Winter Wheat Yield Prediction Based on the ASTGNN Model Coupled with Multi-Source Data

Abstract

Timely and accurate prediction of winter wheat yields, which is crucial for optimizing production management, maintaining supply–demand balance, and ensuring food security, depends on interactions among numerous factors, such as climate, surface characteristics, and soil quality. Despite the extensive application of deep learning models in this field, few studies have analyzed the effect of the large-scale geospatial characteristics of neighboring regions on crop yields. Therefore, we present an attention-based spatio-temporal Graph Neural Network (ASTGNN) model coupled with geospatial characteristics and multi-source data for improved accuracy of winter wheat yield estimation. The datasets used in this study included multiple types of remote sensing, meteorological, soil, crop yield, and planting area data for Anhui, China, from 2005 to 2020. The results showed that multi-source data led to higher prediction performance than single-source data, and enabled accurate prediction of winter wheat yields three months prior to harvest. Furthermore, the ASTGNN model provided better prediction performance than two traditional crop yield prediction models (R² = 0.70, RMSE = 0.21 t/ha, MAE = 0.17 t/ha). Therefore, ASTGNN enhances the accuracy of crop yield prediction by incorporating geospatial characteristics. This research has implications for improving agricultural production management, promoting the development of digital agriculture, and addressing climate change in agriculture.

1. Introduction

Agricultural production plays a crucial role in promoting global economic development, reducing poverty, and ensuring food security. The accurate prediction of crop yields not only helps farmers and agricultural managers to optimize production management, improve production efficiency, and reduce production costs but also serves to maintain the market supply–demand balance and stabilize the prices of agricultural products [1,2]. Based on crop yield prediction data, governments can formulate and adjust agricultural policies and take reasonable agricultural security measures to ensure food security [3,4]. Therefore, the timely and accurate prediction of winter wheat yields has great significance.

Crop yields depend on diverse factors, including environmental conditions (e.g., soil and meteorological conditions) [5,6,7,8], crop variety [9,10,11], management practices (e.g., fertilization, irrigation, and crop rotation) [12,13], and plant diseases [14,15,16]. The correlation between these variables and crop yields is usually non-linear, requiring environmental data (e.g., climate and soil data) to reveal the physical conditions associated with crop growth. In recent decades, numerous studies have focused on crop yield prediction. Traditional crop yield estimation methods (e.g., sampling surveys and field observations) are inefficient, unsuitable for large-scale yield prediction, and hindered by insufficient accuracy [17,18,19]. In contrast, the continued development and innovation of intelligent and efficient remote sensing technologies has provided high-resolution, large-scale, and spatially continuous environmental information on crops. Additionally, remote sensing data can present different spectral reflectance information of crops at different key growth stages, providing a useful reference for crop yield estimation. Therefore, remote sensing technologies are among the most effective solutions for crop yield estimation [20,21,22,23,24,25,26]. Previous studies have used remote sensing data and meteorological data alone, or a combination of multi-source data, for crop yield prediction [27,28,29]. However, multi-source data may differ in quality and accuracy because they are typically acquired from different platforms or sensors, which adversely affects the accuracy and stability of the final prediction results [30,31,32,33]. In terms of large-scale crop yield prediction, few studies have determined the relative performance of prediction models based on multi-source data and single-source data.

