Mapping Small-holder Rice Farms in Bhutan using Earth Observation Datasets and Machine Learning Model.
Abstract
Although recent advancements in remote sensing technology have the potential to map agricultural land extent, there are several difficulties in accurately representing small-holder rice fields. In Bhutan, the scarce availability of satellite data, ground data, and dynamic terrain makes it difficult to map croplands. This study addresses these challenges using the Open Science tool and Earth Observation. We used Collect Earth Online (CEO), developed by the NASA-USAID SERVIR program and Food and Agriculture Office (FAO) for data collection, and the available cropland dataset from the Regional Land Cover Monitoring System (RLCMS) developed by the SERVIR Hubs. A stratified random sampling approach was used to collect training datasets across the five central rice growing districts (Punakha, Wangdue Phodrana, Sarpana, Paro, and Samtse). A total of 5000 sample plots were collected with a 30 X 30 meter minimum mapping unit. The training points were then used in a Random Forest framework within Google Earth Engine (GEE) to map rice fields. Finally, the rice phenology cycle was used to aggregate the results from the growing season to produce the final composite map. Initial accuracy of 86% was obtained with the F1-score and Kappa scores of 86% and 72%, respectively. This research is part of the SERVIR's work on STEM engagement and Service Development co-developed by SERVIR and Bhutanese partners.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMGC32N0764B