Irrigated rice-field mapping in Brazil using phenological stage information and optical and microwave remote sensing

Abstract

Irrigated rice-field mapping methodologies have been rapidly evolving as a result of advanced remote sensing (RS) technology. However, current methods rely on extensive time-series data and a wide range of multi-spectral bands. These methods often struggle with classification accuracy with contaminated satellite data due to environmental factors or acquisition device constraints, e.g., cloud cover, shadows, noise, and the temporal and spectral resolution trade-off. Our goal is map irrigated rice-field by using a suitable satellite image band composition instead of time-series data. We divide the growth cycle into different rice phenological stages: beginning, middle and end of season, as well as the season transition periods. Near-infrared (NIR), short-wave infrared (SWIR) and red bands of MultiSpectral Instrument - MSI/Sentinel-2 (optical RS), along with polarizations of VV (vertical–vertical) and VH (vertical–horizontal) of Sentinel-1 C-band Synthetic Aperture Radar (SAR) (microwave RS), were used to create ten different false-color image composites. Ground truth maps from two consecutive growth seasons (2017/2018 and 2018/2019) served as references. We applied a modified version of the Fusion Adaptive Patch Network (FAPNET), named as Patch Layer Adaptive Network (PLANET) convolutional neural network (CNN) to obtain binary rice mapping, which was evaluated using the traditional Mean Intersection over Union (MIoU) and Dice coefficient. Analytic results show that the end of season is the most suitable for obtaining a reliable classification based on optical and SAR sensors. Although complex rice-field pose challenges, our predictions consistently scored a MIoU above 0.9. We conclude that choosing the right phenological stage for rice mapping combined with deep learning model can greatly improve the classification results. These results indicate that the choice of composition significantly impacts classification accuracy, especially in more complex environments.