Surface Water Information Extraction From High Resolution Remotely Sensed Image Based On Integrated Learning

Since the 21 st century,the ability of multi-scale,all-round and real-time dynamic observation of the earth's surface coverage has been further improved,and the ability of remote sensing to obtain images has been increasingly improved.Human beings have entered a new era of earth observation characterized by massive data,multi-source remote sensing,multi-temporal,all-weather information acquisition and rapid and realtime automatic data processing.And we can process real-time data rapidly and quickly.The surface water resources play an important role in the evolution of the earth's environment and human daily life.How to effectively use the rich remotely sensed image data to extract and analyze the surface water has been a widely discussed problem for researchers.Integrated learning algorithms is learning by building and combining multiple learners of the same or different kinds.It combines multiple learners with different combination methods,and usually can obtain a better learning effect and better generalization performance than a single learner.In reality,it is easy to find a kind of simple and inefficient learner.It is of great practical significance to build the simple and inefficient learner into a strong learner with strong learning ability,stability and generalization ability through the integrated framework.In this paper,Decision Tree algorithm and three kinds of integrated learning algorithms including Random Forest(RF),Gradient Boosting Decision Tree(GBDT)and AdaBoost are used to extract water information from the two meter resolution fusion remotely sensed image of GF-1,one meter resolution fusion remotely sensed image of GF-2 and two meter resolution fusion remotely sensed image of GF-6 in Dongting Lake area.On this basis,the design of sample collection proportion is carried out to determine the best sampling proportion,and then the experimental analysis of the feature combination of the best extracted water body is carried out,and the following conclusions are obtained:(1)Decision Tree algorithm and three kinds of integrated learning algorithms are used to extract water information from three highresolution remotely sensed images in Dongting Lake area.The experimental results show that the performance of GBDT is the best.It can complete the classification task in a relatively short time and obtain the best classification effect.Its average classification time for three kinds of remotely sensed images is 0.309 hours.Its average overall accuracy is 0.976,and the average Kappa coefficient is 0.951.Among the four classification algorithms,the quantitative accuracy of GBDT is the highest.(2)In the comparative test of sampling proportion based on GBDT,the quantitative accuracy evaluation shows that 0.1% of the four sampling proportion,0.1%,0.2%,1% and 2%,achieves the best accuracy.Its overall accuracy is 0.991,and kappa coefficient is 0.982.And take the shortest classification time.Experiments show that it is not the larger the sample size,the better the classification effect.On the premise of ensuring the comprehensive and complete collection of water types,0.1% sampling ratio can obtain the optimal classification effect.(3)The GBDT integrated learning algorithm is used to collect samples at 0.1% of the sample collection ratio.The water body extraction test is carried out on the two meter resolution remotely sensed image after the fusion of GF-1 in Wuhan area.The principal component analysis is carried out on the image and NDVI and NDWI indexes are calculated.From the analysis of experimental results,it can be concluded that the combination of the four converted components of principal component analysis and NDVI and NDWI bands is the best way to extract and classify water body,and the accuracy of the method is improved compared with that only using four spectral bands.