Statistical models and process-based crop models are two common methods for predicting crop yields [34,35,36,37,38,39,40]. Process-based models simulate the physical processes of crop growth, allowing for the exploration of interactions between yield and environmental factors. However, these models require extensive field data and are computationally intensive [41,42]. In contrast, statistical models establish empirical relationships between yield and predictor variables using existing data, with regression analysis being a widely used method in early applications of remote sensing for yield estimation [43,44,45]. Traditional static regression and mechanistic models are limited in applicability and face significant uncertainties, making it difficult to provide reliable yield predictions [46,47]. While machine learning methods, such as decision trees, random forests, multiple regression, and association rule mining, have been widely applied in crop yield prediction, they typically treat yield as a latent function of input variables such as weather and soil conditions, which may fail to fully capture the complex relationships between these variables [46,48]. Deep learning is increasingly being applied across various fields within agriculture [49,50,51,52]. Deep learning, a hierarchical machine learning method, is particularly well-suited for handling unstructured and unlabeled data. It has demonstrated superior performance in agriculture, especially in feature extraction from large datasets and modeling nonlinear relationships compared to traditional machine learning methods [53]. Given that crop yield prediction depends on a variety of complex factors influencing crop growth, deep learning offers a powerful capability to extract key features from existing data, thereby improving the accuracy of yield predictions [54,55,56,57,58]. For example, Nevavuori et al. [59] used convolutional neural networks (CNNs) based on normalized difference vegetation index (NDVI) and RGB data acquired by unmanned aerial vehicles to predict wheat and barley yields with high accuracy. Moreover, Jiang et al. [60] presented a long short-term memory network (LSTM) method for crop yield prediction, which performed well in learning temporal characteristics compared with machine learning methods such as random forest and least absolute shrinkage and selection operator algorithms. Wang et al. [61] investigated the impact of different combinations of remote sensing, meteorological, and soil data on model performance, finding that incorporating soil data improved the model’s ability to capture yield spatial variability. Tian et al. [62] explored combinations of the Vegetation Temperature Condition Index (VTCI) with remote sensing data and VTCI with both meteorological and remote sensing data to determine the optimal wheat yield estimation model. Zhang et al. [63] demonstrated that combining various input data, including vegetation indices (VIs) and climate indices (CIs), enhanced crop yield prediction, identifying the best combinations of CIs, EIs, and VIs. Sharma et al. [64] applied the CNN-LSTM method for crop yield estimation, comparing models with and without background information (e.g., water bodies, farmland, and urban landscapes), and found that incorporating background information improved yield predictions. Gastli et al. [65] combined Gaussian Processes (GP) with CNN-LSTM to predict crop yields in California, showing better performance with GP-enhanced CNN-LSTM compared to CNN and LSTM alone. Sun et al. [66] used CNN-LSTM for seasonal crop yield predictions and compared it with CNN and LSTM methods, finding that CNN-LSTM performed better over five years, with MODIS surface reflectance significantly affecting model performance.

However, previous studies often treat counties as independent units, neglecting the geographical relationships between neighboring counties, which prevents effective capture of the interactions between regions. In reality, neighboring counties or areas may share similar environmental conditions, and this interconnection significantly impacts crop yield. Traditional deep learning methods, such as CNNs, do not effectively consider and learn the relationships between a node and its neighboring nodes in an image [67]. Graph Neural Networks (GNNs) can capture the interdependencies between a sample and the samples it interacts with, providing a feasible solution to the aforementioned problem [68]. Furthermore, attention mechanism-based methods can better identify the spatio-temporal characteristics of crops [69,70], thus improving the accuracy of crop yield prediction. Therefore, the objectives of this study are as follows. (1) To design a novel Attention-based Spatiotemporal Graph Neural Network (ASTGNN) model that integrates Graph Neural Networks and attention mechanisms, addressing the limitations of traditional methods in neglecting neighboring geospatial features. (2) To incorporate multi-source data, including satellite remote sensing, meteorological, and soil data, into the model to provide more comprehensive environmental information, thereby significantly improving the accuracy of winter wheat yield estimation. (3) To compare the performance of LSTM, CNN-LSTM, and ASTGNN models in capturing spatiotemporal cumulative data features, and to assess the effectiveness of multi-source versus single-source data in crop yield prediction.

2. Materials and Methodology

2.1. Study Area

Anhui is an important agricultural and grain-producing region in eastern China (Figure 1). The province stretches from 114°54′ to 119°37′ (east longitude) and from 29°41′ to 34°38′ (north latitude). Climatically, it spans temperate and subtropical zones, and abounds with mountainous and plain areas, with four distinctive seas and abundant rainfall, making it suitable for winter wheat growth. Winter wheat in the study area is usually harvested from late October to early November and from late May to early June the following year.

