Employing optical and SAR imagery for enhanced mapping of vegetation and crops in challenging environments

Abstract

Multi-temporal optical and synthetic aperture radar (SAR) imagery has made significant achievements in vegetation mapping and monitoring in regions with low cloud coverage and homogeneous vegetation distributions. However, existing studies face challenges in mapping abrupt vegetation disturbance and crop cultivation in challenging environments such as urban, cloud-prone and terrain-fragmented environments. First, frequent cloud cover and rainy events usually lead to missing observations in optical imagery and unintended noises in SAR time series, resulting in low data availability and quality in cloudy regions and posing challenges in accurate mapping of abrupt vegetation disturbance caused by natural disasters such as typhoons. Second, the high heterogeneity of vegetation planting and cultivation in urban and mountainous areas leads to the severe mixture of observed satellite signals. The complex landscape with the mosaic of greenspaces and buildings in urban environments may weaken signals of typhoon-caused vegetation changes, while smallholder farming in mountainous areas produces a host of small rapeseed parcels with irregular shapes and complex crop planting structures, increasing the omission and commission error in mountainous rapeseed mapping. Third, terrain-caused uncertainties are persisting bottlenecks for SAR-based applications in mountainous areas due to invalid observations in shaded areas and the foreshortening effect caused by terrain relief, causing large variations in SAR backscattering intensity and instability in the temporal pattern of SAR signals. Last, existing satellite-based methods for rice and rapeseed mapping are not robust for cross-regional and large-scale applications as existing studies rely on prior knowledge and training samples and fail to harmonize differences among sites. Therefore, this thesis endeavours to investigate how to better utilize optical and SAR remote sensing for enhanced mapping of vegetation and crops in challenging environments by addressing the following key questions: (1) Can optical and SAR images be used for quantifying abrupt vegetation disturbance in complex urban landscapes? (2) How can we achieve the fusion of SAR time series and available cloud-free optical data for large-scale paddy rice mapping in cloud-prone environment? (3) How can SAR time series overcome the terrain-caused uncertainties and serve for rapeseed mapping in terrain-fragmented environment? To address question 1, I investigated the feasibility of multi-temporal optical and SAR in urban vegetation damage mapping affected by super typhoon Mangkhut. The vegetation damage maps from Sentinel-2 have an overall accuracy of 97% using the very high-resolution WorldView-3 images as reference data and outperformed than that from Sentinel-1 in complex urban landscapes. But SAR-based method can provide complementary information about the canopy loss in dense vegetated area, suggesting they should collaborate for vegetation damage assessment in urban environment. To address question 2, I proposed a novel SAR-based Paddy Rice Index (SPRI) to quantify the probability of land patches planted in paddy rice. SPRI fully uses the unique features of paddy rice during the transplanting-vegetative period in the Sentinel-1 time series. With the assistance of cloud-free Sentinel-2 images, SPRI can be calculated for each cropland object with adaptive parameters. Results show that the SPRI was able to produce an accurate classification map with an overall accuracy of over 88% and an F1 score of over 0.86 at all five sites with diverse climate conditions and cropping systems. Compared with the existing SAR-based rice mapping methods, SPRI performed much better in heterogeneous agricultural areas where rice is mosaiced with other crops. SPRI does not need any prior knowledge or reference samples and has high flexibility and applicability to support paddy rice mapping in large areas, especially for cloudy regions where optical remote sensing data is often not available. To address question 3, I proposed a phenology-and terrain-adapted method for rapeseed mapping in cloudy and mountainous regions using dual-aspect Sentinel-1 time series. I first reconstructed the Sentinel-1 time series using an iteratively valley-filling method to eliminate unintended noise from terrain and frequent rain. After investigating the backscattering mechanisms of different phenological stages of canola growth, I concurrently quantified the shape and intensity pattern of a given pixel with adapted phenological and terrain parameters based on high-confidence sample sets extracted from a strict discriminator automatically. Our method showed superiority in canola mapping in cloudy mountainous regions over existing methods. In sum, in this thesis, I (1) assessed the feasibility of two typical remote sensing data on quantifying the abrupt vegetation disturbance in complex urban landscapes, (2) proposed a robust index combining SAR time series and available optical images for large-scale rice mapping, and (3) explored the application of dual-aspect SAR time series on the rapeseed mapping in cloudy and mountainous regions. Overall, this thesis provided a comprehensive investigation of better mapping and monitoring vegetation in challenging environments using optical and SAR remote sensing.