Next Article in Journal
Direct Estimation of Forest Aboveground Biomass from UAV LiDAR and RGB Observations in Forest Stands with Various Tree Densities
Previous Article in Journal
Assessing Intraspecific Variation of Tree Species Based on Sentinel-2 Vegetation Indices Across Space and Time
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 12
10.3390/rs17122095
remotesensing-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

Crop Classification Using Time-Series Sentinel-1 SAR Data: A Comparison of LSTM, GRU, and TCN with Attention

by
Yuta Tsuchiya
and
Rei Sonobe
*
Graduate School of Science and Technology, Shizuoka University, Shizuoka 422-8529, Japan
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(12), 2095; https://doi.org/10.3390/rs17122095
Submission received: 12 May 2025 / Revised: 15 June 2025 / Accepted: 16 June 2025 / Published: 18 June 2025
(This article belongs to the Special Issue Crop Classification Using Synthetic Aperture Radar and Optical Imagery)
Browse Figures
Versions Notes

Abstract

This study investigates the performance of temporal deep learning models with attention mechanisms for crop classification using Sentinel-1 C-band synthetic aperture radar (C-SAR) data. A time series of 16 scenes, acquired at 12-day intervals from 25 April to 22 October 2024, was used to classify six crop types: beans, beetroot, grassland, maize, potato, and winter wheat. Three temporal models—long short-term memory (LSTM), bidirectional gated recurrent unit (Bi-GRU), and temporal convolutional network (TCN)—were evaluated with and without an attention mechanism. All model configurations achieved accuracies above 83%, demonstrating the potential of Sentinel-1 SAR data for reliable, weather-independent crop classification. The TCN with attention model achieved the highest accuracy of 85.7%, significantly outperforming the baseline. LSTM also showed improved accuracy when combined with attention, whereas Bi-GRU did not benefit from the attention mechanism. These results highlight the effectiveness of combining temporal deep learning models with attention mechanisms to enhance crop classification using Sentinel-1 SAR time-series data. This study further confirms that freely available, regularly acquired Sentinel-1 observations are well-suited for robust crop mapping under diverse environmental conditions.

1. Introduction

Accurate and timely crop type mapping is essential for effective agricultural monitoring, food security planning, and sustainable land management [1,2,3]. As global challenges such as climate change, population growth, and resource scarcity intensify, the demand for reliable agricultural land use information becomes increasingly important [4,5]. Crop maps are used by governments, researchers, and agribusinesses to assess production potential, plan harvest logistics, evaluate environmental impacts, and implement agricultural policies [6,7,8]. Recently, remote sensing technologies have become indispensable for large-scale crop classification [9]. These systems provide non-invasive, repeatable, and cost-effective means of monitoring vast agricultural areas with spatial and temporal consistency. Optical sensors, such as those on Landsat or Sentinel-2 platforms, have been widely used to track vegetation dynamics, phenological stages, and crop health through reflectance patterns in visible and near-infrared wavelengths [10,11,12]. However, optical imagery is highly sensitive to cloud cover, limiting its effectiveness in regions with frequent or persistent clouds—especially during the growing season when timely observations are most needed.
Synthetic aperture radar (SAR), in contrast, offers significant advantages in challenging conditions [13,14]. SAR systems, including those onboard Sentinel-1, operate in the microwave spectrum, enabling the capture of high-resolution imagery regardless of weather or lighting conditions [15,16,17]. This all-weather, day-and-night capability makes SAR particularly valuable for consistent agricultural monitoring. Moreover, SAR signals interact with both the structural and moisture characteristics of vegetation, providing information that complements or even surpasses optical data in certain agricultural applications [15,18]. In regions with frequent rainfall during the crop growing season—such as the Asian monsoon area—acquiring cloud-free optical images is a major challenge [19]. In contrast, SAR satellites enable consistent observation regardless of weather conditions. Therefore, we attempted crop classification using only SAR data from Sentinel-1.
Recent advancements in computational power and machine learning have further enhanced the utility of SAR data for crop classification [20,21,22]. Traditional machine learning algorithms, such as random forest and support vector machines, have been successfully applied to classify crops using SAR features [20,21,23,24,25]. However, these models often fail to fully capture the temporal dynamics that characterize crop growth cycles. As a result, deep learning models—particularly those designed for sequential data—have gained attention for their ability to model complex temporal patterns.
Recurrent neural networks (RNNs), including variants such as long short-term memory (LSTM) and gated recurrent units (GRUs), are foundational for sequential data analysis due to their ability to capture temporal dependencies and are capable of capturing sequential dependencies and learning noise-tolerant temporal patterns from raw data, they offer the advantage of being easy to apply [26,27]. These architectures are particularly effective for time-series analysis, natural language processing, and spatiotemporal tasks [27,28]. Despite their ability to process sequences, traditional RNNs struggle with long-term dependencies, especially when input sequences are lengthy or when the most informative parts of the sequence are unevenly distributed across time. Attention mechanisms have emerged as a solution to this challenge [29,30]. Originally introduced in neural machine translation, attention mechanisms allow models to focus on relevant parts of the input sequence during each output step [31]. Unlike standard RNNs, which compress all past information into a fixed-length context vector, attention assigns dynamic weights to each time step based on its relevance to the task [32,33]. In crop classification using SAR time-series data, this selective focus is particularly valuable. Crops exhibit distinct phenological characteristics at various growth stages. Attention mechanisms enable the model to prioritize time periods that contribute more to class discrimination—such as flowering or peak biomass stages—while down-weighting less informative periods, such as early growth or post-harvest phases. This improves classification accuracy and enhances model interpretability by revealing which temporal features are most influential. Furthermore, attention helps mitigate the limitations of recurrent structures in capturing long-range dependencies [34,35]. While LSTM and GRU networks introduce gating mechanisms to preserve information over time, they still face challenges with long sequences due to issues such as vanishing gradients. Attention solves this by allowing the model to bypass irrelevant intermediate steps and focus directly on key temporal features, resulting in more stable and efficient training.
Another practical advantage of attention-enhanced RNNs is their adaptability across architectures [27,34,36]. Attention can be applied to unidirectional or bidirectional RNNs and seamlessly integrated with hybrid models such as temporal convolutional networks (TCNs), further boosting performance [37,38,39]. Empirical studies consistently show that attention-augmented RNNs outperform non-attention models in various time-series classification tasks, including human activity recognition, speech processing, and agricultural monitoring. Additionally, attention mechanisms foster greater interpretability in AI, a growing need in fields such as remote sensing and environmental monitoring [40,41]. The attention weights can be visualized to identify the temporal contribution of input features, helping domain experts understand model decisions and validate its behavior against known crop growth patterns.
In conclusion, integrating attention mechanisms into RNNs significantly enhances their performance, robustness, and interpretability. For agricultural applications using temporal SAR data, this combination offers a promising approach for accurate and explainable crop classification across diverse environmental conditions. When applied to dense time-series data, such as Sentinel-1’s 12-day intervals, these models can effectively distinguish between crop types based on their unique seasonal growth patterns [41]. Given its free access, extensive coverage, and regular revisit cycle, Sentinel-1 C-band SAR is an ideal source of data for temporal modeling in agricultural applications. By combining Sentinel-1 time-series observations with temporal deep learning models, researchers can achieve robust crop classification—even in cloudy or data-limited environments.

