Optimising rice mapping in cloud-prone environments by combining quad-source optical with Sentinel-1A microwave satellite imagery

Abstract

Optical satellite images are prone to cloud contamination in the tropics and subtropics especially during the monsoon season which is contemporaneous with rice cultivation. Producing rice field maps in such areas using a single optical satellite source is therefore very challenging. To obtain an adequate number of usable optical images that captures the seasonal profile of rice fields for their discrimination, this study explored an operational approach to combining quad-source optical satellite data that include Landsat-8, Sentinel-2A, China’s environment and disaster monitoring and forecasting satellite constellation (HJ-1 A and B), and Gaofen-1 over a test site located in southeast China at two rice growing seasons (2016 and 2017). To minimise inter-sensor differences, a spectral index (SI) dataset containing cross-calibrated Enhanced Vegetation Index (EVI) images from all optical sensors and Modified Normalised Difference Water Index (MNDWI) images from Landsat-8 and Sentinel-2, was derived. As more accurate rice maps have been obtained with the synergistic use of optical and microwave data, the aforementioned datasets were combined with the vertical transmitted and horizontal received (VH) and vertical transmitted and vertical received (VV) polarisation images of Sentinel-1A. The resultant optical, microwave, and combinations of optical and microwave data were used as inputs to the Support Vector Machine (SVM) and Random Forest (RF) algorithms available in ENVI 5.3 and ENMAP Box 2.0, respectively to discriminate rice fields from other land-cover types. Results showed that Sentinel-1A data produced higher mapping accuracies than the quad-source optical datasets. This is attributed to the higher number of Sentinel-1A images at the early stages of rice growth (vegetative phase), a period where changes in rice satellite signals are most abrupt and therefore most diagnostic for discriminating the rice crop from other land-cover types. Additionally, combinations of optical and microwave data produced higher overall accuracies than when used separately, and the highest overall accuracies for both years were achieved with the application of RF on a combination of VH, VV and SI (VHVVSI), yielding 98.43% in 2016 and 96.73% in 2017. Moreover, combining VH with SI (VHSI) produced overall mapping accuracies that are more equivalent to the above (98.40% in 2016 and 96.53% in 2017) than with the combination of VV and SI (VVSI) which yielded 97.22% in 2016 and 93.92% in 2017. This indicates that VH is not only more complementary to optical satellite imagery in rice discrimination, but its combination with SI alone can produce rice distribution maps of the highest possible accuracies in cloud-prone environments.