In the lush, green landscapes of Bhutan, where terraced rice fields cascade down the foothills, a technological revolution is quietly taking place. The tiny Himalayan kingdom, known for its commitment to environmental conservation and Gross National Happiness, is leveraging deep learning and high-resolution satellite imagery to map rice cultivation areas with unprecedented accuracy. This pioneering work, led by Biplov Bhandari, a researcher affiliated with Woolpert Digital Innovation, the Earth System Science Center at The University of Alabama in Huntsville, and the SERVIR Science Coordination Office at NASA Marshall Space Flight Center, is set to transform how Bhutan, and potentially other nations, approach agricultural planning and food security.
Bhandari’s study, published in the ISPRS Open Journal of Photogrammetry and Remote Sensing, focuses on Paro, one of Bhutan’s top rice-yielding districts. The research employs publicly available high-resolution satellite imagery from Planet Labs, funded by Norway’s International Climate and Forest Initiative (NICFI). The goal? To provide geospatial products that can support decision-making in the context of economic and population growth, and their impacts on food security. “As Bhutan’s economy grows, so does the demand for accurate, reliable data to inform our agricultural policies,” Bhandari explains. “Our work aims to bridge that gap by harnessing the power of deep learning and remote sensing.”
The study compares two deep learning approaches: point-based (DNN) and patch-based (U-Net) models. Four different models per approach were trained, each incorporating various data sets, including Red, Green, Blue, and Near-Infrared (RGBN) channels, Elevation data, and Sentinel-1 data. The results were clear: the U-Net model consistently outperformed the DNN model, with the best-performing U-Net model achieving an F1-score of 0.8563 during training and 0.6582 during independent validation.
But the implications of this research extend far beyond Bhutan’s borders. As the global population continues to grow, so too does the demand for food. Accurate mapping of crop cultivation areas can help governments and organizations plan for increased food production, optimize resource allocation, and mitigate the impacts of climate change. “This technology has the potential to revolutionize the way we approach food security,” Bhandari says. “By providing accurate, up-to-date information on crop cultivation, we can help policymakers make informed decisions that benefit both people and the planet.”
Moreover, the study demonstrates the potential for combining deep learning approaches with traditional survey-based methods currently utilized by the Department of Agriculture (DoA) in Bhutan. This fusion of old and new could pave the way for more accurate and efficient agricultural planning. The research also highlights the effectiveness of using regional land cover products, such as SERVIR’s Regional Land Cover Monitoring System (RLCMS), as a weak label approach to address class imbalance and improve sampling design for deep learning applications.
The findings also underscore the importance of independent model evaluation, as performance metrics varied significantly across different evaluations. This highlights the need for rigorous testing and validation to ensure the reliability of deep learning models in real-world applications.
As Bhutan continues to integrate technological approaches into its agricultural planning, the lessons learned from this research could shape future developments in the field. By providing a robust framework for mapping rice cultivation areas, Bhandari’s work lays the groundwork for similar initiatives in other regions, potentially transforming global food security efforts.