The study was conducted in Dacope Upazila, located in the Khulna district of southwestern Bangladesh, between latitudes 22.57˚N and longitudes 89.51˚E (Figure 1). The total area of the upazila is approximately 991.58 square kilometers, with an elevation ranging from 2 to 3 meters above mean sea level [18]. This region lies adjacent to the southern boundary of the Sundarbans mangrove forest, covering an area of 11790.13 hectares, which significantly influences its ecological and agricultural landscape [19][20]. The climate in Dacope is humid subtropical, with hot summers and cold winters [21]. Average maximum temperatures range between 27˚C and 34˚C, while minimum temperatures vary from 16˚C to 27˚C. Over 75% of the annual rainfall occurs during the monsoon season (June to September), creating favorable conditions for rainfed T. Aman rice cultivation [22].

Figure 1. Location of the study area marked in orange (left) and red (right) color.

During the wet season (July-December), T. Aman rice occupied nearly 95% of this area. Farmers typically grow traditional, long-duration rice varieties with tall plant stature that are tolerant to stagnant water. These varieties are usually transplanted in August or September and harvested in December or early January. Due to their late maturity, there is a limited opportunity for a subsequent dry-season crop. During the dry season (January-May), a substantial portion of the agricultural land remains fallow, primarily due to elevated soil salinity, inadequate access to quality irrigation water, and delayed drainage conditions [23].

The region’s soil is predominantly silty-clay in the top 0 - 30 cm layer, transitioning to clay at 30 - 60 cm depth [24]. Soil salinity levels fluctuate significantly between seasons, ranging from 1 - 3 dS·m⁻¹ in the wet season to 6 - 12 dS·m⁻¹ in the dry season [23]. Approximately 70% - 75% of the arable land falls within medium-high to low elevation categories, while the remaining 25% - 30% lies within medium-high to high elevation zones [25]. This heterogeneity in elevation and salinity gradients in Dacope makes it a critical zone for studying rice cultivation under the challenging conditions of coastal regions.

Ground truth (GT) data collection took place between July and November in 2019, including GPS coordinates, varietal information for rice crops, and key phenological dates such as transplanting and harvesting. A total of six T. Aman rice varieties (BRRI dhan49, BRRI dhan75, BRRI dhan76, BR23, and local aromatic rice Sadamota) were recorded. Each variety of field growth stage was documented for temporal profile analysis. At least 60 samples were collected for each class (rice varieties, mangrove, homestead, waterbodies, and urban) to ensure statistical robustness.

To monitor rice growth dynamics, five Sentinel-1 SAR SLC (Single Look Complex) datasets were acquired at approximately 15-day intervals during the 2019 wet season (12 July to 22 September) from the Alaska Satellite Facility (ASF). Although the Sentinel-1 constellation typically provides a 6- or 12-day revisit frequency, images were selected at roughly 15-day intervals due to limited data availability and quality constraints over the study area. This interval ensured consistent temporal coverage that corresponded closely with key phenological stages of T. Aman rice while avoiding missing or low-quality scenes, thereby maintaining a balanced and representative time series for growth monitoring. Data preprocessing was conducted using the Sentinel Application Platform (SNAP), a widely used software suite offering specialized tools for SAR image analysis [26]. The key steps in the preprocessing pipeline are illustrated in Figure 2.

Preprocessing began with the conversion of SLC to Ground Range Detected (GRD) products using SNAP’s “S1 SLC to GRD” operator. This step enabled the derivation of σ˚ backscatter coefficients for both VH and VV polarizations, which are crucial for analyzing vegetation and surface water dynamics [27]. Due to the coherent nature of SAR data, speckle noise is an inherent issue. To address this, multi-looking was applied, followed by a Refined Lee Filter with a 5 × 5 kernel [28], balancing noise reduction with edge preservation. This adaptive filtering technique is particularly effective in preserving spatial heterogeneity in rice fields, where planting density and water levels may vary [29].

Next, radiometric calibration was performed to convert digital numbers into physically meaningful backscatter values in decibels (dB), enabling comparisons across multiple acquisition dates. To reduce terrain-induced distortions, geometric corrections were applied using the Range Doppler Terrain Correction algorithm. Each scene was orthorectified using a high-resolution Digital Elevation Model (DEM) to ensure accurate ground geometry. All scenes were then precisely co-registered using SNAP’s image co-registration tool, which is essential for pixel-level alignment in time-series analyses. After radiometric and geometric corrections, VH and VV datasets were temporally stacked into a multi-date composite. To enhance signal consistency across dates and further suppress residual speckle noise, the Empirical Mode Decomposition (EMD) technique was employed [30]. This method was used to decompose complex radar backscatter time-series into intrinsic mode functions, isolating vegetation signals from short-term noise. In rice monitoring, it enhances temporal backscatter clarity by emphasizing key phenological changes such as tillering, heading, and ripening, improving the precision of variety discrimination. To minimize the effects of varying imaging conditions (e.g., changes in incidence angle, rainfall, or soil moisture), normalization of backscatter values was carried out. This process ensured consistency in the temporal profile across all acquisition dates.

These comprehensive preprocessing steps, including speckle filtering, radiometric and geometric corrections, normalization, and temporal smoothing, resulted in a clean, temporally consistent dataset. The final product was a multi-date stack of vertical transmit horizontal receive (VH) and vertical transmit vertical receive (VV) backscatter coefficients, optimized for time-series analysis. This dataset formed the basis for detecting rice phenology, tracking growth dynamics, and ultimately classifying rice varieties and other land cover types in the coastal area of Khulna.

Figure 2. Flowchart of preprocessing SAR data by SNAP.

To analyze the temporal backscatter behavior of rice and other land cover types, ground truth (GT) points were collected through visual interpretation of high-resolution imagery and field knowledge. These GT points represented various land cover classes, including rice fields, fallow land, water bodies, homesteads, mangroves, and urban areas. The points were digitized in QGIS and exported as vector shapefiles for use in subsequent analysis. Temporal backscatter profiles were generated using Sentinel-1 SAR data, focusing on both VH and VV polarization modes. In SNAP, two sets of RGB temporal composites were developed for six acquisition dates during the 2019 wet season. The first composite covered 20 July, 5 August, and 20 August, while the second included 20 August, 5 September, and 25 September. These composites allowed visual inspection of backscatter variation across rice growth stages and supported temporal feature extraction. The vector shapefiles were imported into SNAP and overlaid on the RGB composites using the Vector Toolbox.

Subsequently, the Analysis Toolbox was used to extract statistical parameters, including mean and standard deviation of the backscatter values (in decibels), for each class and acquisition date. The extracted temporal profiles enabled the detection of distinct phenological stages in rice cultivation, such as transplanting (characterized by low VV backscatter due to water presence), vegetative growth (marked by a progressive increase in VH backscatter due to biomass accumulation), and reproductive/maturity stages (with stabilized or slightly reduced backscatter [31]. These signatures also revealed clear distinctions between rice and non-rice land covers, supporting the development of class-wise backscatter thresholds. This analysis provided a robust foundation for subsequent classification of land cover types and rice varieties using decision-tree and machine learning approaches.

To support accurate land cover classification and rice variety discrimination, SAR backscatter signatures were systematically extracted and analyzed for each major land cover type present in the study area. These classes included rice fields, fallow land, water bodies, homesteads, mangrove forests, and urban settlements. Using the temporal VH and VV composites produced from Sentinel-1 data, pixel-wise backscatter values (σ˚) were extracted for each class based on ground truth (GT) locations. The analysis was conducted across all six acquisition dates to capture the full seasonal variability in backscatter responses. For each land cover class, the lower and upper bounds of the σ˚ values were determined for every acquisition date, enabling the identification of characteristic backscatter ranges.

In the case of rice fields, phenological interpretation of these values was also performed. The lowest VV backscatter values typically occurred shortly after transplanting, when water was dominant in the field, while the highest VH values were recorded during the peak vegetative stage due to increased canopy density and volume scattering [32]. These observations allowed the construction of a temporal envelope for rice, capturing its distinct growth stages in terms of radar backscatter. For non-rice classes such as water bodies and urban areas, relatively stable backscatter values across dates were observed, while mangroves and homesteads exhibited moderate temporal variability. This step of signature extraction and analysis provided essential quantitative thresholds, which were later used in the knowledge-based classification framework to differentiate between rice varieties and other land cover types based on SAR backscatter behavior over time [33].

