Next Article in Journal
Exploring the Influence of Environmental and Crop Management Factors on Sorghum Nutrient Composition and Amino Acid Digestibility in Broilers
Previous Article in Journal
Resting Posture Recognition Method for Suckling Piglets Based on Piglet Posture Recognition (PPR)–You Only Look Once
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Agriculture
Volume 15
Issue 3
10.3390/agriculture15030231
agriculture-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Long Short-Term Memory Neural Network with Attention Mechanism for Rice Yield Early Estimation in Qian Gorlos County, Northeast China

by
Jian Li
1,†,
Yichen Xie
1,2,†,
Lushi Liu
2,*,
Kaishan Song
2 and
Bingxue Zhu
2
1
College of Information Technology, Jilin Agricultural University, Changchun 130118, China
2
State Key Laboratory of Black Soils Conservation and Utilization, Northeast Institute of Geography and Agroecology, Chinese Academy of Sciences, Changchun 130102, China
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Agriculture 2025, 15(3), 231; https://doi.org/10.3390/agriculture15030231
Submission received: 26 December 2024 / Revised: 9 January 2025 / Accepted: 19 January 2025 / Published: 21 January 2025
(This article belongs to the Section Digital Agriculture)
Browse Figures
Versions Notes

Abstract

:
Rice is one of the most extensively cultivated food crops in Northeast China. Estimating pre-harvest rice yield is important for accurately formulating field management strategies and swiftly assessing overall rice production. This can be achieved using a pixel-scale model, which estimates crop yield based on information from each pixel. Previous studies predominantly employed remote sensing indices, climatic data, and yield statistics of rice across either single or all growth periods for yield estimation. These approaches are limited by a delay in yield estimation and are insufficient in the exploration of time-series feature variables at the pixel scale. This study presents the development of a novel deep-learning framework designed for the early estimation of rice yield in Qian Gorlos County, Northeast China. The framework utilizes a long short-term memory neural network integrated with an attention mechanism (ALSTM). In this framework, the heading stage–milk ripening stage is the time window for early yield estimation, and the vegetation indices Normalized Difference Red Edge (NDRE), Green Chlorophyll Vegetation Index (GCVI), and Normalized Difference Water Index (NDWI) from the rice transplanting to the milk ripening stage are time-series feature variables. The yield estimation precision is R2 = 0.88, RMSE = 341.82 kg/ha, MAE = 280.29 kg/ha, outperforming LASSO (R2 = 0.71, RMSE = 567.10 kg/ha, MAE = 487.38 kg/ha), RF (R2 = 0.79, RMSE = 506.70 kg/ha, MAE = 418.90 kg/ha), and LSTM (R2 = 0.83, RMSE = 451.11 kg/ha, MAE = 326.31 kg/ha). The ALSTM introduced in this paper demonstrates its robustness after being generalized to the 2022 growing season. It can forecast the current-year rice yield two months prior to harvest, providing a valuable reference for developing field management strategies to enhance rice productivity.

1. Introduction

As one of the most significant crops worldwide, rice represents 41% of overall grain production [1]. China holds the position of the largest rice producer globally, generating approximately 206 million annual tonnes, occupying 28% of the total rice production worldwide. However, rice production in China has been facing stagnation in recent years [2,3]. To increase production, developing optimal field management strategies with the help of a rapid and precise early rice yield estimation approach, enabling predictions of the current-year yield is necessary [4,5]. This will allow growers to implement timely remedial actions if the estimated yield is less than expectations [6].
While seasoned agronomists can approximately assess yields through the observation of crop morphological characteristics in a non-invasive manner, such estimates derived from empirical knowledge are inherently subjective and unsuitable for addressing large-scale issues [7]. Remote sensing benefits for delivering timely, comprehensive, and spatio–temporal data across extensive spatial scales. It has been effectively utilized for estimating crop yields in recent decades [8,9,10]. Methods for estimating crop yield through remote sensing can be classified into two categories [11]: process-oriented crop simulation models and empirical regression models. Process-oriented crop simulation models depend on yield estimation outcomes at a single-point scale [12]. This model requires broad input variables, including the soil moisture and vegetation index, calculated from remote sensing data, fertilizer conditions, plant density, and daily weather data, which are not applicable to various crops, regions, and years due to both computational costs and limitations in data requirements [13,14]. Conversely, empirical regression techniques utilize straightforward input dimensions derived from regression techniques. Crop yields are generally forecasted by analyzing empirical correlations between vegetation indices (VIs) and actual yields recorded in the field [15]. There is no necessity for intricate crop parameters [16]. However, the prior prediction of crop yield exhibits a highly complex non-linear relationship among input variables. As a result, empirical models are only suitable for specific crop varieties, locations, and years, resulting in limited generalization across larger geographical areas [17].
Recent developments in computer hardware, algorithms, and data availability exhibit the effectiveness of machine learning (ML) in agricultural analyses [18]. The issue of inadequate generalization capabilities found in conventional empirical models can be successfully addressed by the machine learning method [19]. Different machine learning models like support vector machines, random forests, artificial neural networks, and Gaussian process regression, have been effectively utilized for yield estimation by building a nonlinear relationship between input environmental variables (such as temperature, precipitation, soil moisture, solar radiation, soil type, and vegetation index like NDVI) and yield [20,21]. These models demonstrate higher accuracy than traditional empirical regression models, which rely on statistical relationships between observed environmental variables and crop yield and avoid using crop growth parameters as variables like process-based crop models [22]. Nonetheless, challenges like overfitting, extended training durations, and a limited number of hidden layers restrict the capacity of machine learning models to address nonlinear problems and generate large-area crop yield predictions [23]. Compared with machine learning methods, the multilayer architectures of deep learning models enhance the capacity of learning features from data and are currently utilized in crop yield estimation research [24]. Long short-term memory network (LSTM) models have been applied to yield estimation, as they can handle time-series data and capture complex, nonlinear relationships. For instance, Tian et al. [25] used an LSTM deep learning model to estimate maize yield at the county level. The study integrated remotely sensed data, including NDVI, soil moisture, and meteorological data, such as temperature and precipitation, with LSTM to predict wheat yield in the Guanzhong region. However, the LSTM model alone could not fully capture the importance of climatic variables in yield prediction. To address this, we introduced an attention mechanism within the LSTM model to assess the significance of various input variables and identify key growth periods, thus enhancing the model’s accuracy in yield predictions [26,27].
Regardless, current yield estimation models still exhibit certain limitations. Existing yield estimation studies either utilize the complete growth period for yield estimation neglecting the early estimation time window, resulting in its inability to estimate the pre-harvest yield, or rely only on the feature variables of a single growth period, neglecting the temporal characteristics, resulting in a less accurate model [28,29]. Furthermore, yield estimation studies based on satellite data primarily focus on the county scale, lacking pixel-scale analyses that estimate crop yield based on information from each pixel. This limitation restricts the accurate development of field management strategies for enhancing yield [30]. In summary, current yield estimation research is limited by delays in yield assessment, inadequate analysis of time-series variables, and conduction in the pixel scale.
This paper proposes a pixel-scale early estimation framework for rice yield, integrating the attention mechanism to the LSTM network (ALSTM). It combines measured rice yield data with time-series vegetation indices to estimate rice yield in Qian Gorlos County, Northeast China. The primary scientific questions of this research are as follows: (1) How can an ALSTM deep learning framework be developed for estimating pixel-scale rice yields, and how does its precision compare to the baseline model? (2) What is the optimal early estimation time window for predicting rice yields? (3) Which vegetation indices are most significant for rice yield estimation? This study aims to answer these questions, enabling the prediction of the current-year yield two months before the rice harvest. The findings provide advantages for grain yield forecasting, grain insurance, disaster evaluation, and international trade discussions, offering guidance for state and pertinent decision-making bodies to formulate field management strategies for yield enhancement.

