DL-RockClassification

Problem description

Deep learning for rock classification, which is similar with semantic segmentation in computer vision.
Input a multi-band remote sensing image, the purpose is to classify each pixel in the image into different classes of rocks.

Model: Conditional GAN

The model has to parts: Generator(a segmentation network) and Discriminator(an adversarial network)

Generator

Generator follows the design of U-net.

In this model, using a list to control the number of filters of each encoder or decoder. And the last layer is using softmax for multi-class classification.
Using AlphaDropout, and dropout rate is lower as network goes deeper, using Conv2DTranspose to do the deconvolution.

Discriminator

Discriminator takes the multi-band remote sensing image and corresponding label as input, outputs a number in range [0,1], which indicates the probability of that the input label is not produced by the Generator.

Phase: Train Discriminator

To train discriminator, we minimize the loss:
$$loss_d(\theta_d) = \sum_{n=1}^{N}l_{bce}(D(X_n, Y_n), 1) + l_{bce}(D(X_n, G(X_n)), 0),$$ where $l_{bce}$ is the binary crossentropy, $D(x,y)$ is the discriminator, $G(x)$ is the generator, $\theta_d$ means the parameters of dicriminator. In this training phase, we try to trian discriminator to recognize the ground truth label from prediction of generator. And the parameters of generator $\theta_g$ are fixed during this phase.

Phase: Train Generator

To train the generator, we minimize the loss:
$$loss_g(\theta_g) = \sum_{n=1}^{N}l_{cce}(G(X_n), Y_n) + \lambda l_{bce}(D(X_n, G(X_n)), 1),$$ where $l_{cce}$ is the categorical crossentropy, $\lambda$ is the weight to balance two part of this loss, $\theta_g$ means the parameters of generator. In this training phase, we try to optimize generator to predict the ground truth distribution and also improve the prediction to fool the discriminator. And the parameters of discriminator $\theta_d$ are fixed during this phase.

Data preprocessing

use arcpy jupyter notebook API to open pre-processing.ipynb
all images are re-sampled into 10m spatial resolution.
for simplicity, the large remote sensing images are cropped into small tiles with size of 256X256.
each band is normalized by (-max)/(max-min)

ASTER

As an advanced multispectral sensor launched onboard Terra spacecraft in December 1999, ASTER covers a broad ragne of spectral region with 14 spetral bands, including three VNIR bands with 15m spatial resoltion, six SWIR bands with 30m spatial resolution, and fice TIR bands with 90m spatial resolution.
Time: 01/04/2014 - 01/06/2014
Cloud coverage: 0.0%-0.0%

Table of ASTER

Band	Central Wavelength(nm)	Spatial Resolution(m)
1	0.5560	15
2	0.6610	15
3N	0.8070	15
3B	0.8070	15
4	1.6560	30
5	2.1670	30
6	2.2090	30
7	2.2620	30
8	2.3360	30
9	2.4000	30
10	8.2910	90
11	8.6340	90
12	9.0750	90
13	10.6570	90
14	11.3180	90

Sentinel-2A

The Sentinel-2A image contains 13 spectral bands in the VNIR and SWIR spectral range, with four bands at 10m, six bands at 20m, and three atmospheric correction bands at 60m spatial resolution. The cloud free image was automatically atmospherically corrected using the Sentinel Application Platform software package provided by ESA.
Time: 01/04/2018 - 01/05/2018
Cloud coverage: 0.0%-0.0%

Table of Sentinel-2A

Band	Central Wavelength(nm)	Spatial Resolution(m)
1	0.4430	60
2	0.4900	10
3	0.5600	10
4	0.6650	10
5	0.7050	10
6	0.7400	20
7	0.7830	20
8	0.8420	10
8A	0.8650	20
9	0.9450	60
10	1.3750	60
11	1.6100	60
12	2.1900	20

Geophysical data

Calculate NDVI

$$NDVI = \frac{NIR - VIS}{NIR + VIS}$$ using Sentinel-2A band 4 as VIS and band 6 as NIR

composite all useful bands into a multi-band raster

band	mean
0	Blue
1	Green
2	Red
3	VRE
4	VRE
5	VRE
6	NIR
7	SWIR
8	SWIR
9	VRE
10	K
11	Kpcent
12	TH
13	U
14	U2
15	magnetic
16	K over TH
17	K over U
18	TH over K
19	TH over U
20	U2 over TH
21	U over K
22	U over TH

layer number and rock type

Layer	Rock
0	Vegetation
1	Unkown Rocks
2	Carbonate_sediment
3	Dolerite
4	Feldspathic_sediment
5	Felsic_volcanic
6	Gneiss
7	Granite
8	Mafic_volcanic
9	Quartz_sediment

chenyz0601 / dl-rockclassification Goto Github PK