2. Materials and Methods

2.1. Study Area

This study was conducted in the Tokachi Plain, located in southeastern Hokkaido, Japan. The study area spans from 142°55′12″E to 143°05′51″E longitude and from 42°52′48″N to 43°02′42″N latitude (Figure 1). The region has a continental humid climate, characterized by warm summers and cold, snowy winters, with an average annual temperature of approximately 6 °C and annual precipitation of about 920 mm. These climatic conditions, combined with fertile volcanic ash soils and a well-developed irrigation system, make the Tokachi Plain one of Japan’s most productive agricultural zones. The area supports a wide variety of crops, benefiting from long daylight hours during the growing season. This study focused on several key crops representative of Tokachi’s mixed farming system: soybeans, azuki beans, kidney beans, maize, sugar beet, potato, and forage grasses. Beans and maize are typically sown in mid-May, while sugar beets and potatoes are planted between late April and early May. Perennial and winter grasses, such as timothy, orchard grass, and winter wheat, are sown in the preceding year. Harvesting times vary by crop: beans are harvested from late September to early November, sugar beets in November, and potatoes from late August through September. Winter wheat is usually harvested from late July to early August. Forage grasses are harvested twice annually—once from late June to early July and again in late August—reflecting their importance in the region’s dairy and livestock sectors.

2.2. Reference Data

Information on the locations of agricultural fields was downloaded from https://open.fude.maff.go.jp/ and used in this study. The number of agricultural fields used for each crop was as follows: 533 for beans, 274 for beetroot, 720 for grass, 451 for maize, 188 for potatoes, and 496 for wheat. The field areas (in hectares) for each crop are summarized in Table 1. The reference data were created based on crops cultivated between June and August 2024, and no changes in the cultivated crops occurred during the period when the Sentinel-1 data were acquired.

2.3. Satellite Data

Regular observations by Sentinel-1A were conducted over the study area, and 16 Sentinel-1A/C-SAR scenes were acquired at 12-day intervals from 25 April to 22 October 2024 (summarized in Table 2). Sentinel-1 offers several observation modes to support a wide range of applications, including Stripmap (SM), Interferometric Wide Swath (IW), Extra Wide Swath (EW), and Wave (WV) modes. Each mode differs in spatial resolution and swath width, designed for specific monitoring needs. Among these, the IW mode is the default acquisition mode over land and is optimized for land cover mapping, agriculture, and interferometric analysis. For this study, data acquired in the IW mode were used. This mode provides dual-polarization (VH and VV) SAR imagery with a 250 km swath width and a spatial resolution of approximately 10 m, making it ideal for capturing detailed agricultural field characteristics across large areas such as the Tokachi Plain. Except for the orbit used in this study, a single observation could not fully cover the entire study area. Although it is possible to obtain data from other orbits, this would require mosaic processing to merge images taken on different dates. To avoid potential differences in classification accuracy caused by variations between images from different observation dates, we used only data from an orbit that could completely cover the study area in a single observation.
The data were obtained as Ground Range Detected (GRD) products from the Copernicus Conventional Data Access Hubs (https://www.copernicus.eu/en/access-data/conventional-data-access-hubs (accessed on 22 April 2025)). Sentinel-1 is a C-band SAR mission developed by the European Space Agency (ESA) under the Copernicus program. It consists of two polar-orbiting satellites, Sentinel-1A and Sentinel-1B, which provide all-weather, day-and-night radar imaging capabilities, with a revisit time of six days when both satellites are operational. The GRD product provides pre-processed SAR backscatter data projected to ground range, suitable for land and vegetation monitoring. Using ESA’s Sentinel Application Platform (SNAP, version 11.0.1), radiometric correction of the satellite data, speckle noise reduction using the Lee Sigma filter, and geometric correction were performed. For the geometric correction, the 10 m mesh digital elevation model (DEM) produced by the Geospatial Information Authority of Japan (GSI) and the Earth Gravitational Model 2008 (EGM2008) were used.