2.2. Dataset

For this study, we collected county-level data on winter wheat, including yields, planting area, satellite-based vegetation indices (e.g., solar-induced chlorophyll fluorescence (SIF), NDVI, and enhanced vegetation index (EVI)), climate conditions, and soil properties (

Table 1. Data used in this study.

Data	Abbreviation	Source	Spatial Resolution	Temporal Resolution
Remote sensing data
NDVI		MODIS (MOD13Q1)	250 m	16 days
EVI		MODIS (MOD13Q1)	250 m	1 days
SIF		GOSIF	0.05°	monthly
Meteorological data
Potential evapotranspiration	pet	TerraClimate	4 km	monthly
Precipitation	ppt	TerraClimate	4 km	monthly
Runoff	q	TerraClimate	4 km	monthly
Shortwave radiation	srad	TerraClimate	4 km	monthly
Maximum temperature	tmax	TerraClimate	4 km	monthly
Minimum temperature	tmin	TerraClimate	4 km	monthly
Soil data
Organic carbon content	soc	SoilGrids	100 m
Sand content	sand	SoilGrids	100 m
Clay content	clay	SoilGrids	100 m
Soil bulk density	bulk	SoilGrids	100 m
Yield data		Anhui Statistical Yearbook	county-scale	annual
Planting area data		National Ecosystem Science Data Center	30 m

). To integrate the growth stages of winter wheat, we selected data from October of the current year to May of the following year. This period encompasses the entire growth cycle of winter wheat. Additionally, monthly data were used to express all temporal characteristics, and the analysis was based on data from 2005 to 2020.

2.2.1. Remote Sensing Data

NDVI can help assess the health status and growth potential of vegetation [71]. By monitoring changes in NDVI values, we can gain insights into the growth conditions and variations in vegetation cover. EVI is an important remote sensing monitoring metric that primarily reflects seasonal and interannual variations in vegetation growth, and can be utilized to assess the health status, growth condition, and coverage of vegetation [72]. The NDVI and EVI were calculated based on MOD13Q1 images acquired from the Google Earth Engine platform. The temporal and spatial resolution of MOD13Q1 images is 16 days and 250 m, respectively. SIF can directly reflect the dynamic changes in the actual photosynthetic activity of plants. SIF data were acquired through the GOSIF data product [73]. The GOSIF data product is a global SIF data product with a resolution of 0.05°, which is produced through machine learning based on discrete SIF observations from the Orbiting Carbon Observatory-2. Drawing on previous studies [74,75,76], we selected three sources of vegetation index satellite data (NDVI, EVI, and SIF).

2.2.2. Meteorological Data

In this study, the selected meteorological data included potential evapotranspiration, precipitation, runoff, shortwave radiation, maximum temperature, and minimum temperature. All meteorological data were obtained from the Terraclimate dataset, which is a global high-spatial-resolution dataset that integrates the high-spatial-resolution WorldClim dataset with coarse-spatial-resolution data from the Climate Research Unit Ts4.0 and the Japanese 55-year Reanalysis [77].

2.2.3. Soil Data

SoilGrids is a global soil dataset [78], which mainly contains data on soil organic carbon content, sand content, clay content, and soil bulk density, with a resolution of 250 m and seven depths (0, 5, 15, 30, 60, 100, and 200 cm). Soil variables at different depths are highly correlated [38]. In this study, we selected the average values of the specified four soil variables at seven depths from SoilGrids.

2.2.4. Crop Yield Data and Planting Area Data

County-level crop yield data from 2005 to 2020 were obtained from the Anhui Statistical Yearbook (http://www.stats.gov.cn, accessed on 3 September 2024), compiled by the Anhui General Bureau of Investigation of the National Bureau of Statistics on the basis of field survey data. Crop planting area data were acquired from the National Ecosystem Science Data Center (http://www.nesdc.org.cn, accessed on 3 September 2024) [79]. We performed a preliminary quality inspection on the crop yield data to identify and remove the following outliers: (i) counties that lacked yield records in certain years during the study period and (ii) values outside the range of average values from 2005 to 2020 plus or minus the doubled variance.

