Chuqur o’qitish (Deep learning). Asosiy vazifalari va algoritmlari

Deep learning in satellite image classification

• Satellite images have complex structure and deep features that require more complex operations for recognitions. One of the most important uses of satellite and aerial imagery is the production of land-use maps and it can be done through a process called image classification.
• Since 2012, CNN-based algorithms have been used as an effective way to solve the classification tasks.

19
A simple convolutional neural network structure

Mostly used datasets for satellite image classification

Dataset	Number of classes	Number of images	Image size/ Size of dataset	Spatial resolution
UC-Merced Land Use[5]	21	100	256x256	0.3
IARPA fMoW dataset[6]	63	~1.000.000	256x256	0.5
Sentinel-2, Theia Land Data Centre[7]	17	Various	64x64	10
Kaggle, DeepSat (SAT-4) Airborne Dataset[8]	4	99999	28x28	low
Kaggle, Swimming pool and car detection[9]	2	3748	224x224	10

20

Common class types of remote sensing images for land-use classification [image samples from Google Earth Engine, 2020]

21

Recent results in Satellite image classification

#	CNN based model	Accuracy	Dataset
1	AlexNet	~94%	UC-Merced Land Use
2	CaffeNet	~94%	UC-Merced Land Use
3	GoogleNet	~97%	UC-Merced Land Use
4	FC-CNN	~95%	IARPA fMoW dataset
5	VGG, ResNet, and Inception CNNs	~98,5%	UC-Merced Land Use

22

U-NET Architecture

23
et al. Olaf Ronneberger, 2015

• The architecture contains two paths (encoder and decoder).

• The encoder is just a traditional

stack of
convolutional and max pooling layers to
identify what image contains.

• The second path (decoder) is the symmetric expanding path which is used to enable precise localization using transposed convolutions.
• It is an end-to-end fully convolutional network (FCN), i.e. it only contains Convolutional layers and does not contain any Dense layer because of which it can accept image of any size.

The basic architecture of image classification based on segmentation using deep neural network

24

Transfer learning with CNNs

25

• ImageNet is an image database contains more than 14M images which is organized according to the WordNet hierarchy.
• It is well known that training network without overfitting the dataset should be relatively large, millions of images.
• Although networks that are trained on the ImageNet dataset are widely used as a source of the initialization for network weights in other tasks

Proposed method

26

• Proposed method addresses the semantic segmentation of satellite images by aiming at classifying each pixel as belonging to an agricultural field, buildings and unused area.

• It takes long time to train U-Net model from the scratch.
• In this case, possible way is using trained models for initializing network weights (VGG-16).

Representation of proposed “UNet+VGG16” model

Results

27

Model	Model accuracy	Processing time	Image dataset	Epochs
UNet	89,8%	1700s	~15.500	10
UNet	94,0%	3371s	~15.500	20
UNet+VGG16	91,0%	1352s	~6.000	10
UNet+VGG16	94,3%	2681s	~6.000	20

Dataset: IITB dataset
U-net Model: Satellite image segmentation U-Net model from https://github.com/manideep2510/eye-in-the-sky
Hardware: NVidia, GeForce GTX 1080 Ti, RAM 64Gb, GPU 11Gb

Foydalanilgan adabiyotlar

Aurelian Geron, Hands on Machine Learning with Scikit-Learn Keras&Tensorflow // Second edition Concepts, Tools, and Techniques to Build Intelligent Systems, 2019, 510 pages

Primoz Potocnik, Neural Networks: MATLAB examples // Neural Networks course (practical examples)© 2012

https://www.guru99.com/deep-learning-tutorial.html

https://www.tutorialspoint.com/python_deep_learning/python_deep_learning_de ep_neural_networks.htm

https://www.mathworks.com/help/deeplearning/examples/create-simple-deep- learning-network-for-classification.html