University of Twente Student Theses


Detecting ratoon rice and mapping its distribution using machine learning algorithm and Sentinel-1 time-series data

Rathore, Jitender (2021) Detecting ratoon rice and mapping its distribution using machine learning algorithm and Sentinel-1 time-series data.

Link to full-text:
(only accessible for UT students and staff)
Abstract:There is a large gap between rice production and consumption in the Philippines. Therefore, ratoon rice is practised to produce more rice on the same land with less labour and fertiliser; and without land preparations. Although mapping and detecting rice as the primary crop using remote sensing has received a deserved attention in the literature, mapping and detecting ratoon rice has been given less attention. This study aims to detect ratoon and non-ratoon rice and its distribution using Sentinel-1 time series data and random forest algorithm in the Leyte, Iloilo, and Agusan del Sur provinces in the Philippines. Field data was provided by the International Rice Research Institute (IRRI) and included field survey data of a total of 317 fields, in which ratooning was practised in 47 fields. The pre-processed Sentinel-1 image data acquired during three growing seasons between 2017- 2019. Temporal backscatter behaviour of 60 rice fields in different polarisations and the Mann-Whitney U test were used to understand the differences between rice fields where ratooning was practised (n=30) and where it was not (n=30). Then, a random forest (RF) machine learning algorithm was used to discriminate between rice and ratoon rice. The predictive performance of the RF classification model was checked by overall accuracy and kappa values. The RF model was also performed to classify ratoon and non-ratoon rice crops using different ratoon growth stages. Finally, the distribution of ratoon and non-ratoon fields were mapped using a validated RF model. In this study, we demonstrated that there is a clear difference in temporal backscatter of ratoon and non -ratoon rice crops. The results of the Mann-Whitney U test revealed that the backscatter of ratoon and non-ratoon fields when ratoon crops are at the flowering and ripening stages are significantly different in VH, VH/VV polarisations. When the random forest (RF) classifier was performed to discriminate ratoon, and non-ratoon classes, an overall accuracy (69.39%) and kappa value (0.39) were obtained. The RF model was calculated at different ratoon growing stages, demonstrated an overall accuracy of 44.44% at vegetative, 66.67% at flowering and 61.11% at ripening stages. The distribution of ratoon and non-ratoon rice fields showed that Iloilo province has the majority of fields with ratoon rice. We concluded that the sentinel-1 time series could detect ratoon and non-ratoon rice at different stages using the RF model. The difference in ratoon and non-ratoon could be studied using VH polarisation and VH/VV ratio at the ripening stage where mean values of temporal backscatter were dB>1. Also, the VH and VH/VV were statistically significant at the flowering stage. Hence more samples are required to study if ratoon and non-ratoon fields can be discriminated at this stage. Further study in provinces where ratoon and non-ratoon are practised should also be explored.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page