This is the first place solution of team WeMoveMountains in the NASA Harvest Field Boundary Detection Challenge. This solution is a single 10-fold modified Regnetv-Unet developed in Pytorch.
MLHub model id: model_nasa_rwanda_field_boundary_competition_gold_v1. Browse on Radiant MLHub.
The dataset description is available on RadiantMLHub. See here.
Muhamed T., Azer K. (2023) “Spatio-Temporal Attention-based Unet for Field Boundary Detection”, Version 1.0, Radiant MLHub. [Date Accessed] Radiant MLHub. https://doi.org/10.34911/rdnt.h28fju
This solution was developped by:
- Supervised Learning
- Segmentation
Our solution is a modified Unet++ with an attention mechanism between every interconnection of the encoder to the decoder (Meaning after each output layer of the encoder model). The encoder is a Regnet (more specifically regnetv_040 available in the Timm library).
The training was done on a Linux system with an Nvidia GPU (A100 80GB). An A600 24GB should be enough to train the model.
Review the GitHub repository README to get started running this model for new inferencing.
- Augmentation
The only augmentation we did was static. It is done before the training and saved into a new folder. We noticed that the model learned better with little (only Flip augmentation) to no augmentation during the training.
- Training procedure
For each tile, all the time-series images are loading as Timestamps x C x H x W and passed to the model. The input now is Batch_size x TimeStamps x C x H x W. It is reshaped to Batch_size*TimeStamps x C x H x W before being fed into the encoder, and then reshaped to Batch_size x TimeStamps x D x H' x W' for the attention pooling mechanism. Finally, the input to the decoder becomes Batch_size x D" x H" x W".
The output file is a csv named output.csv and should be available in the data/output folder. Each row of the csv file correspond to a pixel of the flattened 256x256 image.