Skip to main content
ARS Home » Pacific West Area » Maricopa, Arizona » U.S. Arid Land Agricultural Research Center » Plant Physiology and Genetics Research » Research » Publications at this Location » Publication #418048
Research Project: Analysis and Quantification of G x E x M Interactions for Sustainable Crop Production

Location: Plant Physiology and Genetics Research

Title: Crop type mapping in smallholder agricultural settings using Sentinel-1 SAR imagery and deep learning

Author
item POLAVARAPU, SANDEEP - University Of Maryland
item IRIGIREDDY, IRIGIREDDY - Oak Ridge Institute For Science And Education (ORISE)
item KULKARNI, CHAITANYA - University Of Maryland
item LAN, SONG - University Of Maryland
item PVNR, KOUTILYA - University Of Maryland
item Bandaru, Varaprasad

Submitted to: Book Chapter
Publication Type: Book / Chapter
Publication Acceptance Date: 8/15/2024
Publication Date: 7/21/2025
Citation: Polavarapu, S., Irigireddy, I.C., Kulkarni, C.S., Lan, S., Pvnr, K., Bandaru, V. 2025. Crop type mapping in smallholder agricultural settings using Sentinel-1 SAR imagery and deep learning. Book Chapter. p. 283-297. https://doi.org/10.1201/9781003396253-13.
DOI: https://doi.org/10.1201/9781003396253-13

Interpretive Summary: This study focused on improving mapping crop types and other major land cover types at fine scale resolution, which is important for understanding environmental changes affecting food security and ecosystems. We used advanced deep learning and radar images from satellites to map land types in central and north-east Thailand, where small farms are common. Extensive survey data was used to train the a Long Short-Term Memory (LSTM) model and validate the trained model. Results showed that model was very accurate (96% accuracy)performing as well or better than previous methods that needed more complex data. This approach shows promise for reliable land mapping, which can help in environmental management and planning.

Technical Abstract: Effective land cover mapping is crucial for understanding environmental changes and their impacts on food security, ecosystems, and rural communities. To address this need, this study explored advanced deep learning techniques and Synthetic Aperture Radar (SAR) imagery to map crops and other major land cover classes across multiple provinces in Thailand where smallholding farms are common. Utilizing Sentinel-1 SAR data, which provides consistent and reliable observations regardless of weather conditions and daylight, we trained a Long Short-Term Memory (LSTM) model using extensive georeferenced labels to classify fundamental land cover (LC) classes. The model was evaluated across a broad geographic region, achieving a high overall accuracy of 95.99%, with average user and producer accuracies of 0.958 and 0.806, respectively. Our estimates were close to reported inventory data. Our land cover mapping algorithm performed equally or better than previously reported remote sensing-based land cover mapping methods, using only SAR and elevation data as opposed to the multiple features and optical data used in previous studies. This study underscores the potential of combining deep learning techniques with SAR imagery for accurate and reliable land cover mapping.