Implements AddCoords and CoordConv modules that augment convolutional layers with normalized spatial coordinates (x, y, optionally radius). This is a self-contained, reusable pattern for any vision task requiring spatial awareness, such as remote sensing or medical imaging. It can be extracted and applied to any CNN with minimal modification.

Provides a full ConvGRU implementation with configurable layers, kernel sizes, dropout, and batch normalization. The ConvGRUCell and ConvGRU modules are well-structured, support multi-layer stacking, and include proper weight initialization. This is a reusable spatiotemporal sequence model that can be extracted for any video or time-series prediction task.

Implements block and grid attention mechanisms as used in MaxViT. The PartitionAttention base class provides a generic framework for partitioning spatial dimensions, applying multi-head self-attention with relative position bias, FiLM conditioning, and MLP, then reversing the partition. BlockAttention and GridAttention subclasses implement specific partition strategies. This is a clean, reusable pattern for efficient window-based attention in vision transformers.

Provides haversine-based Gaussian weight computation and multi-satellite merging for geospatial data. The gaussian_merge_multi function takes a list of xarray DataArrays, latitude/longitude grids, and satellite subpoints, and produces a blended field using Gaussian center-weighted averaging. This is a reusable data fusion pattern for combining overlapping satellite observations, applicable to any multi-sensor geospatial dataset.

Implements a PointToGridTranslator that converts sparse weather station observations (lat/lon point measurements) into a dense 2D grid using Lambert Conformal Conic projection. It handles averaging of multiple stations per pixel, variable selection, and fill values. This is a reusable pattern for gridding any sparse point data (e.g., weather stations, sensor networks) onto a regular grid, with configurable projection and resolution.