Transfer Learning from Deep Features for Remote Sensing and Poverty Mapping
Authors: Michael Xie, Neal Jean, Marshall Burke, David Lobell, Stefano Ermon
AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that these learned features are highly informative for poverty mapping, even approaching the predictive performance of survey data collected in the field. Our transfer model significantly outperforms every model except the survey model in every measure except recall. |
| Researcher Affiliation | Academia | Michael Xie, Neal Jean, Marshall Burke, David Lobell, and Stefano Ermon Department of Computer Science, Stanford University {xie, nealjean, ermon}@cs.stanford.edu Department of Earth System Science, Stanford University {mburke,dlobell}@stanford.edu |
| Pseudocode | No | The paper describes the model architecture and training process in text but does not include any explicit pseudocode blocks or algorithm listings. |
| Open Source Code | No | The paper does not provide any explicit statements about releasing source code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | The aggregate dataset D2 consists of over 330,000 images, each labeled with an integer nighttime light intensity value ranging from 0 to 631. We further subsample and bin the data using a Gaussian mixture model, as detailed in the companion technical report (Xie et al. 2015). |
| Dataset Splits | Yes | The random cropping model achieved a validation accuracy of 70.04% after 400,200 SGD iterations. In comparison, the fully convolutional model achieved 71.58% validation accuracy after only 223,500 iterations. All models are trained using a logistic regression classifier with L1 regularization using a nested 10-fold cross validation (CV) scheme, where the inner CV is used to tune a new regularization parameter for each outer CV iteration. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., specific GPU or CPU models). It only mentions 'Both models were trained in roughly three days'. |
| Software Dependencies | No | The paper mentions that 'all networks are trained with Caffe (Jia et al. 2014)', but it does not provide a specific version number for Caffe or any other software dependencies. |
| Experiment Setup | Yes | Both CNN models are trained using minibatched gradient descent with momentum. Random mirroring is used for data augmentation, along with 50% dropout on convolutional layers replacing fully connected layers. The learning rate begins at 1e-6, a hundredth of the ending learning rate of the VGG model. All other hyperparameters are the same as in the VGG model as described in (Chatfield et al. 2014). |