2. Materials and Methods

2.1. Study Area

The study area, Qian Gorlos County, is situated in the western region of Jilin Province, Northeast China (Figure 1). Qian Gorlos County belongs to a temperate continental monsoon climate and is in the semi-humid and semi-arid transition zone. The measured depth of soil freezing is 1.8 m. The typical annual average rainfall is 450.6 mm, with a peak of 612.4 mm [31]. The mean annual evaporation rate is 1696.5 mm, with a relative humidity of 62.4% in the air. The annual total amount of sunshine is 2801.6 h. These climatic conditions create an optimal environment for rice cultivation. The rice cultivation area in Qian Gorlos County covers approximately 10.55% of the total area allocated for major crops [31,32]. Over the past five years, the average annual rice production in this area is 675,800 tons. The planting pattern consists of single-season rice, transplanted during the first half of May and harvested at the end of September. The growth cycle of rice encompasses six essential growth periods: the transplanting stage, tillering stage, jointing stage, heading stage, milk ripening stage, and yellow ripening stage [33] (Table 1).

2.2. Data

2.2.1. Remote Sensing Data

This study is based on Sentinel-2 multispectral satellite images obtained from the ESA Copernicus Data Centre (https://scihub.copernicus.eu/ (accessed on 5 January 2024)). Sentinel-2 is a new generation of multispectral imaging satellites, initiated by the European Commission and the European Space Agency, and is widely utilized for environmental monitoring purposes. Sentinel-2, consisting of 2A and 2B, offers high temporal and spatial resolution and provides a revisit period of 10 days for each satellite, benefiting from two complementary satellites [34,35,36]. The remote sensing images consist of 13 bands ranging from visible and near-infrared to short-wave infrared, with a maximum spatial resolution of 10 m (Table 2). The L1C-level data products were downloaded from ESA, and the conversion of L1C data into L2A-level surface reflectance remote sensing data was accomplished using the Sen2Cor-2.5.0 plug-in. The obtained Sentinel-2 images were subsequently subjected to pre-processing, including atmospheric correction and radiometric calibration.

2.2.2. Rice Yield Measurement

The research team collected 235 actual rice yield measurement points within the rice growing area of Qian Gorlos County during the yellow ripening stage in 2022 and 2023, of which 100 points were recorded in 2022 and 135 points in 2023. To avoid sampling points in mixed pixels and enhance the representativeness of the sampling points, the selection was primarily focused on extensive rice distribution areas. The pixel size of the Sentinel-2 image is 100 m2; therefore, we partitioned the 100 m2 pixel into four equal regions. The center of each region was designated as the sampling point, with a sample area of 25 m2. One square meter of rice was harvested in each of the five directions from the center of the sample square: southeast, northwest, and north. Upon returning to the laboratory, the harvested rice underwent threshing, drying, and weighing processes. The weight for an area of 1 m2 of rice was determined based on a moisture content of 14.5%. The cumulative mass of the five orientations of the sample square was utilized to determine the rice yield of the 25 m2 sample square. The rice yields from the four sample squares were summed to calculate the yield of the pixel (Figure 2).

2.2.3. Mapping of Rice Spatial Distribution

Determining the crop area is essential for estimating crop fields. The period from late June to August in Northeast China is the peak growth period for rice, during which various physiological parameters, such as greenness and water content, attain their maximum levels. In this time window, to achieve an optimal crop classification, we developed month-by-month time-series data of the Land Surface Water Index, Enhanced Vegetation Index, Normalized Difference Vegetation Index, and Relative Vegetation Index as feature variables from June to August 2023 relying on 10 m spatial resolution Sentinel-2 images. A decision tree model was then constructed to delineate the rice planting area in Qian Gorlos County. The findings indicate that the overall classification precision of the rice spatial distribution map reaches 86.4%, with the total 78,133 ha rice area in Qian Gorlos County in 2023 (Figure 1).

2.3. Methods

We propose a deep learning framework that integrates the attention mechanism with LSTM for the early estimation of rice yield. Utilizing advanced data analytics and deep learning methodologies, we provide novel insights for pixel-level early yield estimation of single-season rice in Northeast China. The sequence of the research process is outlined below (Figure 3): (1) The Pearson correlation coefficient method is used to determine the relationship between vegetation indices and crop yields across various growth periods, allowing for the identification of vegetation indices that are particularly responsible for rice yield. (2) During various rice growth periods, the vegetation index identified in the prior step serves as an independent variable, while the measured rice yield acts as a dependent variable. This setup facilitates the development of the ALSTM time-series yield estimation model alongside three baseline models: Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and Long Short-Term Memory (LSTM). The cross-validation method was employed using 135 sampling points in 2023. (3) Compare the yield estimation results of ALSTM and the baseline models across various time series using the coefficient of determination (R2), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) metrics. (4) Identify the optimal time window for early rice yield estimation and validate the ALSTM yield estimation framework presented in this paper. The ALSTM rice yield estimation framework introduced in this paper is further validated using 2022 data, employing 135 sample points for robustness validation.

2.3.1. Calculation and Screening of Vegetation Index

Previous studies show that the sensitivity of vegetation indices to crop yield differs across various growth periods of crops. DVI, EVI, EXG, GCVI, GNDVI, NDRE, NDVI, NDWI, OSAVI, and RVI demonstrate significant sensitivity to crop yield and are frequently utilized for both qualitative and quantitative assessments of crop yield [37,38,39,40] (Table 3). To select vegetation indices that are sensitive to rice yield, we initially adjusted the temporal resolution of Sentinel-2 images from 10 days to each growth period interval of rice, following the primary growth periods of single-season rice in Northeastern China (Table 1). The conversion involved calculating the average reflectance from all Sentinel-2 images corresponding to each growth period. Following this, as detailed in Table 3, the vegetation indices for the five growth periods were calculated.
The Pearson correlation coefficient r is capable to assess the linear correlation between variables [41]. In this study, we use it to assess the relationship between rice yield and vegetation index, aiming to screen the vegetation indices that are most sensitive to rice yield among those analyzed. The absolute value of r lies within the range of 0–1. When the absolute value approaches 1, the correlation strength between the x and y variables increases. Conversely, as the value approaches 0, the correlation between the two variables x and y diminishes. The method of calculation is presented in Equation (1).
r = ∑ j = 1 n ( O j − O ¯ j ) ( y j − y ¯ j ) ∑ j = 1 n ( O j − O ¯ j ) 2 ∑ j = 1 n ( y j − y ¯ j ) 2
where O j is the estimated yield. O ¯ j is the average value of estimated yield. y j is the measured yield. y ¯ j is the average value of the measured yield. j is the sampling point number; n is the number of sampling points.
Table 3. Vegetation indices commonly used for rice yield estimation.
Table 3. Vegetation indices commonly used for rice yield estimation.
VIAbbreviationFormulaSource
Difference Vegetation Index DVINIR − RKaufman and Tanre [42]
Enhanced Vegetation IndexEVI2.5 × (NIR − R)/(NIR + 6 R − 7.5 B + L)Corti, et al. [43]
Excess Green IndexEXG2G − R − BE. Meyer, et al. [44]
Green Chlorophyll Vegetation IndexGCVINIR/G − 1Shuai, et al. [45]
Green Normalized Difference Vegetation IndexGNDVI(NIR − G)/(NIR + G)Daughtry, et al. [46]
Normalized Difference Red Edge Vegetation IndexNDRE(NIR − REG)/(NIR + REG)Li, et al. [47]
Normalized Difference Vegetation IndexNDVI(NIR − R)/(NIR + R)Li, Miao, Feng, Yuan, Yue, Gao, Liu, Liu, Ustin and Chen [47]
Normalized Difference Water IndexNDWI(G − NIR)/(G + NIR)McFeeters [48]
Optimized Soil Adjust Vegetation IndexOSAVI(1 + 0.16) × (NIR − R)/(NIR + R + 0.16)Wan, et al. [49]
Ratio Vegetation IndexRVIR/NIRPearson and Miller [50]
Note: B, G, R, REG, and NIR represent the reflectance of blue, green, red, red edge, and near-infrared bands, respectively. VI represents Vegetation Index. L represents a constant used to correct for background noise (typically set to 10,000).

