Comparing Deep Learning models for mapping rice cultivation area in Bhutan using high-resolution satellite imagery

Abstract

Crop type and crop extent are critical information that helps policymakers make informed decisions on food security. As the economic growth of Bhutan has increased at an annual rate of 7.5% over the last three decades, there is a need to provide geospatial products that can be leveraged by local experts to support decision-making in the context of economic and population growth effects and impacts on food security. To address these concerns related to food security, through various policies and implementation, the Bhutanese government is promoting several drought-resilient, high-yielding, and disease-resistant crop varieties to actively combat environmental challenges and support higher crop yields. Simultaneously the Bhutanese government is increasing its utilization of technological approaches such as including Remote Sensing-based knowledge and data products into their decision-making process. This study focuses on Paro, one of the top rice-yielding districts in Bhutan, and employs publicly available Norway’s International Climate and Forest Initiative (NICFI) high-resolution satellite imagery from Planet Labs. Two Deep Learning approaches, point-based (DNN) and patch-based (U-Net), models were used in conjunction with cloud-computing platforms. Four different models per Deep Learning approaches (DNN and U-Net) were trained: (1) Red, Green, Blue, and Near-Infrared (RGBN) channels from Planet, (2) RGBN and Elevation data (RGBNE), (3) RGBN and Sentinel-1 data (RGBNS), and (4) RGBN with Elevation and Sentinel-1 data (RGBNES). From this comprehensive analysis, the U-Net displayed higher performance metrics across both model training and model validation efforts. Among the U-Net model sets, the RGBN, RGBNE, RGBNS, and RGBNES models had an F1-score of 0.8546, 0.8563, 0.8467, and 0.8500 respectively. An additional independent model evaluation was performed and found a high level of performance variation across all the metrics (precision, recall, and F1-score) underscoring the need for practitioners to employ independent validation. For this independent model evaluation, the U-Net-based RGBN, RGBNE, RGBNS, and RGBNES models displayed the F1-scores of 0.5935, 0.6154, 0.5882, and 0.6582, suggesting U-Net RGBNES as the best model across the comparison. The study demonstrates that the Deep Learning approaches can be used for mapping rice cultivation area, and can also be used in combination with the survey-based approaches currently utilized by the Department of Agriculture (DoA) in Bhutan. Further this study successfully demonstrated the usage of regional land cover products such as SERVIR’s Regional Land Cover Monitoring System (RLCMS) as a weak label approach to capture different strata addressing the class imbalance problem and improving the sampling design for Deep Learning application. Finally, through preliminary model testing and comparisons outlined it was demonstrated that using additional features such as NDVI, EVI, and NDWI did not drastically improve model performance.