Monitoring and Estimation of Horticultural and Agricultural Crop Cultivation Area through Integration of Sentinel-2 Sensor Time Series and Machine Learning Algorithms (Case Study: Mojen Region)

Abstract
Background and objectives: Accurate monitoring of cropping patterns and cultivated area, particularly in mountainous regions with fragmented and small-scale landholding structures, is essential for optimal water resource management and agricultural planning. However, the lack of precise and up-to-date statistics poses significant challenges for planning in these areas. Remote sensing technology, with its extensive coverage capacity and ability to provide timely data, offers an effective solution for monitoring cultivated area. This research was conducted with the objective of monitoring and estimating the cultivated area of dominant horticultural and agricultural crops in the Mojen region by integrating Sentinel-2 satellite image time series and machine learning algorithms.
Materials and methods: The study area of this research is the city of Mojen. Located on the southern slopes of the Alborz mountain range with specific topographic conditions, it features a complex cropping pattern consisting of dense orchards (such as apricot, apple, and cherry) and agricultural lands (primarily wheat). The lack of accurate and up-to-date data on the cultivation area of each of these crops has hindered planning for the optimal allocation of the region's limited water resources. Consequently, access to an automated, accurate, and low-cost method based on remote sensing technology for monitoring the cropping pattern in this region is not only a scientific need but also an operational necessity for sustainable water resource and agricultural management. This research, focusing on this regional challenge, sought to fill the existing information gap. In this study, multi-temporal Sentinel-2 satellite images were used to identify the cultivated area of lands in the Mojen water market by utilizing the phenological cycle. Initially, the types of crops in the region, including apricot, apple, cherry, and wheat, along with their phenological stages, were identified. Subsequently, from all orchards and farms in the region, 91 apple orchards, 254 apricot orchards, 15 cherry orchards, 35 wheat fields, and also 50 areas of barren and uncultivated land were selected as training samples to prevent spectral interference. Furthermore, the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Normalized Difference Water Index (NDWI) were used as the most practical vegetation indices for identifying vegetation covers, crop type, and the greening status of the region during different periods. Finally, with the help of field surveys and supervised models including Support Vector Machine (SVM), Neural Network (Neural Net), Minimum Distance (MD), and Maximum Likelihood Estimation (MLE), the satellite images were classified, and the crop types and agricultural land map were obtained.
Results: To verify the accuracy of the results, the generated maps were cross-referenced with field surveys. The analysis of vegetation index profiles clearly demonstrated the distinct separation of the profiles, with each of the apple, apricot, and cherry orchards, as well as the wheat fields, being discernible as separate entities with different intensities and colors. This effectively illustrated the phenological dynamics of the region's dominant crops—including Leaf Area Index (LAI), canopy type and structure, moisture content, and the identification of vegetated water bodies—throughout their growth cycle from onset to harvest and dormancy. A cross-validation method was employed to assess the accuracy of the classification models, utilizing the statistical metrics of Kappa coefficient and overall accuracy. Additionally, the values of the two errors—omission error (indicating pixels of a class that were incorrectly excluded from that class) and commission error (indicating pixels incorrectly included in a class)—were calculated to determine producer's accuracy and user's accuracy, respectively. For the Neural Network, Support Vector Machine (SVM), Maximum Likelihood, and Minimum Distance methods, the Kappa coefficients were 0.79, 0.78, 0.66, and 0.60, respectively, with corresponding overall accuracies of 89%, 88%, 81%, and 79%. Examination of these values, along with the commission and omission errors, indicated the superior performance of the Neural Network and Support Vector Machine methods compared to the other two. The results obtained using the Neural Network method showed that from the total 2000 hectares of land in Mojen, the areas covered by apricot, apple, and cherry orchards were 1442, 333, and 52 hectares (72.1, 16.7 & 2.6 percent), respectively, and the area under wheat cultivation was 173 hectares (8.6 percent). This represents a 94% accuracy compared to the statistics from the Agricultural Jihad Organization. Field investigations revealed that factors such as the limited intercropping of peach, nectarine, sour cherry, and walnut trees within some apricot and apple orchards—coupled with the region's vastness and similarities in agricultural calendar phases—contributed to estimation errors for these crops. Furthermore, the presence of unremoved weeds in a small number of orchards in early spring led to errors in distinguishing these orchards from wheat farmlands. The lowest error rate was observed in the estimated area under wheat cultivation, attributable to its autumn sowing season, during which no other crops are simultaneously cultivated in the region.
Conclusion: In summary, the integration of multi-temporal satellite imagery and phenological indices offers a reliable approach for identifying agricultural crop types and estimating cultivated areas with acceptable accuracy. From a management and monitoring perspective, this methodology is highly valuable due to its distinct advantages, including access to remote regions, extensive spatial coverage, rapid and easy data acquisition, and sufficient precision for generating accurate agricultural land-use maps. Furthermore, employing this technology across various research and operational contexts can provide critical insights for improving irrigation management, optimizing water use in farms and orchards, and mitigating water stress. By adopting this approach, stakeholders can enhance the accuracy of agricultural planning and monitoring across different sectors, while minimizing costs and enabling continuous, large-scale oversight.