2.3.2. ALSTM Neural Network

The ALSTM model developed in this study is a deep learning framework designed for learning time-series features in remote sensing and yield data. It consists of two main components [51]: the LSTM network and the attention mechanism (Figure 4). The LSTM network effectively handles time-series data processing, whereas the attention mechanism resembles the human selective visual attention process which prioritizes key components of the input sequence data that influence the output data, thereby facilitating the acquisition of essential information from the input sequence data [52]. The ALSTM framework consists of six distinct layers: an input layer, an LSTM layer, an attention mechanism layer, an output layer, and two dropout layers. The input layer is composed of the vegetation index of rice corresponding to each growth period. The dropout layer modifies the parameters by severing connections from the neurons with a specified probability during the training of the model. This method is designed to light the risk of model overfitting. The attention mechanism layer, implemented after the LSTM layer, serves to allocate probability values that reflect the impact of the crop growth process on yield. This document details the utilization of the TensorFlow framework for the construction of the ALSTM model. The input data are scaled to the range of 0–1 through Max–Min normalization. The 5-fold cross-validation technique is employed to allocate 80% of the yield measurement data for training purposes while reserving 20% for model validation. This approach aims to minimize variance in model selection and enhance the reliability of the model.
After conducting multiple training sessions on the network, the optimal parameters are established as follows: The batch size is configured to be one-fifth of the total sampling points of the crops, while the number of training epochs for the network is established at 1000. The LSTM unit is configured to 10. The starting learning rate is established at 0.0006. The optimization method utilized is Stochastic Gradient Descent. The loss function is the mean squared error. To prevent overfitting during network training and to enhance the network’s generalization capability, a dropout layer is incorporated before the LSTM layer and after the attention mechanism layer, with the dropout parameter configured to 0.1.

2.3.3. Baseline Models

We also selected LASSO, RF, and LSTM as baseline models to facilitate the comparison and evaluation of the ALSTM framework regarding the precision of rice yield estimation. LASSO is indicative of the conventional approach to linear regression. At this stage, the random forest model is the most common approach for yield estimation in the field of machine learning. The LSTM framework is a deep learning model for time-series analysis that resembles ALSTM but without an attention mechanism.
LASSO is a traditional linear regression model that reduces the model parameters by the L1 approach [53]. The model parameter values can be minimized towards 0 to compress the parameters unrelated to the dependent variable. This approach aims to enhance prediction accuracy and model explanatory power while lowering the risk of overfitting. The L2 penalty coefficient established in this study is 0.002.
RF is based on the concept of integrated learning, merging several decision trees for classification or regression to enhance model performance and mitigate the risk of overfitting in a singular model [54]. Every node within the tree of a random forest model signifies a random selection of features. Limiting the number of extracted features enhances the model’s generalization ability. RF has been extensively utilized in crop yield estimation due to its aforementioned characteristics, demonstrating its capability to effectively estimate crop yields on a large scale. Following the optimization of hyper-parameters, this study establishes the number of trees (Ntree) in the Random Forest (RF) model at 500, and the maximum number of nodes (Maxnodes) in each tree is set to 300.
LSTM is a deep learning model designed for modeling time-series data [55]. It features an architecture that utilizes time-transferred information units to represent the temporal logic of high-dimensional data. Meanwhile, it features the temporal unit architecture characteristic of a recurrent neural network, which is adapted to the cumulative process of crop growth. Its capability to recognize patterns in extended time-series data and to effectively model intricate non-linear relationships has been shown. This paper presents an LSTM architecture that includes an input layer, two hidden layers with LSTM units, and a final output layer. This model comprehensively analyzes the intricate time-series correlations among features and acquires knowledge of the growth characteristics of rice across five distinct growth periods from remotely sensed images.

2.3.4. Model Evaluation

To mitigate overfitting, we employed k-fold cross-validation during model training. The samples are randomly partitioned into k sub-datasets. In each iteration, k − 1 of these sub-datasets are used for modeling, while the remaining sub-dataset serves as the validation set. The final precision of the model is determined by calculating the average of k modeling accuracies. The value of k is set to 5 in this document. In 2023, a total of 135 sampling points were collected and categorized into an average of five sub-sample sets. Four of these sets are designated for training purposes, while the remaining set serves as the validation set. The process of cross-validation is executed five times, with each sub-sample undergoing validation once. The final precision of the model training and validation sets is determined by averaging the evaluation results from the five training and validation sets. We select R2, RMSE, and MAE as the metrics for assessing model accuracy. R2 measures the proportion of variation in the dependent variable explained by the independent variable. A higher R2 indicates a better model fit. RMSE quantifies the difference between predicted and actual values, with a lower RMSE indicating better model precision. MAE calculates the average error in predictions, with a lower MAE indicating higher accuracy.

3. Results and Analysis

3.1. Correlation Between Rice Yield and Vegetation Index

The spectral reflectance data for red, green, blue, near-infrared, and red-edge bands were extracted from Sentinel-2 images. Subsequently, various spectral indices including DVI, EVI, EXG, GCVI, GNDVI, NDRE, NDVI, NDWI, OSAVI, and RVI were calculated to develop a model for estimating rice yield. Table 3 presents the specific calculation formula. The correlation coefficient r was utilized to examine the relationship between rice yield and different vegetation indices during each growth period. The heat map illustrating the correlation between vegetation indices and measured crop yield indicates that NDRE, GCVI, NDWI, OSAVI, and NDVI exhibit a stronger correlation with rice yield during the growth period of GP3-GP5, achieving correlation values exceeding 0.65 [56,57]. This performance surpasses that of the other five vegetation indices, in which GNDVI demonstrates the weakest correlation (Figure 5). In the interim, the relationship between the vegetation indices of GP3, GP4, and GP5 and the measured yield are stronger than those of GP1 and GP2, with an order of GP4 > GP3 > GP5. The most significant correlation is shown between the heading stage and the milky ripening stage (Figure 5).
We conducted autocorrelation analyses of different vegetation indices in GP4, which reveal strong correlations between NDRE and NDVI, as well as between NDRE and OSAVI, with values of 0.75 and 0.76, respectively (Figure 5). The three vegetation indices incorporate near-infrared bands. The reflectance of the vegetation NIR band is primarily determined by the internal structure of the leaves. This reflectance serves as a key indicator of vegetation health and is crucial for modeling the growth status of surface crops, assessing the degree of cover, and understanding crop growth dynamics. NDRE contains both the near-infrared band and the red edge band, which is characterized by a rapid change in vegetation reflectance at the intersection of near-infrared and red light. This feature effectively indicates the condition of plant pigmentation and overall health. Consequently, NDRE was retained while NDVI and OSAVI were excluded due to their autocorrelation with NDRE. Ultimately, among the five vegetation indices, NDRE, GCVI, NDWI, OSAVI, and NDVI, the most pertinent to rice yield, NDRE, GCVI, and NDWI were chosen for the building of the estimation model.

3.2. Time-Series for Early Yield Estimation

Understanding key growth periods is essential for accurate yield estimation. We compared the performance of ALSTM against three baseline models: LASSO, RF, and LSTM. This comparison focuses on the effects of different time series on yield estimation. The time-series data were from GP1 to GP2, GP3, GP4, and GP5. We established the ALSTM estimation framework alongside the baseline models, employing NDRE, GCVI, and NDWI vegetation indices as the feature variables. The R2, RMSE, and MAE of the four aforementioned models were compared across various time series (Figure 6).
The results indicate that the R2 values of the ALSTM model for rice yield estimation across various growth periods surpassed those of the LASSO, RF, and LSTM models. Additionally, the RMSE and MAE for the ALSTM model are lower than the LASSO and RF models (Figure 6). The significance of vegetation indices for yield estimation differs across various growth periods. The four models consistently suggest GP4 as a turning period in the precision of yield estimation, as illustrated by the pink background bars in Figure 6. The performance of the models increases before GP1-GP4. After GP1-GP4, when rice growth enters the milky and yellow ripening stages, the model performance decreases without significant fluctuations (shown by the yellow background bars in Figure 6). The performance of the three baseline models demonstrates continued improvement with the expansion of the time series. However, the performance of these models at GP1-GP5 exhibits minimal enhancement compared to GP1-GP4. The ALSTM model displays enhanced performance as the time series expanded before GP1-GP4; however, the R2 value at GP1-GP5 shows a decline of 0.1 in comparison to GP1-GP4. The results indicate that an extended time series does not necessarily lead to a more favorable yield estimation. The enhancement of baseline model performance is minimal although incorporating GP5 into the time series. To achieve early yield prediction before rice maturity and meanwhile maintain model accuracy, we utilized GP1-GP4 as time-series data.
The ALSTM framework shows superior accuracy compared to LASSO, RF, and LSTM. It effectively captures long time-series data, outperforming baseline models, and exhibits heightened sensitivity to the impact of vegetation indices on yield estimation across various growth periods. GP4 represents a crucial phase for the formation of rice yield and the accuracy of yield estimation. This period serves as the time window for estimating yield. The ALSTM framework utilizing the GP1-GP4 time series demonstrates superior performance, achieving accurate yield predictions (R2 > 0.8) from 1 to 2 months before rice harvest.

