Overview
EuroSAT is a collection of RGB and multispectral images from the Copernicus Sentinel-2A satellite, capturing various land cover types across Europe. Each RGB image is accompanied by 10-band multispectral images and appropriate land use labels. An AI model capable of classifying these images would allow satellites to downlink only relevant images (e.g., urban areas), enabling faster data transfer with reduced volume. Given the hardware constraints on satellites, such a model needs to be highly optimized for power consumption and resource utilization.
Goals
The aim was to create an AI model that can successfully classify EuroSAT images, allowing for the downlinking of only relevant images for specific use cases. The model had to be deployable on-board the satellite, necessitating its implementation on a resource-constrained platform such as FPGA
Solution
Our solution involved creating a pipeline that handled the training of prototype models while enabling deployment on various resource-constrained platforms. By specifying the target FPGA, the deployment pipeline manages the necessary transformations to port the ML model to the target device. This process utilizes Quantization Aware Training (QAT) to streamline the model. Post-training, additional transformations are performed to convert the model into hardware-compatible layers suitable for FPGA deployment.
Results
- Easier deployment to supported FPGAs
- Faster prototyping and validation
- Successful classification of EuroSAT images with optimized model performance on resource-constrained platforms
- Efficient downlinking of relevant images, reducing data transfer volume and enhancing operational efficiency
