Data Aspects

• I use Sentinel 2 data of train coordinates both locations ( Fergana , Orengurg ).
• also Cloud Probability from sentinel 2 used.
• Timeseries span about 4 years.
• GEE used collection with CLOUD_COVER_PERCENT 20 ( similar as test data ).

Feature Engineering

Sentinel 2 bands used B2 to B12 ( simiar as test data ).

Added Normalized Difference indexes such as ( ex: BNDVI, MNDVI, NDVI, NDVI705, RENDVI, RNDVI, GNDVI, WI2, WI1, VrNIRBI, VIG, VI700, etc..).

Also Transform 4 year timeseries to aggregated of each month.

that data structure like ( sample_size, months, features )

Model Architecture

Since data structure is 2D, it best to use conv layer, so an Simple stack of convolution with dropout layer

Insight

• Very few samples in train coordinate mislabel note by [@ZaakciiRu](https://zindi.africa/competitions/geoai-challenge-for-cropland-mapping-in-dry-environments/discussions/28437) but I didn't manualy remove any sample. so to counterbalance loss function with label_smoothing 0.1 value.
• very heavy dropouts used during training.