3.3. Weight Analysis of Feature Variables Based on Attention Mechanism

A deep learning model operates as a ‘black box’; it is essential to identify the features of the input variables that are pertinent to the target vectors and enhance the interpretability of the deep learning model [58]. The visualization of the attention mechanism offers insight into the model’s decision-making process, thereby enhancing the model’s transparency and credibility. Relying on visualization, we can identify which components of the model are receiving attention during data processing. Examining the attention distribution allows researchers and developers to comprehend how the model retrieves valuable information from the input data, which is crucial for troubleshooting and enhancing model performance. Feature importance analysis helps improve the interpretability of the deep learning model by revealing which input variables have a greater impact on the model’s predictions and how these variables vary across different growth periods. This insight allows us to understand the model’s behavior and refine the decision-making process.
Figure 7 illustrates the weights of the vegetation indices derived from the ALSTM model for GP1 through GP5. It indicates that the weights of various vegetation indices during the same growth period exhibit similarities; however, the weights of the same vegetation index across different growth periods show significant variations. The weights assigned to each vegetation index at GP1 and GP2 are comparatively lower, falling within the range of from 0.35 to 0.49. The weights of each vegetation index at GP3 and GP4 show an increase, although GP4 exhibits higher weights (0.82–0.83) than GP3 (0.71–0.76). Nevertheless, the weights of vegetation indices at GP5 (0.58–0.68) show a subsequent decrease. The analysis demonstrates that GP3 and GP4 have a greater impact on rice yield than GP1, GP2, and GP5, and GP4 is the most significant contributor to the yield. This aligns with the findings presented in Section 3.2, indicating that GP4 is a crucial phase for rice yield development and the accuracy of estimations. Additionally, it confirms that GP1-GP4 can serve as the time-series input for the ALSTM yield model.

3.4. Robustness Validation of the ALSTM Yield Estimation Framework

We assessed the reliability of the ALSTM rice yield estimation framework by applying it to data from 2022. The total rice yield predicted by the model in Qian Gorlos County was compared with the official released yield. Initially, the rice planting area in 2022 was extracted through the methodology outlined in Section 2.2.2, obtaining a total area of 79,786 ha. The ALSTM estimation framework introduced in this paper, along with the three baseline models (LASSO, RF, and LSTM), were then employed for the rice yield estimation in 2022. The total rice production in Qian Gorlos County in 2022, as calculated from the four models, was 700,230 t, 741,420 t, 720,825 t, and 638,445 t, respectively. The yield estimation results are ultimately compared with the official yield statistics to evaluate the robustness of the model performance. The Statistics Bureau of Qian Gorlos County reported that the total rice production for the year 2022 was 686,500 t. Compared with the actual production, the ALSTM results display an increase of 2%, LASSO shows an increase of 8%, the RF model shows an increase of 5%, whereas LSTM shows a decrease of 7%. This overestimation or underestimation is mainly due to the problem of mixed pixels coupled with unfavorable climatic conditions During the rainy season and cloudy weather, the calculation of vegetation indices will be affected. Additionally, paddy fields often coexist with various other land cover types, such as cash crops, fruit trees, and shrubs adjacent to roads. This can result in the misinterpretation of rice characteristics alongside other land cover patterns, potentially inflating yield modeling outcomes. In summary, the estimates produced by the ALSTM model are the closest to the government statistical yields in 2022, which demonstrates greater accuracy compared to the baseline model, indicating the reliability of the ALSTM framework introduced in this paper.
In addition, we evaluated the precision of ALSTM and baseline models utilizing 135 samples from 2022, and generated scatter plots along with yield spatial distribution maps (Figure 8). The results indicate that the R2, RMSE, and MAE values for ALSTM are 0.85, 115 kg/ha, and 74 kg/ha, respectively, demonstrating superior precision compared to the other three baseline models. The yield spatial distribution map indicates that the three baseline models did not accurately estimate the yield in the high-value zone (Figure 9). The vegetation index in areas with high yield values may reach saturation, making it unable to respond to changes in yield. In cases where certain vegetation indices exhibit saturation and show minimal variation with yield, the ALSTM model demonstrates superior capability compared to the three baseline models in extracting significant yield-related information from the integration of multiple vegetation indices and time-series data.

4. Discussion

4.1. Advantages of ALSTM in Yield Estimation

In agriculture, knowing the impact of phenological features on crop yields is important for farmers and agriculturists in the development of sustainability programs. This study integrates the Attention mechanism with the LSTM model to extract key yield-related climatic feature variables, constructing a more reliable and robust model for estimating rice yields in regions like Northeast China. The results indicate that the accuracy of the ALSTM time-series yield estimation model surpasses that of the baseline model. This can be ascribed to two benefits of the attention mechanism [59]. First, the attention mechanism is capable of allocating varying attention (weights) to distinct input variables, assigning higher weights to significant information while lower weights to disregard irrelevant data. This benefit is evident from our results, as the ALSTM model assigns higher weights to key climatic features during critical growth periods (e.g., GP3 and GP4) while minimizing the impact of less relevant data, as shown in the weight distribution in Figure 7. Additionally, this mechanism can continuously modify the weights, allowing for the selection of important information across diverse contexts. Our findings also demonstrate this benefit, as the attention mechanism helps the ALSTM model focus on crucial data during key periods, ensuring effective use of time-series data even with longer sequences. This addresses challenges associated with recurrent neural networks (RNNs) [60], such as performance decline with longer input time-series data and inefficiencies in computation resulting from the sequence of temporal feature inputs.
Precedents exist for the application of the ALSTM model in yield estimation. The authors of [60] developed a novel deep-learning framework utilizing ALSTM for the estimation of winter wheat yield in the Guanzhong Plain. The results indicated that the ALSTM model enhanced estimation accuracy and successfully predicted winter wheat yield 20 days earlier than the LSTM model. The ALSTM framework developed by Tian, Wang, Tansey, Han, Zhang, Zhang, and Li [52] is suitable for county-level applications; however, the rice yield estimation framework presented in this paper is designed for pixel-level analysis, providing enhanced precision for the formulation of field management strategies.

4.2. Important Growth Periods for Rice Yield Estimation

The fertility cycle of rice begins with seed germination and culminating in the production of new seeds. The final crop yield is determined by the accumulation of organic matter across various growth periods throughout the reproductive cycle. Each growth period contributes differently to the overall yield. Evaluating the ideal lead time for estimating rice yield becomes essential for ensuring food security [61].
The critical growth period of single-season rice in Northeast China includes the transplanting stage, tillering stage, jointing stage, spike-tapping stage, milk-ripening stage, and yellow ripening stage. The transplanting phase indicates the arrangement of rice seedlings, the density of planting, and the initial growth condition. During the tillering stage, a significant quantity of rice leaves develops, and the physicochemical characteristics of these leaves indicate the nutritional condition of the rice plant. The jointing stage represents the primary phase of growth for the rice stem base, rather than the key period for the accumulation of organic matter in crop yield organs. Throughout this phase, rice stalks exhibit significant vertical growth, while the leaves of the rice plant remain comparatively small. The presence of bare ground influences the spectral reflectance data. The accuracy of the rice yield estimation model utilized during the jointing period is insufficient. The heading stage represents a critical phase of development characterized by enhanced reproductive growth and nutrient uptake in crops, both of which occur with increased vigor and coevolution. Throughout this phase, the rate of dry matter accumulation in rice increases and attains its maximum level. During the milky ripening stage, the quantity of grains in the spike of the rice plant is essentially established. The increase in the dry weight of grains occurs at a rapid pace, while the accumulation of dry matter takes place gradually. Nutrients synthesized by the leaves are continuously transported to the grains in significant quantities [33,62,63]. Consequently, GP4 represents the most pivotal phase in establishing rice yield during the heading and milky ripening stages. The findings align with the results indicating that the yield estimation model utilizing the time-series data up to GP4 demonstrates the highest precision, with the vegetation index weights also peaking at GP4 (Figure 6 and Figure 7). During these two critical rice growth periods, the impacts of collapse, water stress, and light are significant in the final formation of rice yield. During the tasseling stage, inadequate water levels are the primary factor influencing rice yield. As rice growth progresses into the yellow ripening stage, there is a continued decrease in the chlorophyll content of the rice leaves [64,65]. The relationship between various vegetation indices and crop yield is diminishing. The reliability and stability of the estimation model decrease after the rice enters the yellow ripening stage for this reason.
Several previous studies have identified the tasseling and milk ripening phases as critical growth periods for estimating rice yield. Yang et al. [66] performed an estimation of rice yield utilizing a Convolutional Neural Network, leveraging unmanned aerial imagery captured during the milk ripening phase. The results indicate that the network trained using the RGB dataset during the milk ripening stage demonstrates outstanding spatial generality. Son et al. [67] conducted a large-scale estimation of rice yields in the Mekong Delta, Vietnam, with the MODIS-enhanced vegetation index and normalized difference index, focusing on the heading stage as the window period. Yang et al. [68] proposed a methodology for estimating rice yield based on milky season data. The framework developed a new strategy and methodology by hierarchical modeling and a normalized weight decision-making strategy. This approach aims to serve as a reference for achieving real-time, accurate yield estimation through unmanned aerial remote sensing in the future. The growth periods utilized for yield estimation in the prior research align with this paper, suggesting that the tasseling and milk ripening phases serve as the critical time windows for rice yield estimation. Additionally, GP1-GP4 can be employed as the time series for early yield prediction.