2.4. Classification Model Structure

LSTM networks are a specialized form of RNNs designed to address the vanishing gradient problem that commonly occurs in traditional RNNs. LSTMs introduce a memory cell and three types of gates—input, forget, and output—that regulate the flow of information [42,43]. These networks are effective at capturing long-term dependencies in sequential data, making them particularly well-suited for tasks like time-series classification. In practical applications, an LSTM layer typically consists of 64 or 128 hidden units, although this can vary depending on the complexity of the data. The activation function within LSTM cells is usually the hyperbolic tangent (tanh) for the internal state and sigmoid for the gates. Although LSTMs tend to require more training time due to their sequential nature and complex parameter structure, they generally perform well on data with long and structured temporal dependencies [44].
A bidirectional GRU (Bi-GRU) is an advanced sequence modeling architecture that processes input data in both forward and backward temporal directions [45]. Unlike a standard GRU, which processes a sequence from the first to the last time step, a Bi-GRU uses two parallel GRU layers—one reading the sequence in the forward direction and the other in reverse [46]. This dual processing enables the model to capture both past and future contextual information at each time step, significantly enriching its understanding of the sequence dynamics. For time-series classification tasks, such as crop type recognition using multispectral satellite imagery, this bidirectional approach is particularly advantageous. For example, to classify a specific time point, it is helpful to know not only the prior reflectance pattern but also how the spectral signals evolve in subsequent observations. A unidirectional GRU, which only considers past information, lacks this capability. In contrast, the Bi-GRU processes the input sequence in both directions and concatenates the two hidden states at each time step, creating a more expressive and temporally comprehensive representation. In practical applications, Bi-GRUs have shown strong performance improvements over standard GRUs, especially when combined with attention mechanisms [47]. The attention layer assigns context-sensitive weights to each time step, and with the enhanced features from the Bi-GRU, these weights are calculated over a richer temporal embedding [48]. This improves both classification accuracy and interpretability, because it allows researchers to identify which time periods contributed most to the prediction. Overall, adopting a Bi-GRU architecture is beneficial when both past and future information are crucial for understanding the signal, which is common in temporal phenomena such as crop phenology, weather-dependent growth cycles, or disease progression in biological sequences.
TCNs offer an alternative to recurrent models by using dilated 1D convolutions to model sequences [49]. These convolutions apply increasing dilation rates—typically powers of two (i.e., 1, 2, 4, etc.)—to exponentially enlarge the receptive field of the network without significantly increasing depth. A typical TCN uses 32 to 64 filters per convolutional layer (64 in this study), a kernel size of 3 to 5 (3 in this study), and residual connections to help with training deep architectures [50,51]. The activation function is usually the rectified linear unit, and dropout may be applied for regularization. Because TCNs process data in parallel rather than sequentially, they are generally faster to train and scale better for large datasets. They are particularly useful for sequences that are fixed length and well-aligned, such as multi-temporal satellite imagery with consistent acquisition intervals.
When combined with these models, attention mechanisms add an interpretable layer that dynamically assigns weights to each time step based on its relevance to the output [52,53]. The attention mechanism typically introduces parameters such as the weight matrix W∈Rd×1 and a bias term b∈RT×1, where d is the dimensionality of the features and T is the number of time steps. These parameters are learned during training. The result is a set of attention weights (or scores) that sum to one, which are applied to the sequence to form a context vector summarizing the most relevant parts of the input. Attention improves both model performance and interpretability, especially in remote sensing, where it can highlight critical time points (e.g., key phenological stages of crops). Attention is particularly effective in models that process sequences of 10 or more steps (50 steps in this study), because it mitigates the tendency to overemphasize recent time points.
LSTM, GRU, and TCN are capable of capturing sequential dependencies and learning noise-tolerant temporal patterns from the raw data, and the frequent acquisition of Sentinel-1 ensures relatively dense temporal coverage, which helps the models detect consistent phenological signals despite the presence of speckle noise or minor temporal irregularities. Therefore, time series interpolation and smoothing algorithms were not used. Cross-validation was used to evaluate performance for each hyperparameter setting, and the combination yielding the lowest loss was selected. Additionally, early stopping and regularization techniques were applied to prevent overfitting during training.

2.5. Classification Process

Figure 2 illustrates the processing procedure. Polygon data was used instead of raster data to focus on changes in scattering patterns between crops. To minimize the impact of mixed pixels (i.e., mixed spectral signals near field boundaries), a 10 m inward buffer was applied to the polygon data containing crop types and field locations. The average values of VH and VV polarization gamma nought backscatter, as well as the polarization ratio (VH/VV), were then extracted for each field from all 16 Sentinel-1 C-SAR scenes using these buffered polygons. For crop classification, a 10-fold cross-validation approach was employed, and the data were divided so that the proportion of the six classes was equal across the ten groups, with 90% of the data used for training and 10% for testing. Within the training set, 20% of the data was further reserved as validation data for model parameter tuning. All processing and modeling were conducted in the Google Colab environment. Classification performance was evaluated using the following metrics: overall accuracy (OA), kappa, producer’s accuracy (PA), user’s accuracy (UA), and the F1 score. Additionally, the McNemar test was used to assess the statistical significance of differences in the classification results.

3. Results

3.1. Temporal Changes in Backscatter Coefficients

The time-series variations in backscatter coefficients for each crop are shown in Figure 3. The temporal dynamics of VH and VV polarization backscatter coefficients reflect both the scattering characteristics of C-band microwaves and the phenological stages of each crop. By 25 April, most fields had not yet been sown or transplanted, and the backscatter signals were primarily influenced by surface conditions, such as residual plant material from the previous growing season and high soil moisture from snowmelt. In contrast, perennial grasses and winter wheat, which had been cultivated since the previous year, showed backscatter mainly from small, emerging plant structures, with less contribution from the soil surface. As a result, the VH backscatter coefficients exhibited little variation among the six crop types, with average values ranging from −16 to −17 dB. However, the VV backscatter for perennial grasses and winter wheat was noticeably lower than for other crops. Following this initial period, both VV and VH backscatter for perennial grasses and winter wheat dropped significantly on 19 May, likely due to early growth stages with sparse canopy coverage, but then increased again by 31 May as the canopy developed. For beans, a distinct decrease in backscatter occurred on 12 June, possibly due to canopy closure, which reduces volume scattering. From then until 6 July, backscatter values generally increased as crop biomass grew. On 18 July, backscatter for perennial grasses and winter wheat declined sharply due to harvesting. For the remaining crops, backscatter continued to rise until around 23 August, reflecting ongoing growth and canopy development. After this peak, backscatter began to decrease as the crops entered their ripening stages and approached harvest. The rate and timing of this decline varied depending on each crop’s maturity and harvest schedule, with most fields showing a decrease in backscatter by 18 September.

3.2. Crop Classification Accuracies

The classification accuracies obtained through 10-fold cross-validation and the results of the McNemar test are summarized in Table 3 and Table 4. Among the models evaluated—LSTM, Bi-GRU, TCN, LSTM + Attention, Bi-GRU + Attention, and TCN + Attention—the TCN with attention model achieved the highest OA of 85.7%. Adding an attention mechanism improved the classification accuracy for both the LSTM and TCN models, but did not improve performance for the Bi-GRU model. The TCN with attention model showed improved performance in terms of PA, UA, and F1 score for all crops except beetroot. In contrast, the LSTM model showed improvements in all three metrics (PA, UA, and F1) only for beans.
The McNemar test revealed no statistically significant differences (p > 0.1) between the following model pairs: LSTM and LSTM + Attention, TCN and Bi-GRU + Attention, and Bi-GRU + Attention and LSTM + Attention. However, a significant difference at the 10% level was observed between LSTM and Bi-GRU. Statistically significant differences at the 0.1% level were found between the following: LSTM and TCN, TCN and LSTM + Attention, and LSTM + Attention and TCN + Attention. For all other model combinations, significant differences were observed at the 1% level.
Figure 4 shows the classification maps of all six schemes.

3.3. Importance of Each Variable

Several patterns were observed when the importance (attention weight) of the backscattering coefficients for VH and VV polarizations, along with the polarization ratio at each of the 16 time points, was visualized as a heatmap (Figure 5). Negative values were due to the gamma naught values sometimes being negative, which reflects the signs of the original features. Across all models, the importance of data from the later stages of crop growth was higher, particularly for data acquired after 28 September, which consistently showed high importance for VH, VV, and the polarization ratio. Notably, the polarization ratio on 23 August had significant importance, acting as a valuable biophysical indicator. In contrast, attention weights before 24 June were lower, suggesting that key features had not yet emerged during the early growth stages. A distinctive feature of the TCN model was the heightened importance of the polarization ratio on 23 August and the data from 4 September.

4. Discussion

Several studies have utilized Sentinel-1 SAR data in combination with optical sensors for crop type classification, achieving overall accuracies ranging from 70% to 85%, depending on factors such as crop classes, study region, and modeling approaches. For example, Inglada et al. [54] used a random forest classifier to map major crops, reporting an OA of approximately 80%. Veloso et al. [55] applied time-series Sentinel-1 data for crop classification in southwestern France, achieving accuracies between 70% and 85%. In this study, an OA of 85.7% was achieved using a TCN with an attention model to classify six major crop types in the Tokachi Plain, Hokkaido. This result is comparable to, or slightly greater than, previous studies, despite the challenge of distinguishing between crops with similar growth patterns, such as beans and maize. Factors contributing to this strong performance include the dense temporal sampling of 16 Sentinel-1 scenes, the use of both VH and VV backscatter coefficients and the polarization ratio as input features, and the adoption of advanced deep learning models that effectively capture temporal dynamics.
The relatively clear seasonal growth patterns in the Tokachi Plain, along with the use of attention mechanisms to focus on key phenological periods—particularly during the late growing season—further improved classification accuracy compared to conventional methods. These findings highlight the potential of integrating high-temporal-resolution SAR data with modern deep learning techniques, especially those employing attention mechanisms, for significantly improving crop classification performance. This approach positions the study at the forefront of SAR-based agricultural monitoring.
In this study, the polarization ratio (VH/VV) was more effective for crop classification than using VH or VV backscatter coefficients individually. This is consistent with previous findings in SAR-based agricultural monitoring. The polarization ratio enhances sensitivity to vegetation structure and moisture variations while reducing the influence of soil surface conditions [56,57]. While VV backscatter is strongly influenced by surface roughness and soil moisture, and VH backscatter is primarily associated with volume scattering from vegetation, the VH/VV ratio normalizes these effects and emphasizes crop-specific canopy characteristics. Bousbih et al. [56] found that the VH/VV ratio was more strongly correlated with vegetation indices such as leaf area index than either VH or VV alone, particularly during the mid to late growing season. Similarly, Abdikan et al. [57] showed that polarization ratios improved the discrimination between different crop types in Turkey, especially during peak biomass periods. The attention weights analysis in this study confirmed that the polarization ratio became increasingly important from 23 August onward, coinciding with the peak growth and maturity phases for most crops. This timing corresponds to when vegetation structures are most differentiated among crops and when soil background effects are minimized by dense canopy cover. Thus, the polarization ratio effectively captured subtle differences in crop biophysical properties during critical phenological stages, contributing significantly to classification performance.
These findings suggest that incorporating polarization ratios can enhance the robustness of crop classification models by improving sensitivity to crop-specific growth dynamics while simultaneously mitigating the confounding effects of soil moisture and surface roughness, which can affect SAR backscatter signals.
In classifying six crop types, the TCN outperformed both LSTM and Bi-GRU. This superior performance can be attributed to TCN’s ability to capture long-range temporal dependencies more effectively through its hierarchical, dilated convolutional structure. Unlike recurrent models, TCN processes all time steps simultaneously and avoids the vanishing gradient problem, which often limits the ability of RNN-based models such as LSTM and Bi-GRU to learn from long time-series data. Given the gradual and seasonal changes in crop development reflected in the SAR time series, TCN’s structure enabled it to model these patterns more efficiently and accurately. The addition of an attention mechanism improved classification accuracy for both LSTM and TCN. In LSTM, attention helped overcome the limitation of sequential memory by allowing the model to focus on important time points, particularly during key growth stages. For TCN, the attention mechanism further enhanced its ability to prioritize critical periods, sharpening its focus on informative features throughout the time series without losing its inherent parallel processing advantage. However, in the case of Bi-GRU, the addition of attention did not result in noticeable improvements. This could be because Bi-GRU already processes the sequence in both forward and backward directions, inherently capturing important temporal dependencies. Thus, the marginal benefit of adding an attention layer to Bi-GRU was limited. Furthermore, Bi-GRU may not have been as effective as TCN in handling long time sequences with subtle variations, meaning that attention alone was insufficient to compensate for its architectural limitations in this task. Overall, these results suggest that model architecture and feature extraction strategies must be carefully selected based on the nature of the temporal patterns in the data. Attention mechanisms are particularly beneficial for models that process sequences sequentially (e.g., LSTM) or in a convolutional manner (e.g., TCN), but not necessarily for bidirectional recurrent models such as Bi-GRU.
The importance of the data acquired after 28 September remained consistently high, indicating that the backscatter characteristics captured during the later growth stages were crucial for distinguishing between the six crop types. The high attention weights assigned to the data after 28 September for VH, VV, and the polarization ratio suggest that the structural and moisture-related changes occurring during the ripening and pre-harvest phases were key features for classification. This aligns with the understanding that microwave backscatter is sensitive to crop biomass, structure, and moisture conditions, all of which vary significantly as crops mature. The importance of the polarization ratio around 23 August suggests that it effectively captured biophysical properties of the crops, possibly related to canopy structure or water content during the late vegetative or early reproductive stages. The polarization ratio appears to be a strong biophysical indicator, significantly contributing to the model’s ability to differentiate crops at this stage. The relatively low attention weights before 24 June imply that the early growth stages, when plants are small and field conditions are relatively homogeneous, provided limited distinctive information for crop classification. This aligns with the expectation that microwave signals during early growth primarily reflect soil conditions and sparse vegetation, making it difficult to differentiate between crop types at that stage.
Among the three models, TCN uniquely placed higher importance on the polarization ratio on 23 August and the data from 4 September. This suggests that TCN may be more effective at capturing subtle temporal patterns in backscatter evolution, benefiting from its ability to model longer temporal dependencies compared to recurrent models such as LSTM and Bi-GRU. Overall, these findings emphasize the importance of selecting appropriate observation periods and features (e.g., polarization ratio) when designing SAR-based crop classification models, especially for operational monitoring applications.
The findings of this study underscore the effectiveness of time-series C-band SAR data, particularly from Sentinel-1, in discriminating between multiple crop types. Building on these results, it is essential to examine the capabilities of both current and upcoming SAR missions to assess their potential for enhancing crop classification through temporal analysis. Sentinel-1, operated by the ESA, remains a foundational platform for SAR-based agricultural monitoring. Its consistent revisit cycle, dual-polarization (VV and VH), and high spatial resolution of 10 m have made it indispensable for time-series analyses across large agricultural areas. As demonstrated in this study, temporal backscatter dynamics captured by Sentinel-1 provide valuable biophysical insights related to crop growth, harvest, and senescence stages.
In this study, we conducted 10-fold cross-validation on data from 2662 fields to classify and evaluate six crop types. However, there was an imbalance in the number of samples per crop type—for example, pasture accounted for 720 fields, while potato fields numbered only 188. Potato fields are characterized by the presence of ridges, which differ from fields of other crops. This characteristic leads to stronger direct and double-bounce scattering components. As a result, the classification accuracy for potatoes reached an F1-score of 70–75%, but the model was more prone to overfitting compared to other crops. The use of attention mechanisms in this study aimed to mitigate the effects of overfitting, and for the TCN model, we confirmed the effectiveness of attention in potato classification. However, it did not surpass the performance of the Bi-GRU model without attention in classifying potatoes. Even for crops with distinctive scattering patterns, a limited number of samples resulted in persistent uncertainty.

