the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
SEA-Rice-Ci10: High-resolution Mapping of Rice Cropping Intensity and Harvested Area Across Southeast Asia using the Integration of Sentinel-1 and Sentinel-2 Data
Abstract. The Southeast Asia region has a vast expanse with diverse tropical climates, making it a prominent centre of rice cultivation, contributing to about 20 % of the world’s rice production and contributes 29 % of global rice methane emissions. As a staple food for many countries, accurate and up-to-date information on the rice harvested area is crucial for addressing food security issues, predicting rice yield and methane emissions, and formulating effective government policies. This paper presents the first detailed study of rice cropping intensity and harvested areas throughout Southeast Asia. Current remote sensing products have not been able to produce up-to-date cropping intensity information due to the variability of local cultivation practices. To address this problem, we integrated Sentinel-1A and Sentinel-2A/B time series data from 2020 to 2021 and developed a local unsupervised classification with phenological labelling (LUCK-PALM) method. We implemented the system on the Google Earth Engine (GEE) cloud-based platform to produce geospatial products of rice cropping intensity and harvested area at a spatial resolution of 10 m called SEA-Rice-Ci10s. The results show that Southeast Asia's total rice growing area in 2020–2021 was 28.5 Mha, with 51 % single cropping, 47 % double cropping, and 2 % triple cropping. These were equivalent to 42.9 Mha of annual harvested area, consisting of Thailand (11.2 Mha), Indonesia (8.4 Mha), Myanmar (8.4 Mha), Vietnam (6.3 Mha), Cambodia (3.9 Mha), the Philippines (3.3 Mha), Laos (0.8 Mha), Malaysia (< 0.5 Mha), and Timor-Leste (0.01 Mha). We compared our rice maps to agricultural statistics data at the district and province levels and existing rice maps for some Southeast Asian countries. The results demonstrate that our map agreed well with countries’ statistics with a linear coefficient of determination (R2) from 0.85 to 0.97. Compared to existing products, our map can resolve small paddy fields of about 0.2 ha in the hilly areas. This information will be useful in addressing food security challenges and improving estimates of methane emissions from rice cultivation. The 10 m paddy rice cropping intensity map for Southeast Asia, SEA-Rice-Ci10, is available on the GEE App (https://rudiyanto.users.earthengine.app/view/seariceci2021), the Climate TRACE platform (https://climatetrace.org/) and the Zenodo repository (https://doi.org/10.5281/zenodo.10707621) (Frisa Irawan et al., 2024).
- Preprint
(4933 KB) - Metadata XML
-
Supplement
(1502 KB) - BibTeX
- EndNote
Status: open (until 01 May 2024)
-
RC1: 'Comment on essd-2024-90', Anonymous Referee #1, 25 Apr 2024
reply
The SEA-Rice-Ci10 study devised a novel approach called the Local Unsupervised Classification with Phenological Labeling Method (LUCK-PALM). This method aimed to accurately quantify and map rice cropping intensity and harvested areas across Southeast Asia from 2020 to 2021, with a 10-meter resolution using time-series inputs from Sentinel-1 and Sentinel 2A/B imagery. The LUCK-PALM method is able to resolve the challenge due to the discrepancy in rice cropping calendar among the country and regions. Compared to agricultural statistics and existing rice maps, the study demonstrates its strength in retrieving field paddy details. The study, being open-source and featuring an open dataset, offers valuable insights into agricultural decision-making and management practices, such as methane emissions from rice cultivation, but the quality of the manuscript could be enhanced by addressing the following major and minor comments regarding methodology and content organization, etc. Please refer to the detailed feedback provided for further improvement.
Major comments:
- As shown in Fig.2 and Section 2.3.5, the “Google Map Very High Resolution and Street View” is the data source for exports to label rice and non-rice areas. Although examples and locations of such street view imagery are given in Fig.3, Fig.17 and Figure.S2/Table.S4, it would be good if more details about the street view collection for validations are provided. For example, how was each street view image acquired and how many validation points did the experts label in each grid or each country/province/district? Were those points selected randomly or the selection process was the same as the 2,000 random samples from each defined area?
- The manuscript provided an assumption (line 219-220) that 25 to 30 clusters output from the unsupervised K-Means classification method would sufficiently represent the spectral data variations in each grid – with 2,000 points and 72 bands. Further explicit explanations or justifications for this assumption should be provided to offer readers a clearer understanding of its methods and enhance transferability.
- The study trained local K-Means models for each grid using 2,000 randomly sampled points, and then assessed the accuracy of the produced maps with respect to agricultural statistics and existing products. However, descriptions about how the study applied the locally trained models in combination with expert knowledge to produce grid-wise maps as well as the compilation of maps into national/provincial/regency scales were missing from the manuscript. Also, as seen from the workflow of the study presented in Fig.2, the “Accuracy Assessment” was before the “Map of rice field extent and cropping intensity,” which could potentially lead to the confusion that if the accuracy assessment was conducted with the 2,000 samples each grid or with the final mapping product.
- The methods presented in the manuscript featured the workflow from Sentinel time series to rice / non-rice crop mapping (in terms of spatial distribution), but it appears that descriptions about how the authors retrieve the cropping intensity of rice (from time series-based spectral profiles?) were less elaborated in methods.
Minor comments:
- In Fig.2, it is good to be concise in workflow illustrations, but how high is “Google Maps Very High Resolution” presented here? Also, the capitalization styles of words in Fig.2. could be more consistent.
- In Section 2.3.2 (line 192-195), what is the difference between using different landcovers from the WorldCover dataset to “filter out” non-croplands and using waterbody, tree, and built-up layers from the same dataset to “mask” non-cropland areas? The goal to facilitate computation and processing, as well as improving model performance is clear here, but the description of this step could be clearer.
- 3 appears to be coarser than other figures provided in the manuscript in terms of resolution.
- 16 could have legends and classification accuracy and/or coefficients of determinant labeled on the map for each region.
- Line 642 has a typo: “penological mapping.”
Citation: https://doi.org/10.5194/essd-2024-90-RC1
Data sets
SEA-Rice-Ci10: High-resolution Mapping of Rice Cropping Intensity and Harvested Area Across Southeast Asia using the Integration of Sentinel-1 and Sentinel-2 Data Frisa Irawan Ginting, Rudiyanto Rudiyanto, Fatchurahman Fatchurahman, Ramisah Mohd Shah, Norhidayah Che Soh, Sunny Goh Eng Giap, Dian Fiantis, Budi Indra Setiawan, Sam Schiller, Aaron Davitt, and Budiman Minasny https://doi.org/10.5281/zenodo.10707621
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
333 | 79 | 20 | 432 | 20 | 13 | 15 |
- HTML: 333
- PDF: 79
- XML: 20
- Total: 432
- Supplement: 20
- BibTeX: 13
- EndNote: 15
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1