4.3. Important Vegetation Indices for Rice Yield Estimation

Vegetation indices provide insights into the physical characteristics of crops during various growth periods, including chlorophyll content, leaf area, and biomass, all of which are associated with yield. They established a foundational framework for estimating crop yields through the application of vegetation indices. The analysis of Pearson’s correlation indicated that NDRE, GCVI, and NDWI exhibited strong correlations with rice yield.
The distinction between the red and near-infrared (NIR) bands serves as a method to assess vegetation greenness, crop density, and crop growth. NDVI is the most widely utilized vegetation index derived from red and near-infrared data for estimating the biophysical characteristics of vegetation. However, NDVI reaches saturation at elevated levels of leaf area index (LAI). The sensitivity of NDVI to LAI diminishes at this stage, with LAI serving as a crucial indicator of crop yield, thereby influencing the accuracy of NDVI in yield estimation. NDRE substitutes the near-infrared band in NDVI with the red-edge band, enhancing sensitivity to LAI, chlorophyll content, and biomass through the incorporation of red-edge spectral information [47]. This aligns with our conclusion that NDRE exhibits a robust correlation with measured yield in crop yield estimation.
GCVI is primarily used in agricultural and environmental research, demonstrating sensitivity to variations in chlorophyll levels within plants. Particularly when LAI exceeds 3, GCVI demonstrates greater sensitivity to variations in rice LAI compared to NDVI. The chlorophyll content of the canopy serves as an indicator of the physiological condition of rice and the intensity of its photosynthetic activity [45]. The variations in GCVI effectively facilitate real-time monitoring of rice growth, enabling accurate estimation of rice yield.
NDWI has the advantage that the near-infrared band exhibits significant absorption and minimal reflectivity in water bodies, whereas it shows high reflectivity in vegetation. The system reacts to the spatial and temporal patterns of surface water conditions by inhibiting vegetation growth and emphasizing the presence of water bodies. The NDWI demonstrates heightened sensitivity to irrigation practices in semi-arid regions characterized by low-density agricultural activities, particularly when analyzing data across broader scales and extended periods. This indicates that NDWI is more beneficial for the long-term monitoring of paddy field crops compared to dry fields. Furthermore, NDWI exhibits a slower saturation rate at increased crop leaf area compared to NDVI and EVI, and it effectively monitors the benefits associated with plant water stress. Consequently, NDWI effectively addresses the impact of drought on rice yield. Furthermore, NDWI can be integrated with other vegetation indices to improve the analysis of the ecological environment [48]. This aligns with the objective of this paper to develop a yield estimation model utilizing three vegetation indices, NDWI, NDRE, and GCVI, concurrently to enhance model accuracy.

4.4. Uncertainties and Prospects

This paper introduces the ALSTM rice time series estimation framework, emphasizing the broad application of deep learning models in agricultural remote sensing. This study identifies GP4 as the optimal time window for early yield estimates, highlighting the model’s ability to capture time series data. This research provides a foundational framework for grain yield forecasting and food insurance practices. However, this study still contains some uncertainties and areas for further exploration.
(1) The validation of this study is limited to a specific region, and broader testing across different geographical and climatic conditions is still lacking. This limits the evaluation of the model’s generalizability and practical applicability. Therefore, future work will focus on assessing the performance of ALSTM in yield estimation across different geographic and climatic regions, further validating its applicability and generalizability. (2) Regarding the selection of model input variables, future research may consider adding parameters such as leaf area index, chlorophyll content, soil moisture, and precipitation on top of vegetation indices to further improve the accuracy and reliability of yield estimation. By expanding the input features, the model can capture more factors related to yield, enhancing the precision of predictions. (3) ALSTM enhances model interpretability by visualizing the weights of input variables, while traditional crop models integrate established knowledge of crop physiology and agronomy, making the influence of input variables on crop growth clearer. Future research will focus on exploring how to combine deep learning models with process-based crop models, thereby improving model interpretability while simplifying the computational process of deep learning models, further advancing yield estimation models. (4) In addition to ALSTM and LSTM models, other time-series models such as GRU and Transformer also show certain advantages in time-series prediction. In the future, we will explore the performance of these models across different dataset sizes and investigate their potential applications in crop yield forecasting. By leveraging the strengths of these advanced time-series models, we anticipate providing new solutions to improve yield estimation accuracy.

5. Conclusions

In this study, a pixel-scale ALSTM rice yield estimation framework based on Sentinel-2 image and ground-truthed yield data was proposed and applied in Qian Gorlos County, recognized as a typical cultivation region for first-season rice in Northeast China. We determined the optimal vegetation indices and the time window for early yield estimation associated with rice yield. The findings indicate that GP4, identified as the heading stage–milk maturity stage, is the early yield estimation time window for rice cultivation in Northeast China. The NVRE, GCVI, and NDWI vegetation indices of GP1–GP4 were incorporated into the ALSTM framework as time-series variables. The resulting R2, RMSE, and MAE values were 0.88, 341.82 kg/ha, and 280.29 kg/ha, respectively. The ALSTM framework demonstrates superior precision compared to LASSO, RF, and LSTM, and also exhibits effective capabilities in capturing time-series data. Finally, the proposed ALSTM early yield estimation model was extended to 2022, yielding R2, RMSE, and MAE values of 0.85, 408 kg/ha, and 339 kg/ha, respectively. The discrepancies between the yield estimation results and the statistical yearbook were under 2%, demonstrating the robustness of the yield estimation framework. The yield estimation framework introduced in this paper for one-season rice in Northeast China is capable of predicting the current-year yield two months before the rice harvest, which offers critical insights for rice yield forecasting, disaster assessment, and policy formulation by relevant decision-making authorities, aiding in the mitigation of yield loss risks before the end of the growing season.

Author Contributions

J.L.: Conceptualization, data curation, formal analysis, methodology, software, and writing—original draft; Y.X.: investigation, methodology, software, and writing—original draft; L.L.: conceptualization, formal analysis, supervision, and writing—review and editing; K.S.: software and validation; B.Z.: software and validation. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (XDA28050400), the Natural Science Foundation of Jilin Province (YDZJ202301ZYTS264), the Common Application Support Platform for National Civil Space Infrastructure Land Observation Satellites (2017-000052-73-01-001735), and Technical Services (Natural Sciences)—Commissioned by enterprises and institutions (20240179).

Institutional Review Board Statement

Not applicable.

Data Availability Statement

The original contributions presented in this study are included in the article; further inquiries can be directed to the corresponding author.

Acknowledgments

The authors would like to express their gratitude to Fei Zhou for helping us improve the English in this article.

