نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه علوم و مهندسی آب، دانشکده کشاورزی و منابع طبیعی، دانشگاه بین المللی امام خمینی (ره)، قزوین، ایران
2 بخش تحقیقات فنی و مهندسی کشاورزی، مرکز تحقیقات و آموزش کشاورزی و منابع طبیعی خوزستان، سازمان تحقیقات، آموزش و ترویج کشاورزی، اهواز،
چکیده
کلیدواژهها
عنوان مقاله [English]
نویسندگان [English]
Accurate mapping of crop types and their spatial distribution, along with estimating cultivated areas, plays a critical role in water resource planning and agricultural land management in arid and semi‐arid regions. In this study, field surveys were first conducted to collect precise geographic coordinates of selected crop fields, including wheat, maize, sesame, and alfalfa, and these reference points were subsequently integrated into the Google Earth Engine environment for alignment with remote sensing data. Sentinel-2 imagery was processed through atmospheric correction and cloud masking, while Sentinel-1 data underwent geometric correction and speckle filtering. A set of spectral indices derived from Sentinel-2, together with backscatter intensity and polarization ratio features extracted from Sentinel-1, was generated and integrated to enhance the separability of crop types with similar phenological characteristics. The collected reference samples were divided into training (70%) and validation (30%) subsets, and three classification algorithms, Random Forest, Support Vector Machine, and Extreme Gradient Boosting, were applied to produce the final crop type map. Model performance was evaluated using the confusion matrix, overall accuracy, and the Kappa coefficient. The results indicated that the Random Forest algorithm achieved the highest performance, with an overall accuracy of 96% and a Kappa coefficient of 0.93, demonstrating superior discrimination of crop types, particularly wheat. The Extreme Gradient Boosting model ranked second with an accuracy of 89%, while the Support Vector Machine exhibited lower performance. Comparison of the estimated cultivated areas with official agricultural statistics further confirmed the high reliability of the Random Forest results.
کلیدواژهها [English]
Accurate mapping of crop types and the estimation of their cultivated areas are essential components of sustainable agricultural management, particularly in arid and semi-arid regions such as southwestern Iran. Rapid population growth, diminishing water resources, and the accelerating impacts of climate change have intensified the need for timely and reliable information on cropping patterns. Remote sensing provides a practical and cost-efficient means for agricultural monitoring, offering repeated temporal coverage and high spatial detail through multisource satellite observations. The combined use of optical and radar imagery has proven especially valuable, as it integrates both spectral and structural information to improve discrimination among crops with similar spectral characteristics.
Although considerable progress has been made in crop mapping using satellite imagery, there remains a lack of consistent and transferable frameworks that combine optical and radar data in an operational manner at basin scale in Iran. The Maroon–Jarrahi Basin, one of the country’s key agricultural regions, faces severe challenges from groundwater depletion and increasing climatic variability. This situation requires an accurate, spatially explicit understanding of crop distribution, particularly for major crops such as wheat, maize, sesame, and alfalfa. The present study seeks to establish a comprehensive classification framework that integrates Sentinel-1 and Sentinel-2 datasets to generate accurate crop type maps and reliable acreage estimates for this critical agricultural region.
The study employed a supervised classification design using multisource satellite data and extensive field observations collected between 2021 and 2024 within the Maroon–Jarrahi sub-basin of Khuzestan Province, southwestern Iran. Field data included precise GPS-located samples of four major crops (wheat, maize, sesame, and alfalfa) collected across several administrative districts.
Sentinel-2 imagery was atmospherically corrected and cloud-masked using the Fmask algorithm. A suite of vegetation indices sensitive to plant biophysical and phenological traits was derived and resampled to a 10-m spatial resolution. For Sentinel-1, radar backscatter data acquired in Interferometric Wide (IW) mode with VV and VH polarizations were processed. Speckle noise was minimized using a 5×5 adaptive Lee filter, preserving textural integrity. The polarization ratio (VV/VH) was further computed to capture variations in canopy structure and surface moisture. All radar and optical layers were fused into a single composite dataset to maximize spectral and structural feature diversity. Moreover, All samples for each class were compiled, with 70% allocated for model training and 30% reserved for independent validation
Classification experiments were conducted in the Google Earth Engine environment coupled with Python-based processing. Three classification algorithms, Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) were calibrated and tested. For RF, model parameters included 100 decision trees, five variables per split, and a minimum leaf population of three. The SVM classifier used a radial basis kernel, with optimized γ values ranging from 0.0002 to 0.0004 and C values between 1 and 100. The XGBoost model underwent grid-based hyperparameter tuning with five-fold cross-validation to identify optimal values of n_estimators, max_depth, learning_rate, and subsample rate.