Based on the extracted SAR backscatter thresholds for each land cover type and rice variety, a knowledge-based decision tree classifier was developed and implemented in ENVI software. This approach was selected because it allowed direct integration of expert-derived thresholds and empirical backscatter ranges obtained from field observations, making it particularly suitable for studies with limited training data and the need for transparent, interpretable classification rules. The classification rules were derived by systematically analyzing the upper and lower σ˚ values from both VH and VV polarizations across multiple acquisition dates [34]. These thresholds captured the temporal variability specific to rice growth stages and other land features, enabling accurate segmentation of the multi-temporal SAR composite into distinct classes. The decision tree was executed to generate a classified map representing rice varieties and various land cover types. To ensure classification reliability, the results were validated using an independent set of ground truth points. Validation metrics, including overall accuracy, kappa coefficient, and confusion matrix, were computed to evaluate performance and identify potential misclassifications [35]. Although the knowledge-based approach provided a solid foundation, integrating supervised machine learning classifiers like Random Forest and Support Vector Machines (SVM) is recommended for future improvements. These approaches would leverage labeled training datasets and apply a training-validation split to capture more complex relationships in the data.

Furthermore, a detailed analysis of VH and VV backscatter trends was conducted to observe structural and moisture-related changes in rice fields over time. The VV polarization exhibited high backscatter values during the transplanting stage due to standing water, followed by a gradual increase as canopy development progressed. In contrast, VH polarization showed increased sensitivity to biomass accumulation and canopy structure, particularly during the vegetative and reproductive stages [36]. These temporal backscatter profiles enabled the identification of key phenological transitions, such as tillering, heading, and ripening, thus facilitating rice varietal discrimination and supporting comprehensive crop growth monitoring throughout the season.

Following the classification process, post-processing and visualization were conducted to enhance interpretability and presentation of the results. Classified maps were visualized in QGIS using false-color composites, which allowed clear differentiation between rice varieties and other land cover classes based on their assigned color codes. To improve spatial context and readability, cartographic enhancements were applied by overlaying administrative boundaries, hydrological networks, and other relevant thematic layers [37]. These overlays provided valuable geographic reference points, facilitating better interpretation of spatial patterns and land use distributions. The final map outputs were systematically refined to ensure clarity, consistency, and visual balance. Map elements such as legends, scale bars, and north arrows were included, and symbology was standardized across all map products. The spatial results derived from the SAR-based classification, thereby supporting informed decision-making in agricultural monitoring and sustainable land use planning.

Standard evaluation metrics (including overall accuracy, precision, recall, and F1 score) were used to measure the SAR-based rice varietal classification performance. These metrics are derived from the confusion matrix, which classifies predictions into true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN). Overall accuracy measures the proportion of correctly classified samples out of the total, calculated as the sum of TP and TN divided by the total number of samples, expressed as:

Overall Accuracy = (TP + TN)/(TP + FP + FN + TN) (1)

Precision, defined as TP/(TP + FP), measures the reliability of positive predictions by indicating the proportion of true positives among all predicted positives, thereby reflecting the rate of false positives. Recall or sensitivity, given by the formula TP/(TP + FN), assesses the effectiveness of the model in capturing true positive instances, thereby highlighting the rate of false negatives. As the harmonic mean of precision and recall, the F1 score is calculated by:

F1 Score = (2 * Precision * Recall)/(Precision + Recall) (2)

F1 score provides a balanced metric that accounts for both false positives and false negatives in evaluating classifier performance [38]. These metrics collectively offer a comprehensive understanding of the model’s classification performance, particularly in distinguishing rice varieties and land cover types using multi-temporal SAR data.

Data from three acquisition dates were co-registered to form a unified multi-temporal dataset, reflecting spatial and temporal variations in rice planting across the study region. To generate a multi-date SAR False Color Composite (FCC) image, three acquisition dates were assigned to the primary color channels: the first date to red (R), the second to green (G), and the third to blue (B). This assignment allowed the creation of a georeferenced multi-temporal FCC dataset, visually representing temporal variations in the study area. Various types of rice varieties could easily be distinguished by the theory of color composite in RGB (Red Green Blue) format. The resulting composite images are presented in Figure 3 and Figure 4.

Figure 3. 1st composite (20 July, 5 August, and 20 August) of RGB image for Sentinel-1 data with ground truth (GT) points.

Figure 4. 2nd composite (20 August, 5 September, and 25 September) of RGB image for Sentinel-1 data with ground truth (GT) points.

