| Popis: |
In the 21st century, the establishment of efficient water resource management is crucial for ensuring world water and food security. Irrigation is a significant artificial process in the hydrological cycle and presents the only way to balance between mentioned issues, where collecting knowledge is essential for developing adaptive and sustainable strategies. Considering that, the precise information about the spatio-temporal distribution of irrigated fields on a national scale is thus the initial key step for agricultural water resource management. With a high spatial, spectral, and temporal resolution, Sentinel-2 provides new possibilities in this field. This research focuses on using multispectral satellite imagery and advanced machine learning models for detecting irrigation and rainfed fields on a plot scale. Dry year images during irrigation season were used for vegetation indices calculation for three crop types: maize, soybean, and sugar beet. These three databases were used separately for training the Random Forest classifier. The results showed high overall accuracy for each three crops where soybean reached the highest 0.91, maize 0.89, while sugar beet reached 0.76. According to the results, the assumption is that the difference in accuracy between crops could be caused by the difference in the geospatial characteristic of the area, amount of data, omission in labeling crop types and rainfed fields. Irrigated agricultural fields present a challenge for classification and mapping considering the heterogeneity of the area, climate impact, and diverse crop types. This study showed that classification could be done using Sentinel-2 images, but further analysis including climate and soil data could improve the classification. This methodology has the potential to produce an annual irrigation map which is very important information for optimizing water use and making sustainable agricultural policy. |