2.3. Methodology

2.3.1. Technological Roadmap

To address the issue of varying pixel sizes in multi-source data, which can introduce uncertainty in validation results [80], we first standardized the spatial resolution of remote sensing data and environmental variables to 1 km and the temporal resolution to one month [81]. Continuous data, such as remote sensing and meteorological data, were resampled to 1 km using bilinear interpolation, while static data, such as soil data, were resampled using nearest neighbor interpolation. We then converted the resampled winter wheat planting area map, with a 1 km spatial resolution, into a binary image. Satellite vegetation indices and environmental variables were masked on an image-by-image basis according to winter wheat planting locations. These data were aggregated into monthly average variables at the county level before being imported into the deep learning model. Figure 2 shows the technical roadmap of this study.

2.3.2. GraphSAGE

GNNs provide a deep learning method for processing graph-structured data. A graph is a data structure that consists of nodes and edges. A GNN identifies the complex relationships and patterns in a graph structure by iteratively aggregating the characteristics of nodes and their neighbors, thereby implementing tasks such as node classification, link prediction, and graph classification. The characteristics of each node are updated at each layer by aggregating the characteristics of its neighbors. The aggregation is expressed as follows:

$$h_v^{\left(k\right)}=\mathrm{Aggregate}\left(\{h_u^{(k-1)},\forall u\in \mathcal{N}(v)\}\right)$$

(1)

where $h_v^{\left(k\right)}$ denotes the representation of node $v$ at the $k$-th layer and $\mathcal{N}\left(v\right)$ denotes the neighbor set of node $v$.

However, most existing GNN methods are based on transductive learning (i.e., learning node embeddings directly on a fixed graph), whereas most graphs continually evolve. When the network structure changes or new nodes are available, transductive learning requires retraining (this is highly complex and may cause embedding to drift), making it difficult to implement transductive learning in machine learning systems that must rapidly generate embeddings for unknown nodes. Hence, transductive learning cannot be directly generalized to unknown nodes.

Graph sample and aggregation (GraphSAGE) is a general framework for learning node embeddings based on node characteristic information and aggregating the local neighbors of nodes [82]. In our study, “vertex” refers to the vertices of the county boundaries, which form the polygonal structure of the county. We use the centroid of the county polygon as the node, representing the central location of the county. The coordinates used are projected geographic coordinates to ensure spatial consistency and accuracy in graph computations. Specifically, GraphSAGE can sample the local neighbors of a vertex and aggregate vertex characteristics, rather than training a separate embedding for each vertex, thereby learning and aggregating neighbor information by sampling a portion of nodes. By learning the information of a node, GraphSAGE can construct an aggregation function based on the characteristics of its neighbor nodes. Moreover, by simulating characteristics and neighbor relationships through the forward propagation process, GraphSAGE can obtain the representation results of new nodes. Given that the adjacency matrices of the target county are sparse and the yield graphs of different counties continually evolve, GraphSAGE is particularly suitable for crop yield prediction. The aggregation function of GraphSAGE can involve mean aggregation, pooling aggregation, or LSTM aggregation, effectively capturing the complex relationships between nodes and enhancing the model’s adaptability to dynamic changes in graph structure. The forward propagation process of the GraphSAGE model is expressed as follows:

$$h_v^{\left(k\right)}=\sigma \left(W\cdot \mathrm{Aggregate}\left(\{h_u^{(k-1)},\forall u\in \mathrm{Sample}(\mathcal{N}(v))\}\right)+b\right)$$

(2)

where $W$ and $b$ denote the learnable weight and bias and $\sigma$ denotes the activation function.

