YonSci / UNECA-Deep-Learning-for-Socioeconomic-Indicator-Prediction

UNECA-Deep-Learning-for-Socioeconomic-Indicator-Prediction

Unveiling the Potential of Transfer Learning in Machine Learning and Deep Learning Modes for Socioeconomic Indicators Prediction (Consumption Expenditure) in Malawi, 2016

Incorporating Survey Information into Satellite Imagery

Introduction

This project is dedicated to revolutionizing the way we predict socioeconomic indicators, specifically focusing on consumption expenditure in Malawi using alternative data sources (big data) and open-source frameworks. Although traditional household surveys are vital for decision-making, they suffer from resource-intensive and time-consuming drawbacks. These challenges, coupled with data gaps, reproducibility issues, and knowledge transfer, hinder our ability to effectively monitor UN Sustainable Development Goals. Our project uniquely integrates World Bank LSMS survey Microdata and satellite imagery captured during both nighttime and daytime. By employing advanced machine learning and deep learning techniques through transfer learning approach, we aim to forecast socioeconomic indicators more accurately.

Our objectives include filling in temporal and spatial data gaps, refining the existing methodology with an open-source framework, and enhancing overall reproducibility and knowledge transfer. In alignment with UN SDGs, particularly SDG 1 (No Poverty), SDG 2 (Zero Hunger: providing insights into food consumption and hunger), and SDG 9 (Industry, Innovation, and Infrastructure) by embracing innovation in the field of machine learning to enhance understanding and prediction of socioeconomic indicators. The project aspires to deliver a cost-effective, accurate, and reproducible AI system tailored for monitoring socioeconomic indicators.

In this project, we showcased consumption expenditure prediction for Malawi for the year 2016. The new approach accommodates off-year predictions, operates as a near real-time monitoring system, and demonstrates scalability for diverse regions. The proposed project approach extends beyond consumption prediction covering other indicators like poverty levels, wealth distribution, population distribution, and electricity accessibility.

Problem❓

1. The lack of reliable and continuous socioeconomic data in developing countries is a major obstacle to monitoring and evaluating sustainable development goals and making informed policy decisions. Obtaining frequent and reliable national-level statistics through surveys is both costly and labor-intensive, which poses a significant challenge for governmental and non-governmental organizations.

2. In recent years, there has been a growing interest in utilizing machine learning and deep learning techniques to estimate socioeconomic indicators. However, one of the main challenges associated with these approaches is the lack of reproducibility and documentation, making it difficult to verify and implement them effectively. This makes it difficult to replicate the results and adapt them to different contexts. To address this issue, it is crucial to focus on enhancing the reproducibility of machine learning projects.

By addressing these challenges, organizations and researchers can develop more accurate, reliable, and reproducible machine-learning models for estimating socioeconomic indicators in developing countries. These models can provide valuable insights into the socioeconomic situation in these countries, helping policymakers and organizations make informed decisions and plan effective interventions.

Solution 💡

1. In response to the first challenge, we present a cost-effective, accurate, and scalable method for predicting socioeconomic indicators. In this project, a novel machine learning/deep learning approach is implemented to predict socioeconomic indicators from survey data, and publicly available nighttime and high-resolution daytime satellite imagery. The methodology was originally established by researchers at Stanford University. and is currently being adopted by several institutions around the world.

2. In response to the second challenge, it is crucial to focus on enhancing the reproducibility of machine learning projects. This can be achieved by ensuring that the code, data, and environment used in the development of the models are well-documented and easily reproducible. By doing so, researchers can not only verify the accuracy and reliability of the models but also adapt them to different contexts and situations, thereby increasing their practical utility and impact. We demonstrated a reproducible approach that provided a step-by-step guideline for data collection, preprocessing, and code implementation.

