|
Determining the area under rice cultivation using machine learning methods and simultaneous use of optical satellite images and radar with virtual aperture (Case study: Babol city)
|
Ali لخخیشقظ   |
|
|
|
Abstract: (10 Views) |
Rice is one of the key agricultural products in Iran and the world, and accurate monitoring of its cultivated area, especially the separation of monoculture and double-cropping, plays a vital role in water resources management and production planning. Mazandaran province, as the country's rice production hub, requires new methods for monitoring agricultural lands due to the spatial dispersion of farms and diverse environmental conditions. In this study, a new framework based on the combination of deep convolutional networks (CNN) and recurrent networks (RNN) was developed, which identifies rice lands and distinguishes monoculture/double-cropping patterns at two levels. In the first step, the U-Net neural network architecture was used to identify rice lands in Babol County. This model was trained in two scenarios: (1) with Sentinel-2 optical data only and (2) with a combination of optical and radar data (Sentinel-2 and Sentinel-1). The results of the experiments showed that the model based on Sentinel-2 optical data was evaluated with an overall accuracy of 96.35% and an F1 score of 0.957. While it was expected that the combination of optical-radar data would improve the results, in the study area, due to the relatively cloudless conditions, the main advantage of SAR data was not very apparent. This, along with the need for more complex preprocessing of these data, could be the reason for the relative decrease in the accuracy of the combined model with an overall accuracy of 94.32 and an F1 score of 0.93. Three regions were selected to test the generalizability of the model: Sari County as an area close to the main area, Lahijan County as an area with a greater distance and more diverse vegetation, and south Tehran as an area without rice cultivation. The generalizability assessment showed that the model provided stable and accurate results in Sari County, while in Lahijan, the accuracy decreased but was still acceptable with an overall accuracy of over 88%. Testing the model in rice-free lands south of Tehran also showed a very low error rate (14.2%), indicating the high ability of the model to distinguish non-rice fields. In the second step, the GRU recurrent network was used to distinguish monoculture and biculture lands. This model was able to provide a very good performance in distinguishing monoculture and biculture rice fields with an overall accuracy of 98.4%, a kappa coefficient of 0.951, and an F1 score of 0.974. |
|
| Keywords: Rice Crop Area, Deep learning, CNN, RNN, Sentinel1, Sentinel2 |
|
|
|
Type of Study: Research |
Subject:
Photo&RS
Received: 2025/10/14 | Accepted: 2026/07/17
|
|
|
|
|
|
|
| Add your comments about this article |
|
|
|