2.3.3. Constructing the ASTGNN Model

The key to crop yield forecasting lies in extracting complex non-linear relationships from historical data. Identifying the dynamic, local, and global characteristics of winter wheat yield variation is a substantial challenge. In this regard, previous models (e.g., CNNs and recurrent neural networks) have certain limitations. Owing to the limited receptive field, CNNs can only obtain local information, whereas recurrent neural networks face problems such as a vanishing gradient and explosion so cannot learn long-term dependent relationships and are only suitable for learning static time-dependent relationships; that is, they are not adaptable to the dynamic crop growth process. Therefore, we employ an ASTGNN framework, which is based on the attention mechanism, GNNs, and LSTM networks (Figure 3). The ASTGNN framework mainly comprises the following aspects: (1) input of multi-source data; (2) aggregation of node relationships between the target county and neighboring counties through the GraphSAGE module in the GNN; (3) two self-attention modules at the spatio-temporal attention layer for further capturing spatial and temporal characteristics; and (4) conversion and output of prediction results at the fully-connected layer.

Remote sensing and meteorological data obviously change in the short term, which is a dynamic change. In contrast, soil attributes remain nearly unchanged over time [83]. Therefore, in Figure 3, we use CNN(a) for extracting the dynamic change information of remote sensing and meteorological data, CNN(b) for extracting the static information of soil attributes that change insignificantly, and LSTMt for extracting the temporal features in the t-th year. H_c,t denotes the optimized embedding value after the information of adjacent counties of county c for the t-th year is aggregated by the GraphSAGE module; Z_c,t denotes the new embedding value optimized by the spatio-temporal attention layer every year; and Y_c,t denotes the predicted yield, which is outputted according to annual embedding values. The self-attention mechanism determines weights by calculating the similarity between queries, keys, and values as follows:

$$Q^t=X_t{W}_Q^t, K^t=X_t{W}_K^t, V^t=X_t{W}_K^t$$

(3)

where $X_t$ denotes the time step t at the temporal attention layer and input characteristics of spatial location t and $W_Q^t,W_K^t$, and $W_K^t$ denote the weight matrices of query, key, and value, respectively. The self-attention score is expressed as follows:

$$\mathrm{Attention}(Q^t,K,V)=\mathrm{softmax}\left({\displaystyle \frac{Q^t{K}^T}{\sqrt{d_k}}}\right)V$$

(4)

where $K$ denotes the key vector matrix of all time steps or spatial locations and $V$ denotes the value vector matrix of all time steps or spatial locations.

2.3.4. Model Training and Statistical Assessment

In this study, the Adam optimization algorithm was used for network training, which has the advantage of being able to adaptively adjust the learning rate to speed up the training process and improve stability. The initial learning rate is set to 0.01, and the total number of training rounds (epoch) is 100. In terms of hardware configuration, an NVIDIA 4090 GPU is used for model training, and the training time is about 3 h. To prevent model overfitting, regularization is introduced during training and combined with Early Stopping to automatically terminate training to ensure the model performs optimally on the validation set.

To quantify the effectiveness of the yield prediction model, we assessed the model performance by calculating the determination coefficient (R²), root mean square error (RMSE), and mean absolute error (MAE).

$${\mathrm{R}}^2=1-{\displaystyle \frac{\sum _{i=1}^n {\left(y_i-{\widehat{y}}_i\right)}^2}{\sum _{i=1}^n {\left(y_i-\overline{y}\right)}^2}}$$

(5)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{\sum _{i=1}^n {\left(y_i-{\widehat{y}}_i\right)}^2}{n}}}$$

(6)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{\sum _{i=1}^n |y_i-{\widehat{y}}_i|}{n}}$$

(7)

where $y_i$ and ${\widehat{y}}_i$ denote the statistical crop yield and predicted crop yield of the i-th county in a specific year, respectively; $\overline{y}$ denotes the annual crop yield in a specific year; and n denotes the total number of counties.