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References

  1. Peng, S.; Tang, Q.; Zou, Y. Current status and challenges of rice production in China. Plant Prod. Sci. 2009, 12, 3–8. [Google Scholar] [CrossRef]
  2. Xin, F.; Xiao, X.; Dong, J.; Zhang, G.; Zhang, Y.; Wu, X.; Li, X.; Zou, Z.; Ma, J.; Du, G. Large increases of paddy rice area, gross primary production, and grain production in Northeast China during 2000–2017. Sci. Total Environ. 2020, 711, 135183. [Google Scholar] [CrossRef] [PubMed]
  3. Bin Rahman, A.R.; Zhang, J. Trends in rice research: 2030 and beyond. Food Energy Secur. 2023, 12, e390. [Google Scholar] [CrossRef]
  4. Son, N.; Chen, C.; Chen, C.; Guo, H.; Cheng, Y.; Chen, S.; Lin, H.; Chen, S. Machine learning approaches for rice crop yield predictions using time-series satellite data in Taiwan. Int. J. Remote Sens. 2020, 41, 7868–7888. [Google Scholar] [CrossRef]
  5. Son, N.; Chen, C.; Cheng, Y.; Toscano, P.; Chen, C.; Chen, S.; Tseng, K.; Syu, C.; Guo, H.; Zhang, Y. Field-scale rice yield prediction from Sentinel-2 monthly image composites using machine learning algorithms. Ecol. Inform. 2022, 69, 101618. [Google Scholar] [CrossRef]
  6. Timsina, J.; Connor, D.J. Productivity and management of rice–wheat cropping systems: Issues and challenges. Field Crop Res. 2001, 69, 93–132. [Google Scholar] [CrossRef]
  7. Yang, G.; Liu, J.; Zhao, C.; Li, Z.; Huang, Y.; Yu, H.; Xu, B.; Yang, X.; Zhu, D.; Zhang, X. Unmanned aerial vehicle remote sensing for field-based crop phenotyping: Current status and perspectives. Front. Plant Sci. 2017, 8, 1111. [Google Scholar] [CrossRef]
  8. Jin, X.; Kumar, L.; Li, Z.; Feng, H.; Xu, X.; Yang, G.; Wang, J. A review of data assimilation of remote sensing and crop models. Eur. J. Agron. 2018, 92, 141–152. [Google Scholar] [CrossRef]
  9. Zhu, X.; Cai, F.; Tian, J.; Williams, T.K. Spatiotemporal fusion of multisource remote sensing data: Literature survey, taxonomy, principles, applications, and future directions. Remote Sens. 2018, 10, 527. [Google Scholar] [CrossRef]
  10. Khanal, S.; Kc, K.; Fulton, J.P.; Shearer, S.; Ozkan, E. Remote sensing in agriculture—Accomplishments, limitations, and opportunities. Remote Sens. 2020, 12, 3783. [Google Scholar] [CrossRef]
  11. Prasad, A.K.; Chai, L.; Singh, R.P.; Kafatos, M. Crop yield estimation model for Iowa using remote sensing and surface parameters. Int. J. Appl. Earth Obs. 2006, 8, 26–33. [Google Scholar] [CrossRef]
  12. Wang, E.; Engel, T. SPASS: A generic process-oriented crop model with versatile windows interfaces. Environ. Model. Softw. 2000, 15, 179–188. [Google Scholar] [CrossRef]
  13. Manivasagam, V.S.; Rozenstein, O. Practices for upscaling crop simulation models from field scale to large regions. Comput. Electron. Agric. 2020, 175, 105554. [Google Scholar] [CrossRef]
  14. Divya, K.L.; Mhatre, P.H.; Venkatasalam, E.P.; Sudha, R. Crop simulation models as decision-supporting tools for sustainable potato production: A review. Potato Res. 2021, 64, 387–419. [Google Scholar] [CrossRef]
  15. Ding, Y.; Zhang, H.; Wang, Z.; Xie, Q.; Wang, Y.; Liu, L.; Hall, C.C. A comparison of estimating crop residue cover from sentinel-2 data using empirical regressions and machine learning methods. Remote Sens. 2020, 12, 1470. [Google Scholar] [CrossRef]
  16. Balaghi, R.; Tychon, B.; Eerens, H.; Jlibene, M. Empirical regression models using NDVI, rainfall and temperature data for the early prediction of wheat grain yields in Morocco. Int. J. Appl. Earth Obs. 2008, 10, 438–452. [Google Scholar] [CrossRef]
  17. Gaso, D.V.; Berger, A.G.; Ciganda, V.S. Predicting wheat grain yield and spatial variability at field scale using a simple regression or a crop model in conjunction with Landsat images. Comput. Electron. Agric. 2019, 159, 75–83. [Google Scholar] [CrossRef]
  18. Benos, L.; Tagarakis, A.C.; Dolias, G.; Berruto, R.; Kateris, D.; Bochtis, D. Machine learning in agriculture: A comprehensive updated review. Sensors 2021, 21, 3758. [Google Scholar] [CrossRef]
  19. Saleem, M.H.; Potgieter, J.; Arif, K.M. Automation in agriculture by machine and deep learning techniques: A review of recent developments. Precis. Agric. 2021, 22, 2053–2091. [Google Scholar] [CrossRef]
  20. Sharma, A.; Jain, A.; Gupta, P.; Chowdary, V. Machine learning applications for precision agriculture: A comprehensive review. IEEE Access 2020, 9, 4843–4873. [Google Scholar] [CrossRef]
  21. Sharma, R.; Kamble, S.S.; Gunasekaran, A.; Kumar, V.; Kumar, A. A systematic literature review on machine learning applications for sustainable agriculture supply chain performance. Comput. Oper. Res. 2020, 119, 104926. [Google Scholar] [CrossRef]
  22. Araújo, S.O.; Peres, R.S.; Ramalho, J.C.; Lidon, F.; Barata, J. Machine learning applications in agriculture: Current trends, challenges, and future perspectives. Agronomy 2023, 13, 2976. [Google Scholar] [CrossRef]
  23. Chlingaryan, A.; Sukkarieh, S.; Whelan, B. Machine learning approaches for crop yield prediction and nitrogen status estimation in precision agriculture: A review. Comput. Electron. Agric. 2018, 151, 61–69. [Google Scholar] [CrossRef]
  24. Kamilaris, A.; Prenafeta-Boldú, F.X. Deep learning in agriculture: A survey. Comput. Electron. Agric. 2018, 147, 70–90. [Google Scholar] [CrossRef]
  25. Tian, H.; Wang, P.; Tansey, K.; Zhang, J.; Zhang, S.; Li, H. An LSTM neural network for improving wheat yield estimates by integrating remote sensing data and meteorological data in the Guanzhong Plain, PR China. Agric. For. Meteorol. 2021, 310, 108629. [Google Scholar] [CrossRef]
  26. Niu, Z.; Zhong, G.; Yu, H. A review on the attention mechanism of deep learning. Neurocomputing 2021, 452, 48–62. [Google Scholar] [CrossRef]
  27. Liu, F.; Jiang, X.; Wu, Z. Attention mechanism-combined LSTM for grain yield prediction in China using multisource satellite imagery. Sustainability 2023, 15, 9210. [Google Scholar] [CrossRef]
  28. Shahhosseini, M.; Hu, G.; Archontoulis, S.V. Forecasting corn yield with machine learning ensembles. Front. Plant Sci. 2020, 11, 1120. [Google Scholar] [CrossRef]
  29. Rashid, M.; Bari, B.S.; Yusup, Y.; Kamaruddin, M.A.; Khan, N. A comprehensive review of crop yield prediction using machine learning approaches with special emphasis on palm oil yield prediction. IEEE Access 2021, 9, 63406–63439. [Google Scholar] [CrossRef]
  30. Jeong, S.; Ko, J.; Yeom, J. Predicting rice yield at pixel scale through synthetic use of crop and deep learning models with satellite data in South and North Korea. Sci. Total Environ. 2022, 802, 149726. [Google Scholar] [CrossRef]
  31. Liu, L.; Xie, Y.; Zhu, B.; Song, K. Rice leaf chlorophyll content estimation with different crop coverages based on Sentinel-2. Ecol. Inform. 2024, 81, 102622. [Google Scholar] [CrossRef]
  32. Ma, X.; Ma, Y. The spatiotemporal variation analysis of virtual water for agriculture and livestock husbandry: A study for Jilin Province in China. Sci. Total Environ. 2017, 586, 1150–1161. [Google Scholar] [CrossRef] [PubMed]
  33. Sánchez, B.; Rasmussen, A.; Porter, J.R. Temperatures and the growth and development of maize and rice: A review. Glob. Change Biol. 2014, 20, 408–417. [Google Scholar] [CrossRef]
