Pixel-level rice planting information monitoring in Fujin City based on time-series SAR imagery

Abstract

As the core issues of agricultural remote sensing, the mapping of crop planting and dynamic monitoring of crop growth in the urban area have always been limited by complex climatic conditions that induced optical image shortages. Considering that the backscattering signals of SAR image data are sensitive to crop phenological period, the paper presents a fully convolutional neural network (FCN) model to realize the wide-area planting thematic mapping at the pixel level. Taking Fujin City of Heilongjiang province in China as a specific study area, we selected rice as the typical research object and obtained urban-scale rice planting thematic mapping using Sentinel-1A image data derived from April to October 2020. Based on the proposed FCN classification method, the rice extraction precision is up to 95.7%, and the kappa coefficient is 0.90. Comparative analysis indicated that the proposed model is superior to the traditional random forest classification model in extraction precision and kappa coefficient. Further analysis showed that the rice growth cycle in Fujin City was consistent with the time series of VH/VV polarization coefficient ratio and had prominent phenological cycle characteristics. Relevant research data and results proved that high-precision planting thematic mapping at an urban scale could be achieved based on time series SAR image data sets. And the proposed FCN classification model is helping to dynamic monitoring of crop growth and related phenological research.