To capture cumulative effects, this study starts annual yield forecasting from 2011 and incrementally incorporates data from 2005 to 2010. Specifically, data from all years prior to the prediction year are randomly divided into training (70%) and validation (30%) sets, while the data from the prediction year itself are used as the test set [58]. For example, when predicting for the year 2011, data from 2005 to 2010 are randomly divided into training and validation sets, while data from 2011 are used as the test set, and this process is repeated for subsequent years (Figure 4). This approach effectively utilizes historical data to develop a model with improved predictive performance while minimizing overfitting. Additionally, it retains the data sequences of each year, facilitating subsequent calibration of spatial changes. To further assess the model’s performance on time series data, we selected the monthly increment data for 2020 for evaluation.

3. Results

3.1. Temporal Distribution of Crop Yield Prediction

Table 2. Annual prediction performance of LSTM, CNN-LSTM, and ASTGNN models from 2011 to 2020.

Year	LSTM			CNN-LSTM			ASTGNN
	R2	RMSE (t/ha)	MAE (t/ha)	R2	RMSE (t/ha)	MAE (t/ha)	R2	RMSE (t/ha)	MAE (t/ha)
2011	0.49	0.44	0.31	0.55	0.34	0.26	0.56	0.35	0.27
2012	0.57	0.35	0.25	0.60	0.28	0.25	0.68	0.26	0.22
2013	0.56	0.35	0.25	0.57	0.31	0.24	0.60	0.28	0.24
2014	0.62	0.28	0.24	0.65	0.27	0.23	0.67	0.27	0.22
2015	0.67	0.27	0.23	0.66	0.27	0.23	0.68	0.26	0.22
2016	0.68	0.26	0.22	0.65	0.27	0.22	0.70	0.23	0.21
2017	0.59	0.31	0.26	0.70	0.24	0.21	0.72	0.21	0.20
2018	0.66	0.27	0.23	0.70	0.24	0.17	0.73	0.21	0.18
2019	0.68	0.27	0.23	0.68	0.27	0.23	0.71	0.22	0.19
2020	0.66	0.27	0.23	0.69	0.26	0.22	0.70	0.21	0.17
Average	0.62	0.31	0.25	0.65	0.28	0.23	0.68	0.25	0.21

and Figure 5 present the yield prediction results. As data accumulated over time, model performance continued to improve. The ASTGNN model exhibited the best robustness; annual and average R² values were higher than those of the LSTM and CNN-LSTM models. Moreover, the annual R² values of the LSTM and CNN-LSTM models were alternately higher in different years, whereas the average R² value of the CNN-LSTM model was higher than that of the LSTM model. The RMSE and MAE values showed the same variation trends. According to the yield prediction performance of the three models at different growth stages using the monthly dataset of 2020 (Figure 6), the ASTGNN model again showed higher prediction performance than the other two models. The prediction accuracy of all three models increased steadily at the early growth stage (before February) and remained relatively stable from March to May. This indicates that the yield prediction results were reliable up to three months before the winter wheat reached maturity.

3.2. Spatial Distribution of Crop Yield Prediction

The predicted winter wheat yield of Anhui in 2020 was highly consistent with statistical data (Figure 7). The winter wheat-producing areas of Anhui are mainly distributed in northern Anhui, whereas winter wheat yields are low in southern Anhui owing to complex topography. Figure 8 shows the yield prediction errors of the three models compared with actual yields, which were less than 5% in most winter wheat-producing areas. Among the three models, the ASTGNN model provided the most accurate prediction results, whereas the LSTM and CNN-LSTM models exhibited large errors. The prediction errors for northern Anhui were relatively large for all models, whereas those for southern Anhui were relatively small. Moreover, large differences in yields between certain counties and their neighboring counties contributed to the prediction errors in LSTM and CNN-LSTM models but had a relatively small impact on the ASTGNN model results. This suggests that prediction accuracy can be improved by considering the geospatial information of neighboring counties.

3.3. Effect of Different Data Sources on Yield Prediction

Here, we used the ASTGNN model to assess the accuracy of winter wheat yield prediction based on multi-source data and single-source data (including remote sensing, meteorological, and soil property data). The prediction accuracy was highest when using multi-source data in 2020 (R² = 0.70, RMSE = 0.21 t/ha, and MAE = 0.17 t/ha) (

Table 3. Accuracy of winter wheat yield prediction based on four types of data sources.

Data Type	R2	RMSE (t/ha)	MAE (t/ha)
Multi-source data	0.70	0.21	0.17
Remote sensing data	0.68	0.28	0.23
Soil data	0.64	0.30	0.26
Meteorological data	0.61	0.33	0.26

and Figure 9). Among the single-source data types, the prediction accuracy was highest when using remote sensing data (R² = 0.68, RMSE = 0.28 t/ha, and MAE = 0.23 t/ha). The accuracy of predictions based on soil property and meteorological data was lower than that based on remote sensing data (R² = 0.64, RMSE = 0.30 t/ha, and MAE = 0.26 t/ha and R² = 0.61, RMSE = 0.33 t/ha, and MAE = 0.26 t/ha, respectively). This is attributed to two reasons: (1) soil property data are static data, failing to reflect real-time changes in crop growth; and (2) meteorological data involve significant spatio-temporal changes, resulting in an indirect and complex effect on crop growth. Despite differences in the prediction accuracy between different types of single-source data, the data of different sources are complementary. This indicates that each single factor is correlated with, but has no decisive effect on, the crop yield prediction accuracy of the ASTGNN model. The overall accuracy of yield prediction can be improved by using a combination of multi-source data.

4. Discussion

We conducted an ablation study on the designed ASTGNN model. As shown in

Table 4. Ablation analysis of the spatiotemporal attention layer.

Ablation Analyses	R2	RMSE (t/ha)	MAE (t/ha)
Attention Deficit Layer (ADL)	0.67	0.23	0.22
Attention Span Layer (MSL)	0.70	0.21	0.17

, when the spatiotemporal attention layer was not included, the model achieved an R² of 0.67, an RMSE of 0.23 t/ha, and an MAE of 0.22 t/ha. After incorporating the spatiotemporal attention layer, the R², RMSE, and MAE improved by approximately 4.48%, 8.70%, and 22.73%, respectively. These results demonstrate that the addition of the spatiotemporal attention layer significantly enhances the model’s accuracy in predicting winter wheat yield.

To date, numerous studies have predicted crop yields using deep-learning models. However, county-scale studies typically only consider the effects of multi-source data (e.g., remote sensing, meteorology, and soil type data) on crop yields, ignoring the effects of geospatial correlation and heterogeneity, which results in large model uncertainty. The accuracy of crop yield prediction can be improved significantly by introducing geospatial information into prediction models based on the graph embedding method, proving the effectiveness of geospatial information in crop yield prediction [67,84]. GNNs also show great potential for addressing spatial correlation [85,86].

A comparison of ASTGNN, LSTM, and CNN-LSTM models showed that model performance continued to improve as data were accumulated over time. That is, the use of more historical data can improve the generalization ability and prediction accuracy of prediction models [87]. In terms of annual crop yield prediction, the ASTGNN model showed higher robustness and accuracy than the LSTM and CNN-LSTM models. This suggests that the ASTGNN model has significant advantages in addressing complex spatio-temporal relationships and multi-source data. In addition, the ASTGNN model achieves higher R-squared and shows better stability than the LSTM and CNN-LSTM models three months before crop maturity. This suggests that the ASTGNN model showed advantages for predicting crop yields at different growth stages, especially before crops reached maturity. This finding is consistent with that of You et al. [88], namely, that prediction models based on spatio-temporal characteristics can better capture dynamic changes in crop growth.

