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.
The model has to parts: Generator(a segmentation network) and Discriminator(an adversarial network)
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 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.
To train discriminator, we minimize the loss:
To train the generator, we minimize the loss:
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)
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%
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 |
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%
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 |
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 | 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 |