The MLAB model library is soon to be released and contains base models for Earth system science deep learning models adapted, pretrained, and quantized for immediate use on Xilinx FPGAs through the Vitis AI framework to be executed on the Xilinx DPU102. The base models have been formulated and trained in tensorflow using NVIDIA GPUs, and quantization and deployment have been done through the Vitis AI stack deploying to a ZCU102 development board for integration testing.
The source code, original paper links, and a short documentation are available below.
| Name | Architecture | Documentation | Source | Source Publication |
|---|---|---|---|---|
| vgg16_vitis | VGG16 | Doc | Src | Paper |
| vgg19_vitis | VGG19 | Doc | Src | Paper |
| resnet34_vitis | ResNet34 | Doc | Src | Paper |
| resnet50_vitis | ResNet50 | Doc | Src | Paper |
| resnet101_vitis | ResNet101 | Doc | Src | Paper |
| resnet152_vitis | ResNet152 | Doc | Src | Paper |
| densenet121_vitis | DenseNet121 | Doc | Src | Paper |
| densenet161_vitis | DenseNet161 | Doc | Src | Paper |
| densenet161_vitis | DenseNet169 | Doc | Src | Paper |
| densenet201_vitis | DenseNet201 | Doc | Src | Paper |
| mobilenet_vitis | MobileNet | Doc | Src | Paper |
The development of the models and their testing has been implemented against the following widely accepted datasets.
| Name | Resolution | Channels | #Images | #Classes | Task* | Ref. |
|---|---|---|---|---|---|---|
| EuroSAT | 256x256 | Multispectral | 27,000 | 10 | C | Source, P1, P2 |
| AID | 600x600 | RGB | 10,00 | 30 | C | Source, P1 |
| UC-Merced | 256x256 | RGB | 2,100 | 21 | C | Source, P1 |
| Resisc45 | 256x256 | RGB | 31,500 | 45 | C | P1 |
| RSI-CB256 | 256x256 | RGB | 24,000 | 35 | C | Source, P1 |
* C - Classification, S - Segmentation, CD - Change Detection
** Multi-label
The following provides a more detailed description of the publicly available datasets employed in this project.
The EuroSAT dataset is composed of aerial image tiles showing varying land-use classes in RGB colors as well as with multispectral bands.
Key Features:
Source, Paper no. 1, Paper no. 2
The UC-Merced Land Use Dataset is composed of aerial image tiles showing varying land-use classes.
Key Features:
The AID (Aerial Scene Classification) Dataset is composed of aerial image tiles showing varying land-use classes.
Key Features:
The Resisc45 Dataset is composed of aerial image tiles showing varying land-use classes.
Website is currently down. - 19.7.2022
Key Features:
The RSI-CB256 Dataset is composed of aerial image tiles showing varying land-use classes.
Info: This dataset does also exist with the resolution 128 x 128 pixels (RSI-CB128).
Key Features: