一区论文

《Automatic land cover classification of geo-tagged fieldphotos by deep learning》

ABSTRACT

 

1.Introduction

Global land covermapping→photographs→field photo collecting及几个图形库  +深度学习

2. Methodology

2.1. Transfer learning

迁移学习是什么：Transfer learning was proposed to apply a pre-trained neural networkto another related problem. A pre-trained neural network will include both itsmodel structure and the network weights trained with large datasets.

两种迁移学习模型修改方式： Fine-tuning means continuing training the pre-trained CNN model with another newdataset, according to the task of interest. This process will adjust thenetwork weights of the pre-trained model to fit its outputs as close to newlabels as possible.（“fine-tuning首先是一个类似MLP的结构，但是除了最后一层的权重和bias是随机初始化的之外，其余的权重是在pre-training中训练好的。用在什么地方都可以”）Featureextraction works by removing the last layer of aCNN model (output layer) and treating the output data of the second last layeras extracted features (also called CNN codes), which are always highdimensional vectors and implicitly represent characteristics of input images. The extracted features then can be analysed by otherclassifying models, such as logistic regression,multinomial logistic regression or support vector machine. In feature extraction, thepre-trained CNN model acts as the image feature extractor in the wholeworkflow.

MLP（多层感知机）http://www.360doc.com/content/15/0414/14/12067640_463128216.shtml

（本实验选用feature extraction，∵the GlobalGeo-referenced Field Photo Library has only nearly 30,000 training samples forlandscape classification, which may cause overfitting ifthe CNN model is fine-tuned until the first layer）

（本实验选用Inception-v3、multinomial logistic regression model，removing the last output layer of the pre-trainedInception-v3 model, the image feature extractor wasthen acquired with the output of 2048 CNN codes (image features). The CNN codeswere then classified by a weighted multinomial logistic regression modelfor land cover type recognition）

2.2. Weighted multinomial logisticregression

网络最后一层多项逻辑回归分类器 softmax

（sigmoid函数只能分两类，而softmax能分多类，softmax是sigmoid的扩展）

模型训练的过程（其中，The process of model training is to find outthe best model parameter  that minimises the difference between thepredicted and the actual probability that samples belong to each category. Thegradient descentmethod is used to search for optimum parameterθ, whichis an iterative algorithm that updates parameterθstep by step）

3. Experiment design

3.1. Data sources
3.2. Model training and validation

the labelled samples are divided into twosubsets: training and testing set……训练过程

a photo belongs to every land cover typesare given by the multinomial logistic regression. The type with the highestprobability will be taken as the predicted type within the photo

4.Experiment result

4.1. Overall accuracy
4.2. Predicted probability
4.3. User and produceraccuracy

5. Discussion

5.1. Model performance
5.2. Sample quality

6. Conclusions

本文贡献：This research is focused on proposing a new approach to applying thedeep learning for geographic and environmental studies. The main contributionis to prove the possibility that artificial intelligence can help with land coverclassification and to evaluate how well the model can perform .

Acknowledgements
References