5. Conclusions

This study evaluated the classification of six major crop types in the Tokachi Plain, Hokkaido, using 16 temporal scenes of Sentinel-1 C-SAR data (VH, VV backscatter coefficients, and polarization ratio) acquired at 12-day intervals. Three deep learning models—LSTM, Bi-GRU, and TCN—were compared with and without an attention mechanism. The results showed that the TCN with attention model achieved the highest OA (85.7%) among the models tested. The addition of attention improved the performance of both the LSTM and TCN models by enhancing their ability to focus on key temporal features, particularly during critical phenological stages. However, the attention mechanism did not significantly improve the performance of the Bi-GRU model, possibly because Bi-GRU already captures temporal dependencies through its bidirectional structure.
Analysis of the attention weights revealed that the later stages of crop growth, particularly after 28 September, provided highly discriminative features across all models and polarization types. The polarization ratio around 23 August was also identified as a valuable biophysical indicator for crop differentiation. Early-season data, before 24 June, contributed less to classification, likely due to limited crop development during that period. Overall, the findings emphasize the effectiveness of combining TCN with an attention mechanism for time-series crop classification using SAR data.

Author Contributions

R.S. and Y.T. conducted the data analysis and wrote the manuscript. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by JSPS KAKENHI, grant number 25K09353.

