Mapping Plastic-Mulched Landcover Using Optical And SAR Data Based On Machine Learning Classifiers

Plastic mulching on farmland has been increasing worldwide for decades due to its superior advantages for improving crop yields.As a big consumer for plastic mulch,China has witnessed the exponential growth of plastic-mulched landcover(PML)in the past three decades.Monitoring PML effectively can provide essential information for precision agricultural production,management and decision making.However,mapping PML accurately with remote sensing data is still challenging due to sophisticated spectral characteristics caused by differences in plastic mulch,crops and soil types.This thesis proposes two approaches on PML detection in Xinjiang,a major crop-producing area in China using optical and radar remote sensing data based on machine learning algorithms(Random forest,Classification and regression tree,Support vector machine):(1)object-based PML classification using integrated optical and radar data;(2)Large-scale PML mapping using multi-temporal optical and radar data on Google Earth Engine platform.This thesis aims to explore the feature importance for PML classification in integrated optical and radar data,evaluate the performance of machine learning classifiers,and construct the methodological framework for PML extraction at a large scale.Main results of this thesis are as follows:(1)The important features for PML detection are NDVI and VH/VV;(2)Compared with using optical data alone,using integrated optical and radar data can complement each other,raising the PML classification accuracy by 1%–3%;(3)By analyzing the phenology-based time series of NDVI and VH/VV features,two important time windows for PML detection are identified: from April 10 to May25;from June 5 to August 10;(4)The proposed large-scale PML mapping framework using multi-temporal optical and radar data has great feasibility and classification performance using three machine learning classifiers with overall accuracies above 95%.Random forest has the most stable classification result with great tolerance in sample errors,achieving the best overall accuracy of 96.62%.