Mapping Rice Fields,Biomass And Leaf Area Index Using Optical And Microwave Satellite Imagery

Both optical and microwave satellite data have been employed in rice monitoring.Optical images are however prone to cloud contamination especially in the tropics and subtropics.Microwave images on the other hand are less affected by clouds and have become complementary to optical images in rice monitoring.Hence the synergistic use of optical and microwave images in rice monitoring has recorded better results as against their independent use.However,substantial challenges exist in the conjugative use of these datasets in rice monitoring by virtue of their varying spatial and temporal resolutions,and the different temporal profiles of these datasets with rice crop growth.Against this backdrop,this thesis investigated the most optimal dataset(among optical and microwave satellite data)and model(among key statistical and machine learning models)combinations at specific growth periods(scenarios)of rice.For the first time,this study explored the combined use of quad-source optical satellite data(Sentinel-2A,Landsat-8 OLI,HJ-1 and GF-1)with the new Sentinel-1 A microwave satellite data in mapping rice field distribution,dry biomass and green leaf area index(LAI),at two growing seasons(2016 and 2017),at an area located in southeast China.Using random forest(RF)and support vector machine(SVM)for mapping rice field distribution,and statistical(linear,quadratic,logarithmic,power and exponential)models and machine learning[RF,SVM,k-nearest neighbor(k-NN)and gradient boosting decision tree(GBDT)]regression models for mapping rice biomass and LAI,this thesis answered the vital research questions of which algorithm(SVM and RF for paddy rice fields),growth period(for rice biomass and LAI),and dataset among the vertical transmitted and vertical received(VV)and vertical transmitted and horizontal received(VH)channels of Sentinel-1A,and the enhanced vegetation index(EVI)and modified normalized difference water index(MNDWI)from optical satellite imagery,most adequately meet the requirements of precision and operational rice monitoring.In mapping rice fields,EVI and MNDWI were combined to form a spectral index(SI)dataset,resulting in seven satellite datasets as follows;SI,VH,VV,VHVV,VHSI,VVSI and VHVVSI.The VHVVSI with RF combination produced the highest overall map accuracies of 98.43%in 2016 and 96.73%in 2017,and is therefore considered optimal for rice field mapping.However,VHVV with RF recording accuracies of 96.8%(2016)and 94.5%(2017)is regarded optimal for operational rice mapping initiatives.In dry biomass estimation,VHVV with RF is optimal at all stages,with an R2 of 0.73 and an RMSE of 462.4 g/11m2.EVI before heading recorded the most accurate biomass estimates with GBDT,with an R2 of 0.82 and an RMSE of 191.8 g/m2.After heading,there is a substantial drop in the performance of EVI,recording an R2 of 0.49 and an RMSE of 364.7 g/m2 with the quadratic model,indicating that optical imagery can produce more accurate biomass estimates in the first half of the growing season.The poor performance of EVI after heading is attributed to the greater attenuation by leaves and ears.Therefore,to improve on biomass estimates,elongation to milking and elongation to maturity periods were investigated for the synergistic use of optical and microwave imagery.In this regard,improved biomass estimates were obtained at flowering and milking with an R2 of 0.84 and an RMSE of 251.2 g/m2 based on kNN with VHEVI.Beyond milking,the most accurate estimates with the combined optical and microwave data were obtained by VHVVEVI with RF,recording an R2 of 0.70 and an RMSE of 382.2 g/m2.It is therefore obvious that the combined use of optical and microwave imagery produced rice biomass estimates that are more accurate than corresponding estimates generated by the sole use of microwave images(VHVV with RF in the all stages scenario).With EVI capable of producing more accurate estimates before heading,its deficiency after heading can be ameliorated with the combined use of optical and microwave datasets.However,despite the improved biomass estimates obtained with the division of the growth period,their resultant biomass dynamic maps were very consistent with those obtained by VHVV with RF at the all stages scenario.Thus,for operational paddy rice monitoring initiatives,the sole use of Sentinel-1A VHVV data can be sufficient for the estimation and dynamic mapping of rice biomass.In the estimation of green LAI,VHVV also proved most optimal at the all stages,recording the most accurate estimates with GBDT with an R2 of 0.82 and an RMSE of 0.68 m2/m2.Similar to biomass,EVI also performed best before heading,recording its most accurate estimates with GBDT with an R2 of 0.82 and an RMSE of 0.82 m2/m2,as against its estimates after heading with the power model,having an R2 of 0.52 and an RMSE of 1.03 m2/m2.This suggests that as opposed to biomass,VHVV can yield more accurate rice LAI estimates than EVI.The reverse being the case with biomass suggests that green LAI could be more sensitive to image quality than dry biomass.It is important to note that the optical images used in this study are not 100%cloud-free.Remnant atmospheric effects may have compromised the ability of optical imagery to outperform its microwave counterpart.Even LAI estimates from the combined use of optical and microwave satellite data from elongation to milking where VHEVI with GBDT recorded an R2 of 0.84 and an RMSE of 0.71 m2/m2,and from elongation to maturity where VVEVI with RF recorded an R2 of 0.69 and an RMSE of 0.77 m2/m2,LAI estimates generated by VHVV with GBDT at the all stages were still of superior accuracy.Dynamic LAI maps generated at the all stages scenario by the VHVV with GBDT model were consistent with the seasonal trends of LAI,and this has further demonstrated that in cloudy areas,microwave data can be used in lieu of optical data.This thesis concludes that in mapping rice fields,the combined use of multisource optical and Sentinel-1 A data only brought slight improvements in accuracy over the sole use of the VHVV dataset of the latter.Additionally,despite the improved biomass estimates with combinations of multisource optical and Sentinel-1 A imagery,dynamic biomass maps obtained by VHVV were consistent with in-situ data.Moreover,LAI estimates by VHVV are not only consistent with in-situ data but also superior to those of the combined datasets.Based on these results,the sole use of the VHVV dataset of Sentinel-IA cannot only meet the data requirements of precision agriculture,but also being obtained from a single satellite sensor which avoids the huge amount of time required in processing multisource satellite data,makes it optimal for an operational scenario where timely data acquisition is imperative.Given this potential of VHVV,its use should therefore be extended to the retrieval of other biophysical parameters such as height,phenology and yield for complete crop information.This is especially important in cloudy areas where the use of optical satellite imagery is highly limited.