Guide for mapping the bulit-up land in the North China Plain

The data can be interactive inspected at GEE-APP

The data can be downloaded at Google-Drive or Baidu-Netdisk (download code: mg85)

---

Instruction for use

1. Copy this repository to you PC

• Clone this repository using 'git clone https://github.com/wangjinzhulala/North_China_Plain_GEE_Organized.git'
• Or Download the Zip

2. Create a Python environment

• Download and install Anaconda
• Navigate to the repository you just downloaded (e.g., "C:/Users/wangj/Desktop/North_China_Plain_GEE_Organized")
• Open "Anaconda Prompt", then type in "cd C:/Users/wangj/Desktop/North_China_Plain_GEE_Organized"
• Type "conda env create --file environment.yml"

---

Why North China Plain?

• North China Plain is one of China's food base, providing over 1/3 of food to feed 1.4 billion Chinese people.
• North China Plain is one of the fastest urbanized regions on this planet, the urban population increase from ~20% to ~60% from 1990 to 2019

What is the value of these mapping methods?

• Using early time remotely sensed imagery (e.g., Landsat 5 TM) achieve low accuracy than recent data (e.g., Landsat 8 OLI/Sentinel 2 MSI)
• We stack the 3-years image together and use a fitting algorithm (Fourier transform) to capture the temporal feature as input for built-up classification
• The result showed that using Fourier features to boost the classification accuracy of using Landsat 5 TM to be close to using Landsat 8 OLI and Sentinel 2 MSI
• We use temporal correction to remove inconsistent classifications and promote all accuracies of 1990-2019 to >94%.

---

Before using the repository

This repository allows you to reproduce the results of the mapping. Some knowledge is required to use the codes:

• Google Earth Engine (GEE) skills, which can be learned from here. If you know Chinese, feel free to see this tutorial
• Python skills, which can be learned everywhere.
• Virtual environment management of Anaconda

After mastering the above skills, please follow the below steps to reproduce the results.

---

General Workflow

--------------------------------------------------The input data--------------------------------------------------

There are five types of input data for the built-up land mapping, all of them (except for Meteorology data) can be accessed at the GEE.

• Spectral: the cloud-free image of Landsat/Sentinel
• Indices: the NDVI/EVI/NDBI computed from Landsat
• Fourier: Coefficients of the Discrete Fourier Transformation on indices images
• Meteorology: China Meteorological Forcing Datahere
• Terrain: Elevation and Slope form the Shuttle Radar Topography Mission

Input type	Source	Spatial resolution	Number of bands	Years
Spectral	Landsat TM	30 m	7	1990-2010
	Landsat ETM+	30 m	9	2011-2013
	Landsat OLI	30 m	11	2014-2019
	Sentinel-2A MSI	10 m	13	2015-2019
Indices	NDVI	30 m	1	1990-2019
	EVI	30 m	1	1990-2019
	NDBI	30 m	1	1990-2019
Fourier	Coefficients of the Discrete Fourier Transformation	30 m	24	1990-2019
Meteorology	China Meteorological Forcing Data	1°	7	1990-2019
Terrain	Elevation	30 m	1	1990-2019
	Slope	30 m	1	1990-2019

--------------------------------------------------The study area--------------------------------------------------

Five middle and eastern provinces of China corresponding to the North China Plain region were selected as the study area. The area spanned 780,000 km2 and five provinces (i.e., Henan, Hebei, Shandong, Anhui, and Jiangsu) and two metropoles (Beijing and Tianjin). The study area is one of China’s fastest developing regions, with the urban population rate (excluding the two metropoles) tripling from ~20% in 1990 to ~60% in 2018. The North China Plain holds a strategic position in China in terms of economic development and food security, generating ~37% of the gross domestic product and ~35% of China’s grain production in 2019.

--------------------------------------------------The workflow--------------------------------------------------

The general work flow is as follows:

• Preprocessing of input image data.
• Creating control points and visually inspect each of them.
• Conduct the classification using Random Forest
• Apply a Temporal-correction to remove inconsistent classifications
• Compare the "overall accuracy" and "area change" between this study and other datasets.

_________________________________________________________

Preprocessing of input image data

----------------- Determine the best Stack-years, Harmonic number and export the Fourier images -----------------

The Discrete Fourier Transformation approximates a series of discrete values by summing up a linear function and several pairs of sinuate functions. A figure captures such fitting as below:

To make the production of Fourier transform more easily deployed, we have packed it into a class in the path North_China_Plain_GEE_Organized/blob/master/Process_1_GEE_Python_Classification/BackGround_modules/Class_1_Make_fourier_imgs.py. After that, every time we want to apply the Fourier transform to a stack of images, we just need to import this class, specify the image-stack to be fitted and other parameters.

The Fourier transform was performed first because we need to figure out how much data will be used in the Fourier transform. For example, if we use 2 years of data for the transform, then we will also need to use two years of data to create the cloud-free image of Landsat. As a result, the images used (aka, the stack-year) for Fourier transform determines the other data are producing.

The code for find the optimum Stack-years and Harmonic number is in North_China_Plain_GEE_Organized/Process_1_GEE_Python_Classification/Sub_Process_1_Data_preparation_Fourier_transformation

There are two Jyputer Notebook files in this path:

Step_1_Detemine the best Harmonic number.ipynb is to Determine the best "Harmonic number" for creating the Fourier images.

Specifically, 100 random points were distributed through the research area, and the mean error between the original value and the fitted value for each point was computed with different harmonic numbers and stack years. The harmonic numbers were set to 1–10, and the stack years were set to 1–5 (where 1 means using only the normalized data from 2015, while 5 means using all the normalized indices from 2015–2019). The harmonic number was determined to be 3, i.e., where the most significant drop in mean error occurred. Fewer harmonic numbers are preferred as they produce fewer coefficients for later classification. The stack year was also determined to be 3 by balancing the data used for the discrete Fourier transform and the mean error decrease. Fewer stack years are preferred because built-up land can be mapped at a higher frequency if fewer data are used for the discrete Fourier transform.

Step_2_Create_Fourier_imgs.ipynb is to export the Fourier images to GEE_assest.

Specifically, we:

1. Loop through 1990-2019 by 3-year intervals;
2. Create the Fourier image using a 3-year stack of NDVI/EVI/NDBI, respectively;
3. Export the Fourier image to Asset with the name "AmplitudePhase{year}"

---------------------------- Create the NDVI/EVI/NDBI images (Indices predictor) ----------------------------

The code for this step is in North_China_Plain_GEE_Organized/Process_1_GEE_Python_Classification/Sub_Process_2_Data_preparation_Create_Normalized_index_images. Only one Notebook in this path, which will:

1. Stack 3-years Landsat data;
2. Compute the mean image of the 3-years Landsat data;
3. Create the NDVI/EVI/NDBI from the mean image

---------------------------- Create the cloud-free images (Spectrum predictor) ----------------------------

The code for this step is in North_China_Plain_GEE_Organized/Process_1_GEE_Python_Classification/Sub_Process_3_Data_preparation_Create_Landsat_Sentinel_Cloud_free_image. Only one Notebook in this path, which will:

1. Stack 3-years Landsat/Sentinel data;
2. Apply the "simpleComposite" to create Landsat cloud-free image,use 'QA' band to to create Sentinel cloud-free image;
3. Export the result to Assest

---------------------------- Create the Meteorology images (Meteorology predictor) ----------------------------

The code for this step is in North_China_Plain_GEE_Organized/Process_1_GEE_Python_Classification/Sub_Process_4_Data_preparation_Prepare_Meterological_data/. Two Notebooks are in this path:

Step_1_Convert_NetCDF_to_individual_TIF.ipynb is to convert the raw NetCDF meteorology data to GeoTiff format. This step is not necessary for the analysis of this study, and it requires the ArcGIS pro Python library.

Specifically, we:

1. Import each NetCDF as multiband tif;
2. Export each band to a single Tif file

Step_2_Composite_Meterology_data_into_3_year_mean_image.ipynb is to compute the mean image of meteorology data in GEE.

Specifically, we:

1. stack 3-years of meteorology data;
2. create the mean image;
3. export the mean image to Asset

---

Creating control points

Because the conversion from non-built-up land to built-up land is unlikely to occur, built-up samples collected using Landsat base-maps from 1990–1992 were used for classification in 1990–2013. The non-built-up samples collected using Google Earth High Definition (HD) maps were used for the classification from 1990 to 2019. Given the 30-year research period, a few sample points may be incorrect. The built-up samples were re-inspected using the Google Earth HD map of 2014 and then used for classification in 2014–2019. We also re-inspected samples to ensure high accuracy in the last two classifications, which will be used as masks to remove inconsistencies in the former classifications.

---------------------------- Collection of built-up samples ----------------------------

The raw built-up samples were taken from the National Settlements Database of China. These records were generated in 2000 and comprised two types of settlement: government sites (including the department offices of provinces, cities, districts, counties, towns, and villages) and the offices of nationally owned companies. The total number of National Settlement points of the study area is 751,411, exceeding the analysis capacity in this study. We randomly subset 5,000 points from the total dataset, then used historical Landsat images to visually check each point and further diminish the number to 4,000 by excluding low-quality points (e.g., those near water bodies or in hilly areas).

Because of the low quality of Landsat data from 1990 to 1992, two false-color base maps (one map created using NDVI, NDBI, and EVI; the other map created from the coefficients of the temporal features) were used to assist with a visual inspection. Each sample point was inspected against all three base maps. We manually nudged their position to the center of nearby built-up land for some points located at positions that could be easily misclassified (such as the edge of a village or a skim road). The visuall check can be see with inspection video

---------------------------- Collection of non-built-up samples ----------------------------

The code for this step is in North_China_Plain_GEE_Organized/Process_1_GEE_Python_Classification/Sub_Process_5_Random_stratify_to_create_non_built_control_samples.

A stratified sampling scheme was used to create the raw non-built-up samples. NDVI was used to stratify raw samples because it can distinguish different land covers effectively, thus promoting even distributed non-built-up samples among different land covers. The raw non-built-up samples were produced with the following procedures. First, NDVI data were produced from the cloud-free image of the research area in 2017–2019. Then 50,000 random points were generated to extract the value of NDVI. Next, the random points were reduced to 5,000, where the histogram of NDVI data was used to stratify the reduction. Finally, these 5,000 points were visually checked.

Non-built-up samples were visually inspected using the Google Earth HD map of 2019. Points located in built-up lands were removed. Points located close to built-up lands were manually nudged to nearby non-built-up land to avoid interference.

---

Conduct the classification using Random Forest

------------------ Before classification, make preparation and determine some parameters ----------------------

To make the classification easier to deployed, we packed it into a python-class in North_China_Plain_GEE_Organized/blob/master/Process_1_GEE_Python_Classification/BackGround_modules/Class_2_Classify_Fourier_Img.py. After that, every time we want to classify some images, we just need to import this class, specify the input images, bands involved in the classification and the training sample.

The code for this section is in North_China_Plain_GEE_Organized/Process_1_GEE_Python_Classification/Sub_Process_6_Before_classification_Feature_selection/. There are 4 steps for this section:

Step_1_Extract_img_value_to_sample_points.ipynb is to extract the input image's value to control points.

Why extract image value to control points? Because by doing so we can directly use the control points with image's value to train the classifier, otherwise a lot of time would be wasted during classifier training on "Extracting image value to points"

Specifically, we:

1. stack all input images into a multiband image;
2. extract the multiband image value to control points;
3. export the points of extraction

Step_2_Determine the best tree number and compare diff bands performances.ipynb runs a sensitive test between tree-number/different input-bands to accuracy.

We used the sklearn.model_selection.GridSearchCV module to test the impacts of tree number on accuracy. We found no accuracy gains were achieved with more than 100 trees. Thus we set the tree number to 100. We also investigated control sample sizes from 0.5% to 99% of the sample and computed the corresponding accuracy. We found that ~50% of the control samples were sufficient to high accuracy. In this study, 75% of the control samples were used for built-up land mapping, among which 70% were used to train the RF classifier. As a result, 52.5% (75% × 70%) of control samples were used to train the RF classifier, which was sufficient for stable classification.

The below two figures demonstrate the incorporation of Fourier input-bands improved the classification a lot.

Step_4_Test_the_sample_size.ipynb run a sensitive test for sample-size to accuracy.

Specifically, we:

1. loop through the sample size from (0.05% to 99%) of the control sample;
2. calculate the accuracy

------------------ Use the determined parameters to perform the classification ----------------------

The code for this setction is in North_China_Plain_GEE_Organized/Process_1_GEE_Python_Classification/Sub_Process_7_Classification_on_img/Step_1_Classification_random_split_10_layers.ipynb, which classifiy the input image to get built-up land maps. Because different training samples produce different classifications, we repeated the classification 10 times with a different sample splitting state (i.e., a seed number set from 0 to 9).

Specifically, we:

1. loop through each year-range (1990-2019 at 3-year intervals);
2. loop through each seed number (0-9) to create 10 classifications with diff samples;
3. export classification to Asset

-------------- Sum the 10 classifications and apply a threshold to produce the final classification --------------------

The code for this setction is in North_China_Plain_GEE_Organized/Process_1_GEE_Python_Classification/Sub_Process_8_Determine_the_threshold_for_the_sum_of_10_random_classification/, there are two Notebooks for this section.

Step_1_Threshold_For_sum_of_10_random_classifications.ipynb is to find out the optimum threshold to produce the final classification out of 10 classifications each with a different training sample

Specifically, we summed the 10 classifications and created the final classification to be the pixels greater or equal to different thresholds. It shows that a threshold of 4 led to the highest accuracy. We further make a map to shows misclassifications of bare lands or farmland rotations being removed after applying the threshold, specifically, we:

1. sum all 10 classifications to get an image with a value from 0 to 10;
2. loop through each threshold, convert the pixel >= threshold to 1, others to 0;
3. calculate the accuracy

Step_2_Create_10_folds_corrected_img.ipynb is to export the final classification computed from the sum image of 10 classifications with diff training sample

Specifically, we:

1. sum all 10 classifications to get an image with a value from 0 to 10;
2. loop through inband combination (this is for making the comparison map of diff in-bands);
3. using 4 as the threshold, the pixel >=4 were converted to 1, others to 0;

Apply a Temporal-correction to remove inconsistent classifications

To make the temporal-correction easier to deployed, we packed it into a python-class in North_China_Plain_GEE_Organized/blob/master/Process_1_GEE_Python_Classification/BackGround_modules/Class_5_Temporal_consistency_check.py. After that, every time we want to temporal correct a series of built-up land maps, we just need to import this class, specify the input maps and mask number.

The below figure shows how the temporal-correction works.

-------------------- Apply the temporal check --------------------------

The code for this setction is in North_China_Plain_GEE_Organized/Process_2_Temporal_Check/Step_1_Temporal_Check.ipynb.

Before applying the temporal correction, the number of classifications (i.e., mask number) used to create the mask should be determined. We ran a sensitivity test to compute overall accuracies for 5 mask numbers. As the mask number increased, overall accuracy improved. We set the mask number to 2 because it is close to the second-highest overall accuracy with fewer classifications used in the mask. If we had set the mask number to 4, the maps of the last four periods would be used as a mask and could not be temporally corrected, which was not practical given there were 10 maps in total.

Specifically, we:

1. loop through "mask classification number";
2. calculate the accuracy and determine the best "mask classification number" to be 2 (where window size is 3);
3. loop throuth the iteration number;
4. calculat the area change with diff iterations and determine the iteration number to be 8;

After determining the mask number, we ran another sensitivity test to determine the temporal correction's iteration number. We found that after 8 iterations, the built-up area remained stable. This pattern can be found for all classifications from 1990 to 2019. As a result, we determined the iteration number to be 8.

The temporal correction has removed inconsistent classification.

Lastly, we get the final bulit-up land maps of the study area:

-------------------- Compare this study with other datasets --------------------------

The code for this setction is in North_China_Plain_GEE_Organized/Process_2_Temporal_Check/Step_2_Area_Accuracy_Comparision.ipynb.

Specifically, we:

1. import other datasets;
2. remap the bulit-up land pixel of all dataset (including this study) to year;
3. loop throuth each Dataset-year bands;
4. calculat the area and accuray;