چكيده لاتين
Monitoring and identifying agricultural crop types is one of the fundamental requirements in resource management, crop planning, water allocation, and macro-level agricultural policymaking. In recent years, remote sensing data, particularly satellite imagery with high spatial resolution and suitable temporal frequency, have provided the possibility of producing accurate and up-to-date maps of cultivated areas. This study aims to evaluate the capability of satellite data for crop type classification and to analyze the performance of transfer learning in both spatial and temporal dimensions, with the goal of generating crop-type maps for the study regions. The first study area, Arak, located in Markazi Province, is characterized by a cold semi-arid climate with relatively limited crop diversity. The second area, Marvdasht in Fars Province, has a temperate mountainous climate and higher crop diversity. These two regions were selected due to their climatic and agricultural differences, allowing a robust evaluation of transfer learning generalization. In the spatial transfer experiment, a combination of optical and radar data, including Sentinel-2 (bands 3, 4, and 5) and Sentinel-1 (VV and VH polarizations), from the 2016–2017 cropping year was used for the two regions. In the temporal transfer experiment, Landsat 8 imagery from the 2014–2015 cropping years was employed to assess the transferability of models trained in one agricultural year to the next within the same region. Spectral indices including NDVI, MNDWI, and BUI were used as input features for the models. All stages of preprocessing, feature extraction, and model implementation were performed in the Google Earth Engine (GEE) cloud platform, which enabled large-scale data processing and model comparison. The Random Forest (RF), Decision Tree (DT), and Support Vector Machine (SVM) algorithms were implemented under four scenarios: supervised and unsupervised learning, with and without Principal Component Analysis (PCA). The results indicated that in spatial transfer learning, transferring the model from Arak to Marvdasht achieved higher accuracy than the reverse direction, with the RF algorithm combined with PCA yielding the best performance. Conversely, in temporal transfer learning, applying PCA—especially in the unsupervised mode—led to a reduction in accuracy due to the loss of certain phenological features of crops. Overall, the RF algorithm demonstrated superior stability and generalization capability compared with DT and SVM in all scenarios. These findings highlight that integrating multi-source satellite data, applying transfer learning, and utilizing the cloud-based processing capabilities of Google Earth Engine can serve as an effective approach for near real-time crop monitoring and reducing dependence on extensive field data across different agricultural regions.
