Efficient Poverty Mapping from High Resolution Remote Sensing Images
Authors: Kumar Ayush, Burak Uzkent, Kumar Tanmay, Marshall Burke, David Lobell, Stefano Ermon45280
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our approach exceeds previous performance benchmarks on this task while using 80% fewer high-resolution images, and could be useful in many domains that require high-resolution imagery. We apply this approach to the task of poverty prediction in Uganda. In our study country of Uganda, we show how our approach can reduce the number of high-resolution images needed by 80%. |
| Researcher Affiliation | Academia | 1 Department of Computer Science, Stanford University 2 Department of Earth Science, Stanford University 3 Department of Electrical Engineering, IIT Kharagpur |
| Pseudocode | Yes | See appendix for the pseudocode and implementation details. |
| Open Source Code | No | The paper mentions 'See appendix for the pseudocode and implementation details' but does not explicitly state that source code for the described methodology is being released or provide a link to a repository. |
| Open Datasets | Yes | Our ground truth dataset consists of data on consumption expenditure (poverty) from Living Standards Measurement Study (LSMS) survey conducted in Uganda by the Uganda Bureau of Statistics between 2011 and 2012 (UBOS 2012). We acquire both high-resolution and low-resolution satellite imagery for Uganda... These images come from Sentinel-2 with 3 bands (RGB) and 10m pixel resolution and are freely available to the public. We use the same transfer learning strategy as in (Ayush et al. 2020) by training an object detector (YOLOv3 (Redmon and Farhadi 2018)) on x View (Lam et al. 2018), one of the largest and most diverse publicly available overhead imagery datasets. |
| Dataset Splits | No | The paper states 'We have N=320 clusters in the survey. We divide the dataset into a 80%-20% train-test split.' It does not specify a separate validation split or how validation was performed beyond general training context. |
| Hardware Specification | No | The paper mentions 'GPU availability' as a real-world constraint but does not specify any particular GPU models, CPU models, or other hardware specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions software like 'YOLOv3', 'Sentinel-2', and 'Gradient Boosting Decision Trees' but does not provide specific version numbers for any of these components. |
| Experiment Setup | Yes | As mentioned in the previous section, we divide each tile into S=4 subtiles of 1000 1000 pixels each (higher values of S led to unstable training with higher variance and less sparse selections). We intentionally change λ to quantify the effect on the policy network. As seen in Fig. 6, the policy network samples less HR tiles (a 0.09 fraction) when we increase λ to 2.0 and the r2 goes down to 0.48. On the other hand, when we set λ to 1.0, we get optimal results in terms of r2, while acquiring only a 0.18 fraction of HR imagery. We use a sigmoid function to transform logits to probabilistic values, sj,k i [0, 1]. We use the self-critical baseline (Rennie et al. 2017), A, to reduce the variance. Finally, in this study we use temperature scaling (Sutton and Barto 2018) to adjust exploration/exploitation tradeoff during optimization time. |