Terraformers - SE Challenge

When starting with a deep learning workflow, it's always cheaper to work with existing pre trained models that can accomplish the results you are looking for. When it comes to Rohingya refugee camps. Our first milestone was to get the footprints of the tents and structures built in these camps. To do so we utilized drone imagery shared by  International Organization for Migration  (IOM).

These images where shared for two different areas that you will see in the maps below.

•  Chakmarkul 20190119 
•  Zone PP 

The next step after acquiring the imagery was to download the available deep learning package (dlpk) from Esri's Living Atlas. It's good to note that this model was trained for building footprint extraction in Africa. This said the results that are expected shouldn't be production results. Which means that we will need to retrain the model with more local training data. In this case refugee tents data from the Rohingya refugee camps in Cox's Bazar.

After Acquiring the imagery and the dlpk. We ran  Detect Object Using Deep Learning  tool in ArcGIS Pro. Results are in the below map.

The initial results as expected where not relevant at all and most of the tents footprints were missing. Therefore the next step was to go through a retraining workflow.

In the below section we first digitized a training dataset that consists of approximately 3800 tents. These are show in the map below.

Once the digitization process was completed. We used the  Export Training Data for Deep Learning  tool in ArcGIS Pro in order to have the training data that the model will be retrained with. Understanding the parameters in the tool is a very important step in getting the best results possible when retraining a deep learning model. Below are some key parameters:

• Class Value Field: This should abide the original class value schema of the initial model. For the sake of this workflow we only had one class which is the building/tent footprint
• Image Format: In the initial pre-trained model the format of the training data was a TIFF. Therefore we used TIFF
• Tile Size X & Tile Size Y: This is the size of each image chip that the model will train on. In the initial model emd file the chip sized where 512 pixels so the retraining happened on the same chip size
• Stride X & Stride Y: This follow the same logic as the Tile Size. Therefore we used a stride size of 128 pixels
• Metadata Format: This is another important parameter. In the initial model metadata file you will see that the training happened with RCNN backbone model. Therefore, the retraining dataset should have the same format therefore we used RCNN Masks

Train Deep Learning Model

After finishing from exporting the training data. The next step was to retrain the Deep Learning Model. For this we used  Train Deep Learning Model  tool in ArcGIS Pro. There's also some important pre-requisites and parameters to take into consideration through this step. Below are the most essential:

• GPU: A GPU powered instance. We used an Azure cloud machine with a 16 gb Tesla V100 GPU. Note that you should specify GPU in the environments of the tool
• Deep learning framework:  Install deep learning framework  is an essential step to make sure that all the python dependent libraries are installed on your machine
• Max Epochs: This parameter will set the number of passes through the training dataset. We set the number to 100 which means the tool will pass 100 times over the training dataset
• Batch size: Batch size is dictated by the GPU of the machine you are using. With a Tesla V100 GPU a batch size of 16 will run smoothly without forcing the GPU to run out of memory. Batch size represent the count of images processed at a time.
• Pre-trained Model: This is a very important parameter since training a detection model from scratch will take longer training time. We used the living atlas referenced dlpk earlier on.
• Stope when model stops improving: This is also a very important parameter to make sure that we don't overfit our training model

Note that the process of retraining on our training dataset took approximately 20.5 hours.

Refugee Tents Footprints Extraction - Cox's Bazar is the final dlpk we produced in the previous step. This was the model used to detect the refugee tents footprints. The model accuracy is 84.2% and it's available publicly for downloading and testing.

In the following ArcGIS Dashboard you will notice 4 different tabs. Each tab workflow is explained below:

• Pre Processed Refugee Tents and Structures: The map shows the raw output of the previously trained deep learning model. This is the base layer used to derive all the following analysis. Note that 7894 tents and structures were detected.
• Post Processed Refugee Tents and Structures: In this section we ran  Regularize Building Footprint  tool in ArcGIS Pro to regularize the shapes of the buildings. Then we removed all footprints that are less than 2 square meters and finally ran through a visual inspection to make sure that no anomalies exist in the dataset.
• Tent Use: To derive the Tents & Structures use two workflows where followed. The first was all structures less than a 10 m are naturally small structures that exist next to bigger tents. These where highlighted then labeled as sanitary facilities. The next step was using  Spatial Join  tool in ArcGIS Pro to join the detected footprints with  Camps Facilities  layer from IOM. This resulted with a more detailed view of the detected footprints use. Finally those footprints that ended up without a use where classified as refugee tents. These where certainly the highest in number is the area being studied is a refugee camp. The number of tents detected and classified under refugee tents is approximately 5.4K. This can be examined in the 3rd tab of the dashboard.
• Population: In the  Rohingya Refugee Emergency at a Glance  storymap shared by UNHCR an average of 1 person living per 8 square meter is reported. Moreover, the latest  refugee population data  reported on the Humanitarian Data Exchange reports around 32000 people living in our studied Area of Interest. Taking this information along with the previously classified refugee tents we were able to estimate the population of refugees living in each tent according to the tent area. Search and Update cursors where used to pythonically derive these numbers. The details of the population is available in the Population tab of the below ArcGIS Dashboard.

Data: OSM Bangladesh

Workflow:

Convert OSM data to point, line, polygon

Create Attribute tags for linear and point features

Generate Junctions

Create Network Feature Dataset

• OSM linear pathway feature
• Junction point

Establish Travel parameters (Distance and Walktime)

• Distance (meters) was established with the shape attribute in the linear pathway file
• Walktime was established as length over time and we used 83 meters/min for calculting walk time. This was derived from the value of 5 Km/hr as described in the  Sphere Minimum Standards and Indicators for Humanitarian Response 

Edit Connectivity to establish a seamless network of linear features in the dataset

Test network dataset on sample points to assure accuracy

Generate Washroom Service Areas (2.5 minute walktime)

Model Setup Environment

Data Processing

1. Parse Unique ID from Service Area polygon name to join the attribute values in the summarize step back to the washroom source data
2. Summarize refugee tents (count and population) and other washrooms (count) with each 2.5 minute catchment
3. Use Spatial Join to summarize the number of wash room locations one refugee tent has access to.
4. Join summarized data back to original washroom and refugee datasets

Closest Facility Analysis

Determining the nearest wash room facility to each refugee tent. We look at the amount of tents that have a nearest wash room within 1 minute and outside of the 1 minute walk time. Some tents were not connected to any complete linear feature across the network due to the study area perimeter. Those tents have been omitted from the Closest Facility Analysis. There were 85 refugee tents that have been excluded. By the numbers this equates to a total population of 381 individuals that have been removed.

In order to augment and validate the network-based analysis, we performed the coverage analysis by incorporating Tobler's Hiking Function based on the slope of the terrain. The analysis was used the Distance Accumulation to generate a walk-time surface based on a 2.5 minute cut-off time.