Data Availability Statement

Satellite data were downloaded from https://www.copernicus.eu/en/access-data/conventional-data-access-hubs (accessed on 22 April 2025), and the base GIS data were downloaded from https://open.fude.maff.go.jp/ (accessed on 22 April 2025).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Victor, N.; Maddikunta, P.K.R.; Mary, D.R.K.; Murugan, R.; Chengoden, R.; Gadekallu, T.R.; Rakesh, N.; Zhu, Y.; Paek, J. Remote Sensing for Agriculture in the Era of Industry 5.0-A Survey. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2024, 17, 5920–5945. [Google Scholar] [CrossRef]
  2. Pokhariyal, S.; Patel, N.R.; Govind, A. Machine Learning-Driven Remote Sensing Applications for Agriculture in India—A Systematic Review. Agronomy 2023, 13, 2302. [Google Scholar] [CrossRef]
  3. Komp, K.U.; Haub, C. Global Monitoring for Food Security and Sustainable Land Management—Recent Advances of Remote Sensing Applications to African and Siberian Show Cases. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2012, XXXIX-B8, 256–270. [Google Scholar] [CrossRef]
  4. Lambin, E.F.; Rounsevell, M.D.A.; Geist, H.J. Are Agricultural Land-Use Models Able to Predict Changes in Land-Use Intensity? Agric. Ecosyst. Env. 2000, 82, 321–331. [Google Scholar] [CrossRef]
  5. Rabbinge, R.; Van Diepen, C.A. Changes in Agriculture and Land Use in Europe. Eur. J. Agron. 2000, 13, 85–99. [Google Scholar] [CrossRef]
  6. Kusumastuti, R.D.; Van Donk, D.P.; Teunter, R. Crop-Related Harvesting and Processing Planning: A Review. Int. J. Prod. Econ. 2016, 174, 76–92. [Google Scholar] [CrossRef]
  7. Sharma, R.; Kamble, S.S.; Gunasekaran, A. Big GIS Analytics Framework for Agriculture Supply Chains: A Literature Review Identifying the Current Trends and Future Perspectives. Comput. Electron. Agric. 2018, 155, 103–120. [Google Scholar] [CrossRef]
  8. Rasheed, N.; Khan, S.A.; Hassan, A.; Safdar, S. A Decision Support Framework for National Crop Production Planning. IEEE Access 2021, 9, 133402–133415. [Google Scholar] [CrossRef]
  9. Atzberger, C. Advances in Remote Sensing of Agriculture: Context Description, Existing Operational Monitoring Systems and Major Information Needs. Remote Sens. 2013, 5, 949–981. [Google Scholar] [CrossRef]
  10. Griffiths, P.; Nendel, C.; Hostert, P. Intra-Annual Reflectance Composites from Sentinel-2 and Landsat for National-Scale Crop and Land Cover Mapping. Remote Sens. Env. 2019, 220, 135–151. [Google Scholar] [CrossRef]
  11. Chaves, M.E.D.; Picoli, M.C.A.; Sanches, I.D. Recent Applications of Landsat 8/OLI and Sentinel-2/MSI for Land Use and Land Cover Mapping: A Systematic Review. Remote Sens. 2020, 12, 3062. [Google Scholar] [CrossRef]
  12. Misra, G.; Cawkwell, F.; Wingler, A. Status of Phenological Research Using Sentinel-2 Data: A Review. Remote Sens. 2020, 12, 2760. [Google Scholar] [CrossRef]
  13. Amitrano, D.; Di Martino, G.; Guida, R.; Iervolino, P.; Iodice, A.; Papa, M.N.; Riccio, D.; Ruello, G. Earth Environmental Monitoring Using Multi-Temporal Synthetic Aperture Radar: A Critical Review of Selected Applications. Remote Sens. 2021, 13, 604. [Google Scholar] [CrossRef]
  14. Akbari, V.; Doulgeris, A.P.; Eltoft, T. Monitoring Glacier Changes Using Multitemporal Multipolarization SAR Images. IEEE Trans. Geosci. Remote Sens. 2014, 52, 3729–3741. [Google Scholar] [CrossRef]
  15. Hussain, M.; Keshavarz, R.; Shariati, N. Advancements in Agricultural Microwave Remote Sensing: An Overview From Indoor to Space-Borne Platforms. IEEE Trans. Instrum. Meas. 2025, 74, 5502622. [Google Scholar] [CrossRef]
  16. García, L.P.; Furano, G.; Ghiglione, M.; Zancan, V.; Imbembo, E.; Ilioudis, C.; Clemente, C.; Trucco, P. Advancements in Onboard Processing of Synthetic Aperture Radar (SAR) Data: Enhancing Efficiency and Real-Time Capabilities. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2024, 17, 16625–16645. [Google Scholar] [CrossRef]
  17. Majidi Nezhad, M.; Nastasi, B.; Groppi, D.; Lamagna, M.; Piras, G.; Astiaso Garcia, D. Green Energy Sources Assessment Using Sentinel-1 Satellite Remote Sensing. Front. Energy Res. 2021, 9, 649305. [Google Scholar] [CrossRef]
  18. Wang, L.; Marzahn, P.; Bernier, M.; Jacome, A.; Poulin, J.; Ludwig, R. Comparison of TerraSAR-X and ALOS PALSAR Differential Interferometry with Multisource DEMs for Monitoring Ground Displacement in a Discontinuous Permafrost Region. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2017, 10, 4074–4093. [Google Scholar] [CrossRef]
  19. Sonobe, R.; Tani, H.; Zadi, M.; Shimamura, H.; Mochizuki, K. Improving Crop Classification Accuracy Using Sentinel-1 C-SAR Data and GAN-Generated Optical Images. J. Appl. Remote Sens. 2025, 19, 024505. [Google Scholar] [CrossRef]
  20. Sonobe, R.; Tani, H.; Wang, X.; Kobayashi, N.; Shimamura, H. Random Forest Classification of Crop Type Using Multioral TerraSAR-X Dual-Polarimetric Data. Remote Sens. Lett. 2014, 5, 157–164. [Google Scholar] [CrossRef]
  21. Sonobe, R.; Tani, H.; Wang, X.; Kobayashi, N.; Shimamura, H. Discrimination of Crop Types with TerraSAR-X-Derived Information. Phys. Chem. Earth 2015, 83–84, 2–13. [Google Scholar] [CrossRef]
  22. Sonobe, R.; Yamaya, Y.; Tani, H.; Wang, X.; Kobayashi, N.; Mochizuki, K.-I. Assessing the Suitability of Data from Sentinel-1A and 2A for Crop Classification. GIsci Remote Sens. 2017, 54, 918–938. [Google Scholar] [CrossRef]
  23. Sonobe, R.; Tani, H.; Wang, X.; Kobayashi, N.; Shimamura, H. Parameter Tuning in the Support Vector Machine and Random Forest and Their Performances in Cross- and Same-Year Crop Classification Using TerraSAR-X. Int. J. Remote Sens. 2014, 35, 7898–7909. [Google Scholar] [CrossRef]
  24. Sonobe, R. Combining ASNARO-2 XSAR HH and Sentinel-1 C-SAR VH/VV Polarization Data for Improved Crop Mapping. Remote Sens. 2019, 11, 1920. [Google Scholar] [CrossRef]
  25. Sonobe, R.; Tani, H.; Shimamura, H.; Mochizuki, K. Addition of Fake Imagery Generated by Generative Adversarial Networks for Improving Crop Classification. Adv. Space Res. 2024, 74, 2901–2914. [Google Scholar] [CrossRef]
  26. Waqas, M.; Humphries, U.W. A Critical Review of RNN and LSTM Variants in Hydrological Time Series Predictions. MethodsX 2024, 13, 102946. [Google Scholar] [CrossRef]
  27. Malhotra, R.; Singh, P. Recent Advances in Deep Learning Models: A Systematic Literature Review. Multimed. Tools Appl. 2023, 82, 44977–45060. [Google Scholar] [CrossRef]
