1 of 17

3/24/2022

Maize classification in Northern Ghana

Prepared by Prof Jianxi Huang, Prof Xuecao Li (CAU)

Presented by Dr Qingling Wu (UCL)

College of Land, China Agricultural University

2 of 17

The principle of classification

3 of 17

Principle of monitoring land cover by RS

Objects on the earth are continuously emitting, reflecting, and absorbing electromagnetic waves. The electromagnetic wave characteristic of different objects is diverse, people can distinguish them using these information.

Remote Sensing

4 of 17

Reflectivity characteristics of different ground objects

5 of 17

Name	Formula
NDVI	(NIR – Red) / (NIR + Red)
LSWI	(NIR – SWIR1) / (NIR + SWIR1)
IRECI	(Edge3 – Red) / Edge1 / Edge2
EVI	2.5 * ((NIR – Red) / (NIR + 6 * Red – 7.5 * Blue + 1))
GCVI	NIR / Green - 1
…	…

Spectral bands for the Sentinel-2 sensors

Cloud, soil

6 of 17

Spectral characteristics of multispectral remote sensing

The spectral albedo value of remote sensing (sensor) at different frequency bands

7 of 17

Classifier

Classifier training: to find the decision boundary to separate different classes

Classes in Spectral Space

K-Means

Random Forest (RF)

Support vector machine (SVM)

Neural networks

…

8 of 17

Random Forest Classifier:

A collection of CART (classification and regression trees) classifiers.
The voting method is adopted to solve the classification method.

Random forest classifier

Similarity

Lots of

high-quality samples

+

Actually, random forest classifier is a collection of CART tree classifiers. Every CART tree classifier will random select samples and features. After the training procedure, the voting method is adopted to solve the classification method. That is to say, if most classifiers choose this place is maize, eventually the output result is maize.

Note: as a supervised classification method, random forest requires a large number of high-quality samples to obtain good classification results.

Of course, the characteristics of ground objects will also affect the classification results. If the differences within the same ground objects are small and the differences between diverse ground objects are large, we will get a good result. On the contrary, we can also take an example, for the classification of maize, the main influence factors are phenology and growth trend. In southwest China, the sowing period of maize is from February to May, during which maize can be sown, resulting in a large difference, and it is difficult to obtain good classification results.

9 of 17

Collect sample data (field experiment)

Location

+

Label

10 of 17

Sample data collection (human interpretation)

11 of 17

Case of maize classification in Ghana

12 of 17

Framework

(number of trees:100)

Original band	Index
Red Edge 1	NDVI
NIR	LSWI
SWIR 1	IRECI
SWIR 2	GCVI

Code:

https://code.earthengine.google.com/4795796bb1f47ff7e9ca9c0aae263c11

13 of 17

Distribution of samples

14 of 17

Spectral curves of crop samples

15 of 17

Crop	others	maize	Users accu.	Overall accu.	Kappa score
others	7823	1234	0.86	0.84	0.68
maize	1800	7975	0.82
Producers accu.	0.81	0.87

training set: 70% samples, validation set: 30% samples

Classification result: https://code.earthengine.google.com/?asset=users/xianda19/classification_result/2021/Ghana/maize_20210501_20211028_70percentSamples

training set: 100% samples

Classification result:

https://code.earthengine.google.com/?asset=users/xianda19/classification_result/2021/Ghana/maize_20210501_20211028_100percentSamples

Accuracy assessment

16 of 17

Classification result of maize

Resolution:

10m

17 of 17

Thank You