn yield estimation. The temporal features are learned by an attention-based long short-term memory network and the spatial features are learned by the multi-task learning (MTL) output layers. The DCN model has been applied to quantify the relationship between meteorological factors and the county-level corn yield in the US Corn Belt from 1981 to 2016. Three meteorological factors, including growing degree days, killing degree days, and precipitation, are used as time-series inputs. The results show that DCN provides an improved estimation accuracy (RMSE=0.82 Mg ha −1 ) as compared to that of conventional methods such as LASSO (RMSE=1.14 Mg ha −1 ) and Random Forest (RMSE=1.05 Mg ha −1 ). Temporally, the attention values computed from the temporal learning module indicate that DCN captures the temporal cumulative effect and this temporal pattern is consistent across all states. Spatially, the spatial learning module improves the estimation accuracy based on the regional specific features captured by the MTL mechanism. The study highlights that the DCN model provides a promising spatial-temporal learning framework for corn yield estimation under changing meteorological conditions across large spatial regions.

Crop yield prediction integrating genotype and weather variables using deep learning

“…The advantages of LSTM based models have been recently established for maize yield prediction at a county level [ 27 ], but the model lacked interpretability. Attention based LSTM along with multi-task learning (MTL) output layers has also been used for maize yield prediction using county level data based on meteorological data (maximum daily temperature, minimum daily temperature, and daily precipitation) [ 45 ]. These studies are important for solving the yield prediction challenge; however, models are based on geospatial data without field-scale farming management data and variety information is indiscernible, and based on limited weather variables.…”

“…Attention based LSTM has been used along with multi-task learning (MTL) output layers [ 45 ] for county level corn yield anomaly prediction only based on meteorological data (maximum daily temperature, minimum daily temperature) without field-scale farming data. Previous work [ 46 ] using deep learning for yield prediction has utilized multi-spectral images to predict yield (instead of leveraging only multivariate time series as input) without considering model interpretability.…”

A high-resolution daily gridded meteorological dataset for Serbia made by Random Forest Spatial Interpolation

“…Daily gridded meteorological datasets are essential input for numerous models and analyses across various research fields. For example, daily meteorological gridded datasets are used in agriculture for estimating yield 5 , 6 , the occurrence of insect pests and disease 7 , and crop growth 8 , as well as in meteorology 9 , hydrology 10 , ecology 11 , climate and climate change analysis 12 , risk assessment 13 , and forestry 14 .…”

Winsorization for Robust Bayesian Neural Networks

“…Machine learning methods present the opportunity to model agricultural data using more complex architectures, using fewer assumptions, and on larger datasets. Artificial neural networks [81][82][83][84][85][86][87][88][89], linear regression [87,88,[90][91][92][93], tree-based models [87,[92][93][94][95][96][97][98], and support vector machines [98][99][100] are some of the most used machine learning algorithms [101,102] for crop yield modeling. In particular, ANNs have been used for tasks such as species recognition, weed detection, or crop quality assessment in [81][82][83][84][85][86][87][88][89] and elsewhere, using a variety of complex features including satellite data.…”