  28. Ouma, Y.O.; Cheruyot, R.; Wachera, A.N. Rainfall and Runoff Time-Series Trend Analysis Using LSTM Recurrent Neural Network and Wavelet Neural Network with Satellite-Based Meteorological Data: Case Study of Nzoia Hydrologic Basin. Complex. Intell. Syst. 2022, 8, 213–236. [Google Scholar] [CrossRef]
  29. Soydaner, D. Attention Mechanism in Neural Networks: Where It Comes and Where It Goes. Neural Comput. Appl. 2022, 34, 13371–13385. [Google Scholar] [CrossRef]
  30. Hernández, A.; Amigó, J.M. Attention Mechanisms and Their Applications to Complex Systems. Entropy 2021, 23, 283. [Google Scholar] [CrossRef]
  31. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. Attention Is All You Need. Adv. Neural Inf. Process. Syst. 2017, 30. Available online: https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf (accessed on 14 June 2025).
  32. Liu, L.; He, J.; Ren, K.; Lungu, J.; Hou, Y.; Dong, R. An Information Gain-Based Model and an Attention-Based RNN for Wearable Human Activity Recognition. Entropy 2021, 23, 1635. [Google Scholar] [CrossRef] [PubMed]
  33. Zhang, X.; Chen, F.; Huang, R. A Combination of RNN and CNN for Attention-Based Relation Classification. Procedia Comput. Sci. 2018, 131, 911–917. [Google Scholar] [CrossRef]
  34. Li, Q.; Zhong, R.; Du, X.; Du, Y. TransUNetCD: A Hybrid Transformer Network for Change Detection in Optical Remote-Sensing Images. IEEE Trans. Geosci. Remote Sens. 2022, 60, 5622519. [Google Scholar] [CrossRef]
  35. Reza, S.; Ferreira, M.C.; Machado, J.J.M.; Tavares, J.M.R.S. A Multi-Head Attention-Based Transformer Model for Traffic Flow Forecasting with a Comparative Analysis to Recurrent Neural Networks. Expert. Syst. Appl. 2022, 202, 117275. [Google Scholar] [CrossRef]
  36. Han, Y.; Li, Z.; Wei, T.; Zuo, X.; Liu, M.; Ma, B.; Geng, Z. Production Capacity Prediction Based Response Conditions Optimization of Straw Reforming Using Attention-Enhanced Convolutional LSTM Integrating Data Expansion. Appl. Energy 2024, 365, 123253. [Google Scholar] [CrossRef]
  37. Sheng, Z.; Cao, Y.; Yang, Y.; Feng, Z.-K.; Shi, K.; Huang, T.; Wen, S. Residual Temporal Convolutional Network With Dual Attention Mechanism for Multilead-Time Interpretable Runoff Forecasting. IEEE Trans. Neural Netw. Learn. Syst. 2025, 36, 8757–8771. [Google Scholar] [CrossRef]
  38. Zhao, Z.; Zheng, Y.; Zhang, Z.; Wang, H.; Zhao, Y.; Li, C. Exploring Spatio-Temporal Representations by Integrating Attention-Based Bidirectional-LSTM-RNNs and FCNs for Speech Emotion Recognition. In Proceedings of the 19th Annual Conference of the International Speech Communication Association (INTERSPEECH 2018), Hyderabad, India, 2–6 September 2018; Volume 2018. [Google Scholar]
  39. Zheng, J.; Zheng, L. A Hybrid Bidirectional Recurrent Convolutional Neural Network Attention-Based Model for Text Classification. IEEE Access 2019, 7, 106673–106685. [Google Scholar] [CrossRef]
  40. Himeur, Y.; Rimal, B.; Tiwary, A.; Amira, A. Using Artificial Intelligence and Data Fusion for Environmental Monitoring: A Review and Future Perspectives. Inf. Fusion. 2022, 86–87, 44–75. [Google Scholar] [CrossRef]
  41. Han, H.; Liu, Z.; Li, J.; Zeng, Z. Challenges in Remote Sensing Based Climate and Crop Monitoring: Navigating the Complexities Using AI. J. Cloud Comput. 2024, 13, 34. [Google Scholar] [CrossRef]
  42. Yu, Y.; Si, X.; Hu, C.; Zhang, J. A Review of Recurrent Neural Networks: Lstm Cells and Network Architectures. Neural Comput. 2019, 31, 1235–1270. [Google Scholar] [CrossRef] [PubMed]
  43. Wani, T.; Qadri, S.A.A. A Review of Deep Learning Architectures for Speech and Audio Processing. Int. J. Sci. Eng. Res. 2020, 11, 63–69. [Google Scholar]
  44. Lipton, Z.C.; Kale, D.C.; Elkan, C.; Wetzel, R. Learning to Diagnose with LSTM Recurrent Neural Networks. In Proceedings of the 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, 2–4 May 2016. [Google Scholar]
  45. Kumar, D.; Aziz, S. Performance Evaluation of Recurrent Neural Networks-LSTM and GRU for Automatic Speech Recognition. In Proceedings of the 2023 International Conference on Computer, Electronics and Electrical Engineering and their Applications, IC2E3 2023, Srinagar Garhwal, India, 8–9 June 2023. [Google Scholar]
  46. Abdelgwad, M.M.; Soliman, T.H.A.; Taloba, A.I.; Farghaly, M.F. Arabic Aspect Based Sentiment Analysis Using Bidirectional GRU Based Models. J. King Saud. Univ. Comput. Inf. Sci. 2022, 34, 6652–6662. [Google Scholar] [CrossRef]
  47. Ullah, H.; Munir, A. Human Activity Recognition Using Cascaded Dual Attention CNN and Bi-Directional GRU Framework. J. Imaging 2023, 9, 130. [Google Scholar] [CrossRef]
  48. Hazmoune, S.; Bougamouza, F. Using Transformers for Multimodal Emotion Recognition: Taxonomies and State of the Art Review. Eng. Appl. Artif. Intell. 2024, 133, 108339. [Google Scholar] [CrossRef]
  49. Lara-Benítez, P.; Carranza-García, M.; Luna-Romera, J.M.; Riquelme, J.C. Temporal Convolutional Networks Applied to Energy-Related Time Series Forecasting. Appl. Sci. 2020, 10, 2322. [Google Scholar] [CrossRef]
  50. Pelletier, C.; Webb, G.I.; Petitjean, F. Temporal Convolutional Neural Network for the Classification of Satellite Image Time Series. Remote Sens. 2019, 11, 523. [Google Scholar] [CrossRef]
  51. Ló, T.B.; Corrêa, U.B.; Araújo, R.M.; Johann, J.A. Temporal Convolutional Neural Network for Land Use and Land Cover Classification Using Satellite Images Time Series. Arab. J. Geosci. 2023, 16, 585. [Google Scholar] [CrossRef]
  52. Liu, S.; Le, F.; Chakraborty, S.; Abdelzaher, T. On Exploring Attention-Based Explanation for Transformer Models in Text Classification. In Proceedings of the 2021 IEEE International Conference on Big Data, Big Data 2021, Orlando, FL, USA, 15–18 December 2021. [Google Scholar]
  53. Gao, Y.; Ruan, Y. Interpretable Deep Learning Model for Building Energy Consumption Prediction Based on Attention Mechanism. Energy Build 2021, 252, 111379. [Google Scholar] [CrossRef]
  54. Inglada, J.; Arias, M.; Tardy, B.; Hagolle, O.; Valero, S.; Morin, D.; Dedieu, G.; Sepulcre, G.; Bontemps, S.; Defourny, P.; et al. Assessment of an Operational System for Crop Type Map Production Using High Temporal and Spatial Resolution Satellite Optical Imagery. Remote Sens. 2015, 7, 12356–12379. [Google Scholar] [CrossRef]
  55. Veloso, A.; Mermoz, S.; Bouvet, A.; Le Toan, T.; Planells, M.; Dejoux, J.F.; Ceschia, E. Understanding the Temporal Behavior of Crops Using Sentinel-1 and Sentinel-2-like Data for Agricultural Applications. Remote Sens. Env. 2017, 199, 415–426. [Google Scholar] [CrossRef]