  34. Wang, Q.; Shi, W.; Li, Z.; Atkinson, P.M. Fusion of Sentinel-2 images. Remote Sens. Environ. 2016, 187, 241–252. [Google Scholar] [CrossRef]
  35. Wang, Q.; Atkinson, P.M. Spatio-temporal fusion for daily Sentinel-2 images. Remote Sens. Environ. 2018, 204, 31–42. [Google Scholar] [CrossRef]
  36. Raiyani, K.; Gonçalves, T.; Rato, L.; Salgueiro, P.; Marques Da Silva, J.R. Sentinel-2 image scene classification: A comparison between Sen2Cor and a machine learning approach. Remote Sens. 2021, 13, 300. [Google Scholar] [CrossRef]
  37. Zhou, X.; Zheng, H.B.; Xu, X.Q.; He, J.Y.; Ge, X.K.; Yao, X.; Cheng, T.; Zhu, Y.; Cao, W.X.; Tian, Y.C. Predicting grain yield in rice using multi-temporal vegetation indices from UAV-based multispectral and digital imagery. ISPRS J. Photogramm. Remote Sens. 2017, 130, 246–255. [Google Scholar] [CrossRef]
  38. Qiao, L.; Tang, W.; Gao, D.; Zhao, R.; An, L.; Li, M.; Sun, H.; Song, D. UAV-based chlorophyll content estimation by evaluating vegetation index responses under different crop coverages. Comput. Electron. Agric. 2022, 196, 106775. [Google Scholar] [CrossRef]
  39. Aslan, M.F.; Sabanci, K.; Aslan, B. Artificial Intelligence Techniques in Crop Yield Estimation Based on Sentinel-2 Data: A Comprehensive Survey. Sustainability 2024, 16, 8277. [Google Scholar] [CrossRef]
  40. Liu, T.; Wu, F.; Mou, N.; Zhu, S.; Yang, T.; Zhang, W.; Wang, H.; Wu, W.; Zhao, Y.; Sun, C. The estimation of wheat yield combined with UAV canopy spectral and volumetric data. Food Energy Secur. 2024, 13, e527. [Google Scholar] [CrossRef]
  41. Taylor, J.A.; Bates, T.R. A discussion on the significance associated with Pearson’s correlation in precision agriculture studies. Precis. Agric. 2013, 14, 558–564. [Google Scholar] [CrossRef]
  42. Kaufman, Y.J.; Tanre, D. Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE Trans. Geosci. Remote Sens. 1992, 30, 261–270. [Google Scholar] [CrossRef]
  43. Corti, M.; Cavalli, D.; Cabassi, G.; Marino Gallina, P.; Bechini, L. Does remote and proximal optical sensing success-fully estimate maize variables? A review. Eur. J. Agron. 2018, 99, 37–50. [Google Scholar] [CrossRef]
  44. Meyer, G.E.; Mehta, T.; Kocher, M.F.; Mortensen, D.A.; Samal, A. Textural imaging and discriminant analysis for distinguishingweeds for spot spraying. Trans. ASAE 1998, 41, 1189–1197. [Google Scholar] [CrossRef]
  45. Shuai, G.; Fowler, A.; Basso, B. Within-season vegetation indices and yield stability as a predictor of spatial patterns of Maize (Zea mays L.) yields. Precis. Agric. 2024, 25, 963–982. [Google Scholar] [CrossRef]
  46. Daughtry, C.S.T.; Gallo, K.P.; Goward, S.N.; Prince, S.D.; Kustas, W.P. Spectral estimates of absorbed radiation and phytomass production in corn and soybean canopies. Remote Sens. Environ. 1992, 39, 141–152. [Google Scholar] [CrossRef]
  47. Li, F.; Miao, Y.; Feng, G.; Yuan, F.; Yue, S.; Gao, X.; Liu, Y.; Liu, B.; Ustin, S.L.; Chen, X. Improving estimation of summer maize nitrogen status with red edge-based spectral vegetation indices. Field Crop Res. 2014, 157, 111–123. [Google Scholar] [CrossRef]
  48. McFeeters, S.K. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. Int. J. Remote Sens. 1996, 17, 1425–1432. [Google Scholar] [CrossRef]
  49. Wan, L.; Cen, H.; Zhu, J.; Li, Y.; Zhu, Y.; Sun, D.; Weng, H.; He, Y. Combining UAV-based vegetation indices, canopy height and canopy coverage to improve rice yield prediction under different nitrogen levels. In Proceedings of the 2019 ASABE Annual International Meeting, Boston, MA, USA, 7–10 July 2019; American Society of Agricultural and Biological Engineers: St. Joseph, MI, USA, 2019; p. 1. [Google Scholar]
  50. Pearson, R.L.; Miller, L.D. Remote mapping of standing crop biomass for estimation of the productivity of the short-grass prairie. In Proceedings of the Eighth International Symposium on Remote Sensing of Environment, Ann Arbor, MI, USA, 2–6 October 1972; p. 1355. [Google Scholar]
  51. Wang, Q.; Hao, Y. ALSTM: An attention-based long short-term memory framework for knowledge base reasoning. Neurocomputing 2020, 399, 342–351. [Google Scholar] [CrossRef]
  52. Tian, H.; Wang, P.; Tansey, K.; Han, D.; Zhang, J.; Zhang, S.; Li, H. A deep learning framework under attention mechanism for wheat yield estimation using remotely sensed indices in the Guanzhong Plain, PR China. Int. J. Appl. Earth Obs. 2021, 102, 102375. [Google Scholar] [CrossRef]
  53. Zou, H. The adaptive lasso and its oracle properties. J. Am. Stat. Assoc. 2006, 101, 1418–1429. [Google Scholar] [CrossRef]
  54. Belgiu, M.; Drăguţ, L. Random forest in remote sensing: A review of applications and future directions. ISPRS J. Photogramm. Remote Sens. 2016, 114, 24–31. [Google Scholar] [CrossRef]
  55. Landi, F.; Baraldi, L.; Cornia, M.; Cucchiara, R. Working memory connections for LSTM. Neural Netw. 2021, 144, 334–341. [Google Scholar] [CrossRef] [PubMed]
  56. Panda, S.S.; Ames, D.P.; Panigrahi, S. Application of vegetation indices for agricultural crop yield prediction using neural network techniques. Remote Sens. 2010, 2, 673–696. [Google Scholar] [CrossRef]
  57. Bala, S.K.; Islam, A.S. Correlation between potato yield and MODIS-derived vegetation indices. Int. J. Remote Sens. 2009, 30, 2491–2507. [Google Scholar] [CrossRef]
  58. Muruganantham, P.; Wibowo, S.; Grandhi, S.; Samrat, N.H.; Islam, N. A systematic literature review on crop yield prediction with deep learning and remote sensing. Remote Sens. 2022, 14, 1990. [Google Scholar] [CrossRef]
  59. Xiong, Z.; Xin, Q.; Zhang, Z.; Gui, H. A Meteorology-Driven Lstm For Predicting Time Series of Vegetation Leaf Area Index In Land Surface Model. In Proceedings of the IGARSS 2023—2023 IEEE International Geoscience and Remote Sensing Symposium, Pasadena, CA, USA, 16–21 July 2023; pp. 4915–4918. [Google Scholar]
  60. Li, P.; Jiang, X.; Jin, G.; Yu, Y.; Xie, Z. ALSTM: An attention-based lstm model for multi-scenario bandwidth prediction. In Proceedings of the 2021 IEEE 27th International Conference on Parallel and Distributed Systems (ICPADS), Beijing, China, 14–16 December 2021; pp. 98–105. [Google Scholar]
  61. Tang, L.; Zhu, Y.; Hannaway, D.; Meng, Y.; Liu, L.; Chen, L.; Cao, W. RiceGrow: A rice growth and productivity model. NJAS—Wagening. J. Life Sci. 2009, 57, 83–92. [Google Scholar] [CrossRef]
  62. Fageria, N.K. Yield physiology of rice. J. Plant Nutr. 2007, 30, 843–879. [Google Scholar] [CrossRef]
  63. Hattori, Y.; Nagai, K.; Ashikari, M. Rice growth adapting to deepwater. Curr. Opin. Plant Biol. 2011, 14, 100–105. [Google Scholar] [CrossRef]
  64. Mosleh, M.K.; Hassan, Q.K.; Chowdhury, E.H. Application of remote sensors in mapping rice area and forecasting its production: A review. Sensors 2015, 15, 769–791. [Google Scholar] [CrossRef]
  65. Huang, J.F.; Tang, S.C.; Ousama, A.I.; Wang, R.C. Rice yield estimation using remote sensing and simulation model. J. Zhejiang Univ. Sci. A 2002, 3, 461–466. [Google Scholar] [CrossRef]
