Journal of Geosciences Insights

ABSTRACT

Agricultural monitoring is pivotal for optimizing crop management, resource allocation, and ensuring sustainable development. This study explores the application of advanced deep learning models for automated agricultural monitoring in Rwanda, focusing on two key tasks: crop type classification and farmland segmentation. Utilizing TensorFlow's MobileNet_V2 model, we achieved an overall classification accuracy of 76.13% in identifying various crop types from satellite imagery. Additionally, the Segment Anything Model (SAM) demonstrated promising results in farmland segmentation, effectively delineating agricultural fields within high-resolution satellite images. Despite challenges in quantitative evaluation due to the absence of ground truth data, the visual outcomes underscore SAM's potential for unsupervised segmentation tasks. The integration of these models offers a comprehensive approach to agricultural monitoring, facilitating informed decision-making for farmers, policymakers, and researchers. Future research directions include model optimization, enhanced data augmentation techniques, and the integration of multi-source data to further improve classification and segmentation performance.

KEYWORDS

1. Agricultural monitoring
2. Deep learning
3. Crop type classification
4. Farmland segmentation
5. MobileNet_V2
6. Segment anything model (SAM)
7. Remote sensing
8. Tensor flow
9. Machine learning, Satellite imagery