[M-AG32] Satellite Land Physical Processes Monitoring at Medium/High/Very High Resolution

[MAG32-P09] Field-based rice crop identification using high-resolution sequential satellite images and recurrent neural networks

*Son Thanh Nguyen¹, Chi-Farn Chen, Huan-Sheng Lin, Cheng-Ru Chen, Youg-Sin Cheng, Chien-Hui Syu, Yi-Ting Zhang, Tsang-Sen Liu, Shu-Ling Chen , Shih-Hsiang Chen (1.Center for Space and Remote Sensing Research, National Central University, Zhongli District, Taoyuan City 320317, Taiwan)

Keywords:Sentinel-1 SAR data, Field-based classification, Deep recurrent neural networks, Taiwan

Rice monitoring is seasonally an important task in Taiwan. Policymakers need information on rice fields in a large area to estimate rice production in a timely manner to address issues of national food consumption and rice grain exports. The main objective of this research is to develop a classification framework for seasonally monitoring rice-planted fields from high-resolution Sentinel-1 Synthetic Aperture Radar (SAR) data using deep recurrent neural networks (RNNs) in Taiwan. We processed the satellite data for 2018 to 2020 through three main steps: (1) data preprocessing to take into account image registration, noise reduction, and conversion of digital pixel values to calibrated backscatter values (dB); (2) field-based image classification; the image segmentation of the sequential vertical transmit and horizontal receive (VH) data was first applied using the multiresolution algorithm embedded in eCognition Developer 9.0^®. The RNNs, a class of deep learning algorithms using training samples, were accordingly applied to estimate the likelihood of rice area in each field. Specifically, the satellite data (2018 and 2019) were used for model training, leaving the 2020 data as inputs for estimating rice-growing fields; and (3) field-based accuracy assessment; a hardening process was first applied to convert the probability of rice in a mixed object to a pure object using a threshold of 0.5, in respect to two classes of rice or non-rice. The field-to-field comparison between the classification map and ground reference map was then performed to assess the RNN mapping accuracy. The results showed that rice-growing fields were successfully identified, with F-score and Kapa coefficient values of 0.8 and 0.78 for the first crop, and 0.62 and 0.61 for the second crop, respectively. The results of change detection between the first and second crops showed that a remarkable proportion of the rice-producing areas was changed to cash crop cultivation in the second crop due to adverse effects of climatic conditions on the second rice crop production. This study ultimately led to the potential application of sequential Sentinel-1 VH data for mapping rice-growing fields using RNNs in Taiwan. Such methods, providing reliable quantitative information on rice growing areas at the field level, could be transferable to other regions around the world with similar cropping conditions for rice crop monitoring.