  66. Yang, Q.; Shi, L.; Han, J.; Zha, Y.; Zhu, P. Deep convolutional neural networks for rice grain yield estimation at the ripening stage using UAV-based remotely sensed images. Field Crop Res. 2019, 235, 142–153. [Google Scholar] [CrossRef]
  67. Son, N.T.; Chen, C.F.; Chen, C.R.; Minh, V.Q.; Trung, N.H. A comparative analysis of multitemporal MODIS EVI and NDVI data for large-scale rice yield estimation. Agric. Forest Meteorol. 2014, 197, 52–64. [Google Scholar] [CrossRef]
  68. Yang, R.; Zhou, J.; Lu, X.; Shen, J.; Chen, H.; Chen, M.; He, Y.; Liu, F. A robust rice yield estimation framework developed by grading modeling and normalized weight decision-making strategy using UAV imaging technology. Comput. Electron. Agric. 2023, 215, 108417. [Google Scholar] [CrossRef]
Agriculture 15 00231 g001
Figure 1. The map of the study area, rice planting areas, and sampling sites.
Figure 1. The map of the study area, rice planting areas, and sampling sites.
Agriculture 15 00231 g001
Agriculture 15 00231 g002
Figure 2. The relationship between sampling point, quadrat, and pixel in actual ground survey.
Figure 2. The relationship between sampling point, quadrat, and pixel in actual ground survey.
Agriculture 15 00231 g002
Agriculture 15 00231 g003
Figure 3. Workflow of pixel-scale early yield estimation of single-season rice in Qian Gorlos County, Northeast China. The red box represents the ALSTM model proposed in this paper, and the blue box represents the baseline models.
Figure 3. Workflow of pixel-scale early yield estimation of single-season rice in Qian Gorlos County, Northeast China. The red box represents the ALSTM model proposed in this paper, and the blue box represents the baseline models.
Agriculture 15 00231 g003
Agriculture 15 00231 g004
Figure 4. The overall structure of the ALSTM model for rice yield estimation (modified from Tian et al. [52]).
Figure 4. The overall structure of the ALSTM model for rice yield estimation (modified from Tian et al. [52]).
Agriculture 15 00231 g004
Agriculture 15 00231 g005
Figure 5. Heat map of correlation coefficient |r| values. (a) Heat map of correlation coefficient between vegetation index and rice yield |r| value in five growth periods. (b) Heat map of vegetation indices related to rice yield in GP4 autocorrelation analysis.
Figure 5. Heat map of correlation coefficient |r| values. (a) Heat map of correlation coefficient between vegetation index and rice yield |r| value in five growth periods. (b) Heat map of vegetation indices related to rice yield in GP4 autocorrelation analysis.
Agriculture 15 00231 g005
Agriculture 15 00231 g006
Figure 6. Model performance and time window analysis under different time series. (a–c) are the performance of evaluation models built under different time series. (a) R2, (b) RMSE, and (c) MAE. (d) is a table for model performance evaluation. The pink color represents the estimation of the GP1-GP4 time series, the yellow color represents the estimation of the GP1-GP5 time series, and the blue color represents the estimation using the ALSTM model. All of the above are evaluation results for the validation set.
Figure 6. Model performance and time window analysis under different time series. (a–c) are the performance of evaluation models built under different time series. (a) R2, (b) RMSE, and (c) MAE. (d) is a table for model performance evaluation. The pink color represents the estimation of the GP1-GP4 time series, the yellow color represents the estimation of the GP1-GP5 time series, and the blue color represents the estimation using the ALSTM model. All of the above are evaluation results for the validation set.
Agriculture 15 00231 g006
Agriculture 15 00231 g007
Figure 7. Importance of feature variables in different rice growth periods based on attention mechanism weights.
Figure 7. Importance of feature variables in different rice growth periods based on attention mechanism weights.
Agriculture 15 00231 g007
Agriculture 15 00231 g008
Figure 8. Model performance comparison in the year 2022. (a–d) are scatterplots of measured yield vs. predicted yield of LASSO, RF, LSTM, and ALSTM.
Figure 8. Model performance comparison in the year 2022. (a–d) are scatterplots of measured yield vs. predicted yield of LASSO, RF, LSTM, and ALSTM.
Agriculture 15 00231 g008
Agriculture 15 00231 g009
Figure 9. Spatial distribution of rice estimation in 2022 by (a) ALSTM, (b) LASSO, (c) RF, and (d) LSTM.
Figure 9. Spatial distribution of rice estimation in 2022 by (a) ALSTM, (b) LASSO, (c) RF, and (d) LSTM.
Agriculture 15 00231 g009
Table 1. Growth period of single-season rice in Northeast China.
Table 1. Growth period of single-season rice in Northeast China.
Serial NumberTime IntervalRice Growth PeriodCharacteristics of the Growth Period
GP16.11–6.25Transplanting stage—Tillering stageAfter rice breeding and germinating, the seedlings were transplanted to a paddy field, and the seedlings just adapted to the paddy field environment
GP26.25–7.11Tillering stage—Jointing stageThe tillering buds emerged and protruded 1–2 cm from the base leaf axils
GP37.11–7.25Jointing stage—Heading stageThe number of rice components is fixed, and the internode of the stem rapidly extends upward, which is still the vegetative growth stage
GP47.25–8.17Heading stage—Milk ripening stageThe seeds and stems, leaves, and ears in the middle and upper parts of the plant are green. Grain inclusions are milky white serous.
GP58.17–9.16Milk ripening stage—yellow ripening stageThe rice grains hardened, and the rice began to change from vegetative growth to reproductive growth
Table 2. Spectral parameters of Sentinel-2 image bands.
Table 2. Spectral parameters of Sentinel-2 image bands.
WavebandSpatial Resolution/mCentral
Wavelength/nm		Bandwidth/nm
B2-Blue 49065
B3-Green 56035
B4-Red1066530
B8-NIR 842115
B5-Red edge 70515
B6-Red edge 74015
B7-Edge of the NIR plateau2078320
B8a-Narrow NIR 86520
B11-SWIR 161090
B12-SWIR 2190180
B1-Coastal aerosol 44320
B9-Water Vapour6094520
B10-Cirrus 137530
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Li, J.; Xie, Y.; Liu, L.; Song, K.; Zhu, B. Long Short-Term Memory Neural Network with Attention Mechanism for Rice Yield Early Estimation in Qian Gorlos County, Northeast China. Agriculture 2025, 15, 231. https://doi.org/10.3390/agriculture15030231

AMA Style

Li J, Xie Y, Liu L, Song K, Zhu B. Long Short-Term Memory Neural Network with Attention Mechanism for Rice Yield Early Estimation in Qian Gorlos County, Northeast China. Agriculture. 2025; 15(3):231. https://doi.org/10.3390/agriculture15030231

Chicago/Turabian Style

Li, Jian, Yichen Xie, Lushi Liu, Kaishan Song, and Bingxue Zhu. 2025. "Long Short-Term Memory Neural Network with Attention Mechanism for Rice Yield Early Estimation in Qian Gorlos County, Northeast China" Agriculture 15, no. 3: 231. https://doi.org/10.3390/agriculture15030231

APA Style

Li, J., Xie, Y., Liu, L., Song, K., & Zhu, B. (2025). Long Short-Term Memory Neural Network with Attention Mechanism for Rice Yield Early Estimation in Qian Gorlos County, Northeast China. Agriculture, 15(3), 231. https://doi.org/10.3390/agriculture15030231

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop