Precise Recognition And Above Ground Biomass Estimation Of Lei Bamboo Forests Using Deep Learning

Recently,artificial intelligence has become one of the hot-spot topics of our world.As the blooming of unmanned aerial vehicles(UAV)and high-resolution satellites,extracting high-level representation from the spectral and spatial information of high spatial resolution imagery is drawing more and more attention.Deep learning and machine learning algorithms such as Deep Convolutional Neural Networks(DCNN),could reach some levels of intelligence concerning remote sensing information extraction,forest parameter retrieving and other comprehensive data analysis by training model with enormous data,which are potential methods for automatically monitoring of forest resource.Subtropical forests,especially bamboo forests,play a crucial role in carbon sinks.However,on the one hand,subtropical forests are very complex.It is hard to recognize a single tree species even from very high spatial resolution imagery.On the other hand,forest above ground biomass(AGB)directly relate to forest productivity and ecosystem function.The precise estimation of AGB has drawn much attention for years.Very high-resolution remote sensing(VHRRS)imagery provides valuable spatial information,but there are lots of difficulties concerning how to make use of these information to construct relationship between spatial information and AGB,obtain more samples,and for AGB biomass automatically estimation.In this study,Worldview-2 the VHRRS imagery is used for Lei bamboo forest AGB automatically monitoring based on Deep Learning methods.The study area locates in Taihuyuan,Lin’an,Hangzhou city.This study is mainly composed of:1.The precise classification of Lei bamboo forests based on Deep learning methods and VHRRS imagery.In this part,the study mainly includes:1)the CNN-based classification of Lei bamboo forests and its comparison to Random Forest(RF)with bands as inputs,2)the DCNN-based classification of Lei bamboo forests and its comparison to Random Forest(RF)with Grey Level Co-occurrence Matrix(GLCM)textures and vegetation indices(VIs)as inputs,and 3)The fusion of CNN and RF for classification。2.CNN-based Lei bamboo forest AGB estimate.In this part,the study mainly includes:1)Analysis the contribution of VIs,GLCM textures,ESDA(Exploratory Spatial Data Analysis)textures for AGB estimate basing on VHRRS,2)apply CNN on Lei bamboo forest AGB estimate and compare it with RF,SVM(Support Vector Machine),ANN(Artificial Neural Network)to explore the potentials of Machine learning algorithms in AGB estimate.The study reaches the following conclusions:1.The Lei bamboo forest classification reaches very high accuracy with an overall accuracy(OA)of 0.942 and a Kappa(Kappa co-efficiency)of 0.922.Moreover,the fusion of CNN and RF vastly improves the performance with an OA of 0.991 and a Kappa of 0.990.According to the variable importance analysis,VIs and GLCM textures based on NIR have higher contributions,especially the GLCM textures with a window size of 13.2.Deep CNN with an input window size of 13 reaches very promising accuracy in Lei bamboo forest AGB estimate with a R~2of 0.943,a RMSE of 0.274 Kg/m~2 and a RMSEr of23.1%,which is apparently better than traditional ANN(R~2=0.885)with a difference of around 6%.Although the RF models based on GLCM textures and the all feature set outperformed CNN,CNN only used bands as inputs and extracted high-level representation textures automatically.Using the abstract information from CNN make AGB estimate much more efficient and accurate.As the rapid growing of deep learning technology,CNN has great potentials in precise AGB estimate.