10:45 〜 12:15
[HTT17-P05] Rice Field Detection Based on Sentinel-1 Multi-temporal Images in Yunlin, Taiwan
キーワード：Synthetic Aperture Radar, Backscatter Coefficient, Paddy Phenology, Wavelet transform
Rice is the primary staple food and plays an important role in the economics and the environment in Taiwan. Most of the rice production is in the western part of Taiwan, for example, Yunlin County. However, ground-based monitoring activities are highly time- and resource-consuming. So a few have been deployed on a large scale. In contrast, monitoring paddy phenology by remote sensing provides large-scale information for crop management and food security. Sentinel-1 Synthetic Aperture Radar (SAR) can provide cost-efficient data which is not affected by clouds. Sentinel-1 multi-temporal backscatter detects the rice field according to the paddy phenology. This study aims to retrieve time series paddy-specific phenology using Sentinel-1 backscatter data and to detect the paddy rice field in Yunlin.
Our study uses 60 Sentinel-1 images which a revisit time is 6 days (1A and 1B) in 2019 and follows three major steps: (1) Extracting specific paddy phenology curves in rice samples. And calculating the mean of the temporal behavior of SAR backscatter coefficients (VH, VV, and the ratio of VV/VH) after doing Wavelet transform to remove the noise and smooth the signal; (2) Defining the start of season (SOS), end of tillering (EOT), and end of season (EOS) in the paddy growing cycle through backscattered signal; (3) Comparing the Cross-correlation coefficients with the specific rice sample field. Moreover, our study adds the cadastral map as the geometric feature for object-based classification to improve accuracy.
In our validation based on object-based classification, the kappa value of Yunlin is 0.64. The overall accuracy is over 0.7 for all townships in Yunlin. In the different townships, the region with the high kappa value is concentrated in the eastern part of Yunlin (Douliu City, Dounan, and Dapi Township, etc.) because of the high density of the rice field. On the other hand, the region with the low kappa value is concentrated in the western part of Yunlin (Tuku and Yuanchang Township, etc.) due to crop diversification. Furthermore, the whole Yunlin‘s Pearson correlation coefficient (r) compared with the kappa index and the size of the area indicates the area of 1000 square meters is the smallest unit of detecting rice field in our study.
Our study uses 60 Sentinel-1 images which a revisit time is 6 days (1A and 1B) in 2019 and follows three major steps: (1) Extracting specific paddy phenology curves in rice samples. And calculating the mean of the temporal behavior of SAR backscatter coefficients (VH, VV, and the ratio of VV/VH) after doing Wavelet transform to remove the noise and smooth the signal; (2) Defining the start of season (SOS), end of tillering (EOT), and end of season (EOS) in the paddy growing cycle through backscattered signal; (3) Comparing the Cross-correlation coefficients with the specific rice sample field. Moreover, our study adds the cadastral map as the geometric feature for object-based classification to improve accuracy.
In our validation based on object-based classification, the kappa value of Yunlin is 0.64. The overall accuracy is over 0.7 for all townships in Yunlin. In the different townships, the region with the high kappa value is concentrated in the eastern part of Yunlin (Douliu City, Dounan, and Dapi Township, etc.) because of the high density of the rice field. On the other hand, the region with the low kappa value is concentrated in the western part of Yunlin (Tuku and Yuanchang Township, etc.) due to crop diversification. Furthermore, the whole Yunlin‘s Pearson correlation coefficient (r) compared with the kappa index and the size of the area indicates the area of 1000 square meters is the smallest unit of detecting rice field in our study.