The SAR backscatter coefficient from various land-cover types was analyzed over time, as depicted in Figure 5. Based on analysis, the σ˚ values for all classes exhibited similar variations with changes in VV polarization. The most significant increasing trend was observed in rice crop areas over time. Homestead areas in the SAR data displayed a consistently high backscatter of approximately −7 dB in VV polarization throughout the study period, appearing consistently bright across all dates with minimal variation in mean backscatter [15]. As illustrated in Figure 5, the river, characterized by its smooth surface acting as a specular reflector, exhibited very low backscatter ranging from around −21 to −23 dB [7]. Additionally, a consistent backscatter trend of −15 dB was noted for the mangrove across different dates.

Figure 5. Temporal profile of different classes in VV polarization.

A time-series approach was applied to Sentinel-1 datasets for extracting rice areas, where the backscatter values for rice crops ranged from −25 to −15 dB (Figure 6). The backscatter values (dB) were low during the transplanting time due to flooded field conditions. Rice is usually transplanted in the field under standing water condition. At that time, rice plant height was very short, and very low backscattered values (dB) were found due to surface reflection [39]. The dB values increased

Figure 6. Temporal profile of different rice varieties in VH polarization.

progressively as the rice plants grew, reaching their peak around the mid-growth or early maturity stage [13].

A decision tree classification map for Synthetic Aperture Radar (SAR) data provides a visual representation of land-cover classification results obtained using a decision tree algorithm. This method segments SAR imagery into distinct categories based on decision rules derived from statistical patterns in backscatter values. The decision tree approach is widely utilized in remote sensing because of its capacity to handle non-linear relationships and accurately classify diverse land features [40].

Figure 7. Decision tree classification of SAR image for different T. Aman varieties and land types.

The classification analysis successfully delineated four major land-cover types within the study area: river, waterlogged areas, homesteads, and mangroves. Moreover, distinct backscatter patterns from multi-temporal SAR data enabled the identification of various T. Aman rice varieties, including BRRI dhan75, BRRI dhan76, BRRI dhan49, BR23, and the aromatic variety Sadamota (Figure 7). The classification utilized VH and VV polarization data from Sentinel-1, allowing precise differentiation among land covers and crop types. Multi-temporal backscatter analysis was crucial in capturing growth dynamics and surface changes over time, which improved classification accuracy [32]. The study highlights the potential of SAR-based decision tree classification for agricultural monitoring and land-use assessment. This method enhances the accuracy of rice mapping, contributing to better resource management and decision-making in rice production.

Figure 8 presents the backscatter temporal profile derived from multi-temporal SAR data, enabling clear differentiation between rice categories and other land features. The mangrove class was easily identified using the VH polarization temporal profile, while other features such as homesteads, waterlogged areas, and fallow land were effectively distinguished using multi-temporal VV polarization. The classification of rice varieties was primarily based on VH backscatter polarization profiles, which helped in accurately distinguishing different rice types.

Figure 8. T. Aman rice variety classification map for SAR image.

Accuracy assessment was carried out using ENVI software. From the classification, the various features vs. various rice varieties showed high and significant accuracy (BRRI dhan49 gave 81%, aromatic rice gave 83%, and BRRI dhan75 gave 82%). The training sites, obtained from ground truth data collection, served as reference points. The results indicated that the higher level of mapping accuracy (more than 90%) was shown for the mangrove, homestead river, and waterlogged area. Whereas the accuracy of all rice types was more than 80%. The overall accuracy and kappa value were obtained 76.3% and 0.734, respectively. The accuracy of the rice varietal classification map was assessed using validation methods including error matrices, kappa statistics, and accuracy indices. This helps assess the reliability of the classification results and identify areas of uncertainty [41].

To advance the operational utility of SAR-based agricultural monitoring, several strategic recommendations emerge from this study. Initially, expanding SAR applications to other crop types and across different seasons will enable comprehensive, year-round surveillance of agricultural ecosystems. Integrating SAR with optical and multispectral datasets can enhance classification accuracy and provide deeper insights into crop health and land-use dynamics. The development of automated classification systems using machine learning models such as Random Forests and Convolutional Neural Networks (CNNs) is essential for scaling analysis and reducing manual intervention. Policymakers should support the adoption of remote sensing technologies by investing in data infrastructure and training programs for local stakeholders. Additionally, real-time monitoring tools and forecasting dashboards based on SAR data can facilitate early detection of crop stress, flooding, and pest outbreaks, enabling timely interventions. Finally, continuous ground truth validation and engagement with farming communities are vital to ensure the relevance and accuracy of remote sensing outputs, fostering trust and practical application in agricultural decision-making.