The spatial distribution of the yield prediction results for 2020 showed the overestimation and underestimation of yields in the high-yield areas of northern Anhui and low-yield areas of southern Anhui, respectively. This is because northern Anhui has a large area of cultivated land, making it difficult to collect yield data and leading to data omissions or deviations. Moreover, the sample size of high-yield areas was limited, and the weighted average method was adopted to calculate the relative yield of each county, which ignored the specific trend in certain counties [58]. Additionally, prediction errors produced by the ASTGNN model were small, especially for southern Anhui, which has complex topography. This indicates that the prediction accuracy of the prediction model can be effectively improved by introducing geospatial neighbor information.

Among different sources of data, the prediction accuracy was highest when using multi-source data. Among single-source data, remote sensing data provided higher accuracy than soil property and meteorological data. This is because remote sensing data can capture the crop growth status (e.g., chlorophyll content and vegetation cover) of large-scale crops in real time and with high spatio-temporal resolution through satellite images. Specifically, remote sensing data can capture subtle changes in crop growth and directly indicate both the current health status and future growth trend of crops. Soil property and meteorological data are also major factors influencing crop growth. However, soil property data typically change slowly over time, making it insufficient to adequately reflect short-term variations within the crop growth cycle. Additionally, meteorological data, which cover a wide spatial scope and vary substantially over time, cannot provide crop growth information as timely and accurately as remote sensing data. Furthermore, multi-source data can provide more comprehensive and accurate crop growth information by combining the advantages of different sources of data (including remote sensing, soil property, and meteorological data) [83,89]. Therefore, our research findings suggest adopting a hybrid modeling strategy to achieve better prediction accuracy.

Although the yield prediction model developed in this study demonstrates strong performance, it has several limitations that need to be addressed. The study’s focus on winter wheat in Anhui Province, China, chosen for its complex geography—encompassing both hilly and plain areas—enhances the representativeness of the research. However, this focus may limit the applicability of the findings to other crops and regions. Future research should expand to diverse geographic areas and crop types to assess the generalizability of the proposed methods. Applying similar models in regions with varying climate and soil conditions could provide valuable insights into their effectiveness and adaptability. Additionally, issues related to data accuracy and security must be addressed [90]. The impact of certain features may be overlooked during the processing of multi-source data, potentially compromising the model’s performance in specific contexts. The model’s intricate structure and numerous parameters result in high computational costs, which can limit its applicability to large-scale datasets. Additionally, the complexity of spatial correlations might be underestimated, leading to prediction errors in both high-yield and low-yield areas, highlighting the need for further optimization to enhance predictive accuracy. Given the complexity of the dataset, future research should further explore the impact of different models and combinations of data sources on yield prediction. We acknowledge that the model’s predictive capability still has room for improvement. Future research should focus on refining multi-source data fusion techniques and enhancing data quality and coverage to minimize biases. Streamlining and simplifying model structures could significantly reduce computational costs and improve feasibility for large-scale applications. Furthermore, incorporating additional geographic and neighborhood features could enhance the model’s adaptability to complex terrains and spatial heterogeneity. Addressing concerns related to data accuracy and security, including the protection of privacy and information security during data collection, storage, and processing, is also crucial. Lastly, improving long-term prediction capabilities by integrating dynamic factors such as climate change and land use changes will bolster the model’s ability to forecast future variations, leading to more accurate crop yield predictions.

5. Conclusions

This study presents an innovative ASTGNN model that effectively integrates Graph Neural Networks and attention mechanisms, addressing the limitations of traditional methods in considering neighboring geospatial features. We combined multi-source data—including remote sensing, meteorological, and soil property data—with the ASTGNN model to predict winter wheat yields in Anhui, China. The ASTGNN model demonstrated superior performance in yield predictions from 2005 to 2020 (R² = 0.70, RMSE = 0.21 t/ha, MAE = 0.17 t/ha), accurately forecasting yields three months prior to harvest. The incorporation of multi-source data and geospatial characteristics significantly enhanced the accuracy and reliability of the predictions. Our research not only enriches the existing literature but also provides valuable insights for accurate yield predictions of other crops. This study has significant implications for improving agricultural production management, advancing digital agriculture, and addressing the impacts of climate change on agriculture.