3. Furthermore, we address challenges related to nighttime satellite imagery retrieval by deploying GEE Javascript and resolving the time mismatches with survey data. Additionally, we tackle issues linked to Planet satellite imagery by updating the Planet downloader scripts to accommodate changes in the Planet Data API parameters, ensuring accurate data retrieval. Furthermore, the implementation of advanced deep-learning models allows us to assess the performance of alternative models, enhancing the robustness of the methodology.

Objective 🎯

The primary aim of this project is to integrate survey data, nighttime satellite imagery, and daytime high-resolution satellite imagery, leveraging machine learning/deep Learning methodologies to forecast socioeconomic indicators using a reproducible framework utilizing open-source tools and resources (i.e. Consumption expenditure for Malawi for the year 2016).​

Specific Objectives

• Fill both temporal and spatial socioeconomic data gaps by inferring from remote sensing data. ​

• Enhance the existing methodology using only open-source tools and publicly available data/resources.​

• Improve the reproducibility and documentation of the existing methodology mainly key stages such as data collection, code implementation, and data analysis for easy knowledge/skill transfer.

Application 💻

• This methodology holds broad applicability, extending beyond consumption expenditure prediction, it can be applied to wealth, poverty, income, and population prediction.
• It facilitates predictions during "off-years" when surveys are not conducted.
• It also enables near real-time monitoring serving as an early-warning system.
• This kind of approach is scalable, the trained model, from one location, can be applied to new regions with similar characteristics.
• The proposed method contributes to the production of frequent and continuous statistical reports on socioeconomic indicators, complementing existing methods used in the National Statistics Offices (NSOs).

Implementation Frameworks and Environment 🖥️

We used Google Colab Pro+ for computing with high-performance GPUs—specifically A100 and V100 with 51.0GB of RAM.

The project utilizes the following framework & tools: ​

1. PyTorch framework using Python programming language: PyTorch is a popular deep learning framework that supports Python, making it easier for developers to work with machine learning models and data processing tasks.

2. Google Earth Engine (GEE) with JavaScript codes: GEE is a cloud-based geospatial data processing platform that allows users to visualize, analyze, and model Earth science data. JavaScript codes can be used to create custom GEE applications and widgets for interactive mapping and data visualization.

3. QGIS: QGIS is a free and open-source geographic information system (GIS) that allows users to create, analyze, manage, and visualize spatial or geographic data. QGIS can be used for various applications, including mapping, spatial analysis, and visualization.

High-level Workflow

​

Procedure 📋

The implementation of this approach includes important steps such as:

1) Survey data collection, and processing

The project used publicly available survey data from the World Bank Living Standards Measurement Study (LSMS) Microdata Library, particularly the Fourth Integrated Household Survey (IHS4), gathered through the National Statistical Office (NSO) of Malawi, during the period spanning from April 2016 to April 2017.

Once, the Survey Data is retrieved and undergoes a data cleaning process to handle missing values. Subsequently, the consumption values were standardized using Purchasing Power Parity (PPP) for Malawi in 2016, and the standardized consumption and Geovariable datasets were merged using their unique ID. The data is then grouped by the enumeration area using coordinates (latitude & longitude), and the daily consumption per person is computed ($/person/day). Following this, statistical information is summarized, and the results are verified with external World Bank benchmarks. Finally, the processed data is used to generate consumption maps, providing a visual representation of the predicted socioeconomic indicators.

For a detailed implementation procedure, go through: Survey_Data_Preprocessing_Malawi_2016.ipynb

Steps:

Daily Consumption per Capita for Malawi for 2016

Basic summary statistics​ of consumption per capita per day

Metric	Value
Count	780
Min	0.724403
Max	80.036069
Median	1.853950
Mean	2.428631
STD	3.262697

2) Nightlight satellite imagery acquisition, and processing

The project also utilized nighttime satellite imagery sourced from the NOAA National Center for Environmental Information and downloaded using the Google Earth Engine (GEE) Javascript code editor. Here you can find the Javascript code.

In the script, we filtered the nighttime satellite imagery using the area of interest (AOI), temporal duration, and the relevant spectral bands required for the analysis. Then we computed the annual composite of the filtered images. The resulting nightlight image was exported as a GeoTIFF file to Google Drive. You can find the annual composite nighttime satellite imagery for the year 2016 here. Subsequently, calculations were performed to create a 10kmx10km box around the central latitude and longitude to retrieve the nightlight values for each cluster. The procedure also involved the computation of summary statistics for the nightlight values and calculating their correlation with the consumption value. The processed data is utilized to create maps depicting nightlight values.

The process also included computing summary statistics of the nightlight values and calculating their correlation with the consumption value. Finally, the processed data is used to generate nightlight value maps.

For a detailed implementation procedure, go through Processing_Nighttime_Satellite_Imagery.ipynb

Steps:

Description of Nightlight satellite imagery

Nightlight satellite imagery

Cluster Boxes

Map of Nightlight Values and Consumption for Malawi for 2016

Correlation between Nightlight values and Consumption

3) Generate download locations for Daytime satellite imagery

We created download locations for daytime satellite images using a combination of systematic and stochastic sampling methods within cluster bounding boxes. Specifically, we generated 50 download locations per cluster, forming a grid of 49 uniformly spaced points (7x7) within the bounding box and adding 1 point through random sampling within the same box. This approach ensures diverse download locations for daytime satellite images. Subsequently, for each set of 50 points, we compiled the image name, image latitude, and image longitude, appending them to the data frame. A total of 39,000 image download locations have been generated.

For a detailed implementation procedure, go through Generate_Image_Download_Locations.ipynb

Steps:

Download Location Dataframe

Download Location Map

4) Undersamping to avoid nightlight values class bias​

Within our dataset, the number of nightlight values having zero is notably higher, potentially causing an imbalance in the data distribution. The objective is to address this by undersampling or reducing instances from areas with zero or minimal nightlight data, aiming to mitigate class imbalance. This approach introduces diversity into the model by selectively removing rows associated with zero nightlights until the target fraction is achieved. A total of 33,900 image download locations have been left after performing the under-sampling.

For a detailed implementation of the undersampling, go through Generate_Image_Download_Locations.ipynb

Steps:

Download Location Undersamped Dataframe

5) Daytime satellite imagery acquisition and processing

The acquisition of high-resolution daytime satellite imagery is performed using the Planet API, which provides images specifically for research and academic purposes. The Planet Scope (PSScene) images have a spatial resolution ranging from 3.7 to 4.1 meters, later resampled to 3 meters for practical use. The process of obtaining Planet Imagery encompasses a series of steps. Initially, we set up the API Key in Planet Explorer. Following this, we apply essential filters such as geometry, date, and cloud filters to download the images properly. The download locations (image latitude and longitude) derived from previous steps serve as inputs for image retrieval, incorporating additional parameters like a zoom level of 14 and a maximum cloud filter of 0.05 (5%). The image acquisition spans the period from 2016 to 2017, culminating in a total of 33,900 downloaded images.

Steps:

For downloading the Planet daytime satellite imagery, go through Download_satellite_images_Planet.ipynb

Sample Planet Images

6) Create nightlight bins and label the daytime satellite imagery

The Gaussian Mixture Model is used to establish nightlight bins/labels to cluster the daytime satellite imagery into three categories based on nighttime values. The GMM-predicted cutoff values of 0.020 and 0.376 delineate a low nightlight bin, a medium nightlight bin, and a high nightlight bin.

For a detailed implementation of the undersampling, go through Nightlights_bins.ipynb

Steps:

Labeled Dataframre

Labeled Cluster Map

Labeled Satellite Images

7) Preparation of training and validation datasets

The preparation of the training and validation dataset involves employing a stratified train-validation split method, ensuring that each cluster group has a random assignment of samples to the train-validation set. This approach mitigates potential sampling issues, preventing situations where certain clusters lack training-validation data and ensuring a consistent sampling distribution. Specifically, an 80-20 split is implemented, with

• 80% of the data allocated for training
• 20% for validation

For a detailed preparation of training and validation datasets, go through Prepare_training_validation_datasets.ipynb

Steps:

Training and Validation Sets

8) Train variants of Convolutional Neural Network (CNN) models using a transfer learning approach

This project uses a novel deep-learning approach through a transfer learning method to predict consumption. Transfer learning is a technique that involves using a pre-trained model as a starting point for a new task. The pre-trained model has already learned to recognize many different features and can be used as a starting point for training a new model on a related task. It involves leveraging knowledge gained from one task to be repurposed for a different but related task. In this particular case, we are using nighttime light as a proxy for socioeconomic indicators.

Using pre-trained models (fine-tuning):

• It reduces computation costs
• It reduces computation time (high accuracy with few iterations)
• It requires less amount dataset(less data)
• It reduces carbon footprint (low environmental costs)
• It avoids training ML models from scratch

Our objectives are:

1. Predict and assign the nightlight bin labels (probability class) of the satellite images (image classification using transfer learning).
2. Simultaneously learn and extract features that are useful for consumption prediction (feature vector extraction).

The training process involves a series of steps using a Deep Convolutional Neural Network namely variants of Virtual Geometry Group (VGG) models such as VGG-11, VGG-16, and VGG-19 as a Transfer Learning framework.

General Architecture of VGG Model

These models are renowned for their capabilities in feature extraction and classification. These models have been widely used in various applications, including image classification, object recognition, and image segmentation. These models were initially trained on the ImageNet dataset that contains over 1.2 million images distributed across 1,000 classes. In the default configuration, the final network can classify images into 1000 object categories. The VGG architecture was introduced by Simonyan and Zisserman in 2014 from Oxford University, Very Deep Convolutional Networks for Large Scale Image Recognition. This model achieved a 92.7% top-5 test accuracy using the ImageNet dataset.

Model	Convolutional Layers	Fully Connected Layers	Parameters (approx.)	Pooling Layers	Input Size	Activation Function	Pre-training	Kernel Size	Stride	Padding
VGG-11	8	3	132 million	Max Pooling	224x224x3	ReLU	ImageNet	3 X 3	1	1
VGG-16	13	3	138 million	Max Pooling	224x224x3	ReLU	ImageNet	3 X 3	1	1
VGG-19	16	3	144 million	Max Pooling	224x224x3	ReLU	ImageNet	3 X 3	1	1

Steps:

1. Locate Image Data Directory: Identify the directory containing image data (data_dir) previously created for the task.

2. Download Pre-trained Models (VGG Models): Obtain pre-trained models, particularly VGG models, needed for the task. Link: https://pytorch.org/vision/main/models/vgg.html

3. Set and Initialize Pre-trained Model: Configure and initialize the pre-trained model with its initial parameters.

Parameter	Value
Model names	vgg11_bn, vgg16_bn, vgg19_bn
Image classes	3 (low, medium, high)
Batch size	8
Number of epochs	30
Feature extracting flag	True
Input size	224x224x3 (H X W X C)

4. Apply Image Transformation/Augmentation: To enhance the variety of training data (image).

Transformation Step	Description
Image Flipping	Flips the image horizontally.
Image Cropping/Resizing	Resizes/crops the input image to the specified size.
Image Normalization	Normalizes the input image by subtracting the mean value and dividing by the standard deviation [0.485, 0.456, 0.406], [0.229, 0.224, 0.225].
Conversion to PyTorch Tensors	Converts images from PIL (Python Imaging Library) format to a PyTorch tensor format.
Rearrange Image Dimensions	Dimensions change from HxWxC to CxHxW (channels first).

5. Create PyTorch Image Dataset: Constructing PyTorch datasets serves to streamline image loading, facilitate efficient memory usage, and seamlessly integrate the image transformation module. Additionally, it automatically assigns labels to images based on the subdirectory structure.

6. Create PyTorch Dataloader: Set up a PyTorch data loader to divide the image dataset into batches for efficient training (Batching), randomize the order of data to prevent model bias (Shuffling), and accelerate data loading by fetching batches in parallel (Parallelizing Data Loading).

Parameter	Value
Batch Size	3
Shuffling of the Data	True
Number of Workers for Data Loading	4

7. Check CPU and GPU Availability: Verify the availability of both CPU and GPU resources. Send the model to the appropriate device based on availability. For this project, high-performance GPUs, namely A100, V100, and T4, were used, each equipped with 51.0GB of RAM.

8. Define Optimizer Function: In this project a Stochastic Gradient Descent (SGD), is used to update model parameters. The purpose of the optimizer function is to iteratively update the model's parameters in a way that minimizes the loss function.

Hyperparameter	Value
Momentum	0.1
Learning Rate	1e-4

9. Define Loss Function: In this project Categorical Cross Entropy (CCE) loss function is used to quantify the difference between predicted and actual values. It Calculates the average cross-entropy loss between the predicted and true class labels. It is commonly used for multi-class classification problems.

10. Train the model: The model training involves executing all previously specified configurations, including initial parameters, image transformation, image datasets, data loaders, designated GPUs, optimizer function, and loss function.

11. Save the Model: The trained models are stored for future uses or deployment.

Models: The trained models can be found in this Google Drive directory:

• VGG11 Trained Model
• VGG16 Trained Model
• VGG19 Trained Model

12. Evaluate the Model: The table below displays the loss during both training and validation, including the accuracy of the model's performance.

Model	Train Loss	Valid Loss	Accuracy
VGG11	0.6014	0.5321	0.7718
VGG16	0.5744	0.5284	0.7705
VGG19	0.5805	0.5175	0.7849

The above results have indicated a good performance of the trained model in terms of correctly identifying each satellite image to its respective nightlight bin category/classes. Overall, the VGG models have been successfully trained using transfer learning to classify the satellite images to their respective nightlight bin classes.

Learning Curves

Learning curves represent the graphical depiction of a model's learning performance as a function of time. Widely employed as a diagnostic tool in machine learning, these curves are particularly useful for algorithms that progressively learn from a training dataset.

Reviewing the learning curves of models during training can be used to diagnose problems with learning, such as an underfit or overfit model, as well as whether the training and validation datasets are suitably representative.

Plot for Training and Validation Loss

Plot for Training and Validation Accuracy

9) Feature Extraction and Aggregation

The feature vectors provide a lot of information about evidence of economic activity or lack of economic activity from satellite images. Feature vectors are a numerical representation of an object in an image. These features detected by the model include objects, edges, textures, and other patterns. In particular, urban areas, nonurban areas, roads, water bodies, agricultural areas, etc.

For feature vector extraction each image passes through the pre-trained VGG model and the final dense layer is used to extract the feature vector from each image in the clustur with the output feature vector size of 4096. Finally, the feature vectors of all images in the cluster are averaged to obtain a single feature vector per cluster. The cluster feature vector and cluster order files can be found here

For a detailed implementation of feature extraction, go through Feature_extraction_aggregation.ipynb

Steps:

Visualizing Feature Maps (Activation Maps)

Visualizing feature maps, also known as activation maps, provides valuable insights into how neural networks interpret and understand input data. Feature maps are representations of learned patterns and structures within the input data at different levels of abstraction.

Sample feature maps from low nightlight bin

Sample feature maps from medium nightlight bin

Sample feature maps from high nightlight bin

10) Building prediction model using Ridge regression model

In this project, a Ridge Regression model is employed to forecast consumption levels utilizing the feature vector extracted from the previous step. Ridge regression is a form of a linear regression model with L2 regularization that prevents overfitting. A Ridge Regression model is a supervised learning algorithm, that predicts a target variable based on one or more predictor features.

The feature vector, computed for each cluster, serves as the input variable, while the consumption level for each cluster is used as the output variable. Standardization or scaling is initially applied to both the input and output variables. Subsequently, a randomized cross-validation technique is employed with a 10-fold cross-validation, to predict consumption levels and evaluate model performance using a weighted R-square.

Steps:

The ridge regression model has a cross-validated mean R-squared value of 0.5. The R-squared value indicates that the model explains 50% of the variance in the data. The mean absolute error (MAE) of 0.72 suggests that, on average, the model's predictions are off by 0.72 units from the actual values. The root mean squared error (RMSE) of 3.1 indicates that the model's predictions are off by 3.1 units on average.

Accuracy Metrics

Metric	Predicted consumption
Cross-validated mean R-squared	0.5
Mean Absolute Error (MAE)	0.72
Root Mean Squared Error (RMSE)	3.1

Cross-validated mean R-squared

For a detailed implementation of the Ridge regression model, go through Predict_consumption_ridge_regression_model.ipynb

Finally, the forecasted consumption levels were visually represented on a map.

Predicted consumption (without transformation)

Predicted consumption (with log transformation)

Map of actual and estimated per capita consumption expenditure

Conclusions/ Wayforward

Scalability and Versatility: The proposed methodology extends its applicability beyond the prediction of consumption expenditure, it can be used to predict a wide range of socioeconomic indicators, including poverty levels, wealth distribution, population density, and access to electricity.

Ease of Adoption and Adaptability:

• The trained CNN/VGG models and extracted features can be used by organizations, NSOs, and researchers to make predictions without the necessity to train the model from scratch and avoid the need for downloading and processing satellite imagery.

• Models developed for specific areas can be seamlessly transferred and applied to new regions with similar characteristics. Several previous studies indicated acceptable model performance in prediction.

Enhanced Reporting Capability: This approach facilitates the generation of comprehensive quarterly and annual reports, addressing both temporal and spatial data gaps. It complements existing official statistical methods, contributing to more detailed socioeconomic reports.

Decision Support Tools: Country-specific dashboards can be built to visualize the spatial and temporal patterns of the model predictions. This feature supports evidence-based planning and interventions, providing valuable decision support for countries and NSOs.

Exploration of Advanced Algorithms: Testing advanced deep learning algorithms, including ResNet, RegNet, EfficientNet, Inception V3, and AlexNet, etc holds the potential to further improve prediction accuracy and model performance.

Hyperparameter Optimization: Fine-tuning the model through optimization procedures, such as identifying optimal batch sizes and epochs, can enhance the model's overall performance, ensuring optimal model configuration.

Integration of High-Resolution Images: In this project, the Planet's high-resolution satellite images are used, incorporating other free open-source satellite images may improve or provide new insight into socioeconomic landscapes.

Scripts 📜

The Script folder contains the following files:

1. Survey_Data_Preprocessing_Malawi_2016.ipynb
2. GEE_Nightlight_2016_2017.js
3. Processing_Nighttime_Satellite_Imagery.ipynb
4. Generate_Image_Download_Locations.ipynb
5. Nightlights_bins.ipynb
6. Download_satellite_images_Planet.ipynb
7. Prepare_training_validation_datasets.ipynb
8. Image_labeling.ipynb
9. Train_VGG11_model_145.ipynb
10. Feature_extraction_aggregation.ipynb
11. Predict_consumption_ridge_regression_model.ipynb
12. Visualize_feature_maps.ipynb

Processed/Intermediate output 🗂️

The data directory contains the processed output generated by the scripts:

1. df_clusters_malawi_2016.csv
2. df_clusters_malawi_nl.csv
3. df_malawi_download_loc.csv
4. df_malawi_loc_labed.csv
5. df_malawi_loc_labed_2016.csv
6. image_download_actual_malawi2016.csv
7. Extracted_feature_index_malawi_2016_VGG145.csv
8. predicted_malawi_2016_VGG145.csv

Folder Structure of the project

 Main Project Folder/   
├── Images                           # Images used in the project and some output images   
├── Model_Output                     # Model output files such as feature vectors
├── Nighttime_Satellite_Imagery      # It contains the Nighttime Satellite Imagery used in this project
├── Processed_Files                  # Intermediate files from the analysis  
├── Scripts                          # Scripts used in the project mainly Jupyter notebooks and javascript codes
├── Survey Data                      # LSMS Survey Data: Fourth Integrated Household Survey data (Consumption Aggregate & Household Geovariables file)
├── README.md                        # Overview of the project  
├── License.txt                      # License for the project

Issues

On the issues page, the primary bugs in the code are documented along with their solutions. Additionally, you can create new issues if you encounter any new bugs.

Packages required

Python Packages

Library	Description
sys	System-specific parameters and functions
os	Operating system interfaces
Numpy	Numerical operations and arrays
Pandas	Data manipulation and analysis
Matplotlib	Plotting library
Seaborn	Statistical data visualization
Plotly	Interactive plots and dashboards
Math	Mathematical functions
Random	Random number generation
Geoio	Geospatial data I/O and processing
IPython	Interactive computing in Python
Rasterio	Geospatial raster data I/O and processing
Utils	Utility functions and tools
torch	PyTorch deep learning library
torchvision	PyTorch computer vision library
time	Time-related functions
copy	Shallow and deep copy operations

Contact

1. Yonas Mersha, Data Science Consultant | African Centre for Statistics (ACS) | United Nations Economic Commission for Africa (UNECA)

2. Issoufou Seidou Sanda, Principal Statistician | African Centre for Statistics (ACS) | United Nations Economic Commission for Africa (UNECA)

3. ANJANA DUBE, Senior Regional Advisor | African Centre for Statistics (ACS) | United Nations Economic Commission for Africa (UNECA)

Acknowledgments

1. This research project has been supported by the Regular Program of Technical Cooperation (RPTC).

2. Special thanks to Jatin Mathur for sharing the create_space, add_nightlights, generate_download_locations, drop_zeros functions, and other code snippets on GitHub. We have integrated certain code snippets, tailoring them to suit the specific requirements of our project.

References

World Bank's Living Standards Measurement Study (LSMS) survey data

National Statistical Office. Malawi - Fourth Integrated Household Survey 2016-2017, Ref. MWI_2016_IHS-IV_v02_M. Dataset downloaded from World Bank Microdata Library: https://microdata.worldbank.org/index.php/catalog/2936/related-materials.

VIIRS nighttime satellite imagery

Elvidge, C.D, Zhizhin, M., Ghosh T., Hsu FC, Taneja J. Annual time series of global VIIRS nighttime lights derived from monthly averages:2012 to 2019. Remote Sensing 2021, 13(5), p.922, doi:10.3390/rs13050922 doi:10.3390/rs13050922

C.D. Elvidge, K. Baugh, M. Zhizhin, F. C. Hsu, and T. Ghosh, “VIIRS night-time lights,” International Journal of Remote Sensing, vol. 38, pp. 5860–5879, 2017.

Planet Satellite Imagery

Planet Labs PBC. (2023). Planet Application Program Interface: In Space for Life on Earth. Retrieved from https://api.planet.com

Related Works

Neal Jean, Marshall Burke, Michael Xie, W. Matt Davis, David Lobell, and Stefano Ermon. 2016. "Combining satellite imagery and machine learning to predict poverty." Science 353, 6301. Combining satellite imagery and machine learning to predict poverty. GitHub

Mapping-Poverty-With-Satellite-Images

Jatin Mathur, Predicting Poverty Replication

COMBINING SATELLITE IMAGERY AND MACHINE LEARNING TO PREDICT POVERTY

Satellite images can map poverty