Mekong River Delta crop mapping using a machine learning approach
Agricultural land use and practices have important implications for climate change mitigation and adaptation. It is, therefore, important to develop methods of monitoring and quantifying the extent of crop types and cropping practices. A machine learning approach using random forest classification was applied to Sentinel-1 and 2 satellite imagery and satellite-derived phenological statistics to map crop types in the Mekong River Delta, enabling levels of rice intensification to be identified. This initial classification differentiated between broad and prevalent crop types, including perennial tree crops, rice, other vegetation, oil palm and other crops. A two-step classification was used to classify rice seasonality, whereby the areas identified as rice in the initial classification were further classified into single, double, or triple-cropped rice in a subsequent classification with the same input data but different training polygons. Both classifications had an overall accuracy of approximately 96% when cross-validated on test data. Radar bands from Sentinel-1 and Sentinel-2 reflectance bands were important predictors of crop type, perhaps due to their capacity to differentiate between periodically flooded rice fields and perennial tree cover, which were the predominant classes in the Delta. On the other hand, the Start of Season (SoS) and End of Season (EoS) dates were the most important predictors of single, double, or triple-cropped rice, demonstrating the efficacy of the phenological predictors. The accuracy and detail are limited by the availability of reliable training data, especially for tree crops in small-scale orchards. A preliminary result is presented here, and, in the future, efficient collection of ground images may enable cost-effective training data collection for similar mapping exercises.
| Main Authors: | , , |
|---|---|
| Format: | Report biblioteca |
| Language: | English |
| Published: |
International Water Management Institute
2022-12-30
|
| Subjects: | crops, mapping, deltas, machine learning, satellite imagery, land use, land cover, farmland, |
| Online Access: | https://hdl.handle.net/10568/127825 https://www.iwmi.cgiar.org/Publications/Other/PDF/mekong_river_delta_crop_mapping_using_a_machine_learning_approach.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Agricultural land use and practices have important implications for climate change mitigation and adaptation. It is, therefore, important to develop methods of monitoring and quantifying the extent of crop types and cropping practices. A machine learning approach using random forest classification was applied to Sentinel-1 and 2 satellite imagery and satellite-derived phenological statistics to map crop types in the Mekong River Delta, enabling levels of rice intensification to be identified. This initial classification differentiated between broad and prevalent crop types, including perennial tree crops, rice, other vegetation, oil palm and other crops. A two-step classification was used to classify rice seasonality, whereby the areas identified as rice in the initial classification were further classified into single, double, or triple-cropped rice in a subsequent classification with the same input data but different training polygons. Both classifications had an overall accuracy of approximately 96% when cross-validated on test data. Radar bands from Sentinel-1 and Sentinel-2 reflectance bands were important predictors of crop type, perhaps due to their capacity to differentiate between periodically flooded rice fields and perennial tree cover, which were the predominant classes in the Delta. On the other hand, the Start of Season (SoS) and End of Season (EoS) dates were the most important predictors of single, double, or triple-cropped rice, demonstrating the efficacy of the phenological predictors. The accuracy and detail are limited by the availability of reliable training data, especially for tree crops in small-scale orchards. A preliminary result is presented here, and, in the future, efficient collection of ground images may enable cost-effective training data collection for similar mapping exercises. |
|---|