  56. Bousbih, S.; Zribi, M.; Lili-Chabaane, Z.; Baghdadi, N.; El Hajj, M.; Gao, Q.; Mougenot, B. Potential of Sentinel-1 Radar Data for the Assessment of Soil and Cereal Cover Parameters. Sensors 2017, 17, 2617. [Google Scholar] [CrossRef] [PubMed]
  57. Abdikan, S.; Sanli, F.B.; Ustuner, M.; Calò, F. Land Cover Mapping Using Sentinel-1 SAR Data. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2016, XLI-B7, 757–761. [Google Scholar] [CrossRef]
Remotesensing 17 02095 g001
Figure 1. Location of the study area in Hokkaido, Japan. The enlarged map shows Sentinel-1A VH polarization imagery acquired on 18 July 2024 (downloaded from Copernicus Open Access Hub), providing spatial details of the study region. The crop map was generated based on GIS vector data obtained from MAFF’s open data portal.
Figure 1. Location of the study area in Hokkaido, Japan. The enlarged map shows Sentinel-1A VH polarization imagery acquired on 18 July 2024 (downloaded from Copernicus Open Access Hub), providing spatial details of the study region. The crop map was generated based on GIS vector data obtained from MAFF’s open data portal.
Remotesensing 17 02095 g001
Remotesensing 17 02095 g002
Figure 2. Overview of the classification process.
Figure 2. Overview of the classification process.
Remotesensing 17 02095 g002
Remotesensing 17 02095 g003
Figure 3. Time-series variations in VH and VV polarization gamma nought values for each crop type.
Figure 3. Time-series variations in VH and VV polarization gamma nought values for each crop type.
Remotesensing 17 02095 g003
Remotesensing 17 02095 g004aRemotesensing 17 02095 g004b
Figure 4. Cropping map.
Figure 4. Cropping map.
Remotesensing 17 02095 g004aRemotesensing 17 02095 g004b
Remotesensing 17 02095 g005
Figure 5. Temporal changes in the importance of each variable.
Figure 5. Temporal changes in the importance of each variable.
Remotesensing 17 02095 g005
Table 1. Summary of the agricultural fields included in this study.
Table 1. Summary of the agricultural fields included in this study.
Number FieldsArea (ha)
MinimumMedianMeanMaximumStandard Deviation
Beans5330.13 2.58 2.97 12.90 1.89
Beetroot2740.20 2.50 2.89 14.36 1.69
Grass7200.08 2.07 2.67 23.70 2.38
Maize4510.05 2.35 2.99 17.85 2.35
Potatoes1880.33 2.84 3.16 9.04 1.83
Wheat4960.09 2.90 3.25 12.83 1.95
Table 2. Characteristics of the satellite data used in the analysis.
Table 2. Characteristics of the satellite data used in the analysis.
Acquisition DateModePolarizationIncidence Angle (°)Pass DirectionLook Direction
NearFar
25 April 2024IWVH/VV30.65 45.87 DescendingRight
7 May 2024IWVH/VV30.65 45.87 DescendingRight
19 May 2024IWVH/VV30.66 45.87 DescendingRight
31 May 2024IWVH/VV30.66 45.87 DescendingRight
12 June 2024IWVH/VV30.65 45.87 DescendingRight
24 June 2024IWVH/VV30.66 45.87 DescendingRight
6 July 2024IWVH/VV30.66 45.87 DescendingRight
18 July 2024IWVH/VV30.66 45.87 DescendingRight
30 July 2024IWVH/VV30.66 45.87 DescendingRight
11 August 2024IWVH/VV30.66 45.87 DescendingRight
23 August 2024IWVH/VV30.66 45.87 DescendingRight
4 September 2024IWVH/VV30.66 45.87 DescendingRight
16 September 2024IWVH/VV30.65 45.87 DescendingRight
28 September 2024IWVH/VV30.65 45.87 DescendingRight
10 October 2024IWVH/VV30.65 45.87 DescendingRight
22 October 2024IWVH/VV30.65 45.87 DescendingRight
Table 3. Classification accuracies obtained through 10-fold cross-validation. The underlines indicate the highest value for each class and each metric. McNemar test results: a significant difference at the 0.1% level was found between the following: LSTM and TCN, TCN and LSTM + Attention, Bi-GRU + Attention and LSTM + Attention, and LSTM + Attention and TCN + Attention.
Table 3. Classification accuracies obtained through 10-fold cross-validation. The underlines indicate the highest value for each class and each metric. McNemar test results: a significant difference at the 0.1% level was found between the following: LSTM and TCN, TCN and LSTM + Attention, Bi-GRU + Attention and LSTM + Attention, and LSTM + Attention and TCN + Attention.
LSTMBi-GRUTCNLSTM + AttentionBi-GRU + AttentionTCN + Attention
PA
Beans80.1%81.1%76.3%83.6%80.2%79.0%
Beetroot81.4%84.4%84.5%82.1%84.1%82.5%
Grass91.4%92.2%90.1%90.6%92.3%92.3%
Maize81.4%83.5%84.6%80.2%84.0%88.3%
Potatoes66.8%76.8%72.6%66.7%77.2%75.3%
Wheat84.4%85.6%85.0%84.6%85.6%87.7%
UA
Beans82.2%84.8%85.6%80.1%86.5%87.6%
Beetroot87.6%88.7%87.2%86.9%88.7%88.0%
Grass86.9%88.8%87.5%87.4%88.6%90.1%
Maize72.7%76.1%74.1%78.3%74.3%74.9%
Potatoes72.9%75.5%69.1%73.4%73.9%69.7%
Wheat89.7%91.1%87.1%88.7%91.1%91.7%
F1
Beans81.1%82.9%80.6%81.8%83.2%83.1%
Beetroot84.4%86.5%85.8%84.4%86.3%85.2%
Grass89.1%90.4%88.8%89.0%90.4%91.2%
Maize76.8%79.6%79.0%79.2%78.8%81.1%
Potatoes69.7%76.1%70.8%69.9%75.5%72.4%
Wheat87.0%88.3%86.1%86.6%88.3%89.7%
OA83.2%85.3%83.4%83.6%85.2%85.7%
Kappa0.7920.8180.7950.7970.8170.823
Table 4. Results of the McNemar test. ., *, **, and *** indicate significance at the 10%, 5%, 1%, and 0.1% levels, respectively. ‘n.s.’ stands for not significant.
Table 4. Results of the McNemar test. ., *, **, and *** indicate significance at the 10%, 5%, 1%, and 0.1% levels, respectively. ‘n.s.’ stands for not significant.
LSTMBi-GRUTCNLSTM + AttentionBi-GRU + AttentionTCN + Attention
LSTM .***n.s.**
Bi-GRU .****
TCN ***n.s.***
LSTM + Attention ******
Bi-GRU + Attention *
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

Tsuchiya, Y.; Sonobe, R. Crop Classification Using Time-Series Sentinel-1 SAR Data: A Comparison of LSTM, GRU, and TCN with Attention. Remote Sens. 2025, 17, 2095. https://doi.org/10.3390/rs17122095

AMA Style

Tsuchiya Y, Sonobe R. Crop Classification Using Time-Series Sentinel-1 SAR Data: A Comparison of LSTM, GRU, and TCN with Attention. Remote Sensing. 2025; 17(12):2095. https://doi.org/10.3390/rs17122095

Chicago/Turabian Style

Tsuchiya, Yuta, and Rei Sonobe. 2025. "Crop Classification Using Time-Series Sentinel-1 SAR Data: A Comparison of LSTM, GRU, and TCN with Attention" Remote Sensing 17, no. 12: 2095. https://doi.org/10.3390/rs17122095

APA Style

Tsuchiya, Y., & Sonobe, R. (2025). Crop Classification Using Time-Series Sentinel-1 SAR Data: A Comparison of LSTM, GRU, and TCN with Attention. Remote Sensing, 17(12), 2095. https://doi.org/10.3390/rs17122095

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