Model evaluation was based on a confusion matrix and a set of standard accuracy metrics, including overall accuracy (OA), Kappa coefficient, precision, recall, and F1-score. A spatially stratified random sampling approach was adopted to minimize spatial autocorrelation between training and validation samples. The final classified maps were exported as GeoTIFF files, and crop areas were computed at both county and basin levels.
The comparative assessment of classifiers revealed clear differences in performance. The Random Forest model produced the highest accuracy (OA = 96%, Kappa = 0.93, F1 = 0.97), demonstrating a strong ability to distinguish spectrally similar crops, especially between wheat and alfalfa, while maintaining resilience to noise and spatial heterogeneity. The RF approach effectively reduced confusion between vegetated and non-vegetated classes, with minimal misclassification of water bodies or algae.
The XGBoost classifier ranked second (OA = 89%, Kappa = 0.83, F1 ≈ 0.90), achieving stable results across most crop classes but showing some confusion between bare soil and sparse vegetation, particularly under post-harvest conditions when spectral signatures converge. The Support Vector Machine yielded the lowest performance (OA ≈ 54%, Kappa = 0.31, F1 ≈ 0.55), largely due to its sensitivity to inter-class spectral overlap, notably between sesame and maize.
Area estimation results confirmed strong agreement between RF-derived crop acreage and official agricultural statistics, with deviations of less than 5% for wheat. For alfalfa, a minor overestimation was observed, attributed to its persistent greenness and mixed-pixel effects. The combined use of Sentinel-1 and Sentinel-2 data enhanced the accuracy of water-related classifications, accurately delineating open water and aquatic vegetation. Visual assessment further showed that the RF-generated maps exhibited the most coherent and spatially consistent patterns, whereas XGBoost maps displayed moderate fragmentation and SVM outputs contained irregular artifacts. Overall, the integration of radar and optical data markedly improved class separability, reinforcing the complementary value of multisensor fusion for crop mapping in arid environments.
This research establishes a coherent and replicable framework for crop classification and acreage estimation based on the integration of optical and radar satellite data. The results highlight the potential of Sentinel-1 and Sentinel-2 fusion to significantly improve the accuracy and reliability of crop mapping in data-scarce arid and semi-arid regions. Among the tested classification approaches, the Random Forest algorithm provided the most consistent and robust outcomes, producing highly accurate and spatially stable crop maps.
The proposed methodology provides a transferable and scalable foundation for regional agricultural monitoring, enabling timely assessment of cropping patterns and supporting more informed water resource management and agricultural policy decisions. Future efforts should incorporate multi-year climatic and phenological datasets to strengthen temporal generalization and move toward dynamic, near-real-time monitoring systems across diverse agro-ecological zones of Iran.
Masoud Soltani: Guiding, Conceptualization, Editing the paper, Controlling the results
Bahareh Bahmanabadi: Programming, Conceptualization, methodology, Writing - Initial draft preparation, performing software/statistical analysis
Ali Mokhtaran: supplying the field data in the Marun Basin, Guiding, Conceptualization
Data available on request from the authors.
The authors would like to express their sincere gratitude to Imam Khomeini International University for providing financial and logistical support that greatly contributed to the success of this research. We also acknowledge the valuable assistance of the Agricultural Engineering Research Institute (AERI) for supplying the field data in the Marun Basin.
The authors avoided data fabrication, falsification, plagiarism, and misconduct. Conflict of interest: The author declares no conflict of interest.
The authors of this article declared no conflict of interest regarding the authorship or publication of this article.
)