AI4Copernicus Agriculture services are a collection of pre-processing harmonization tools and deep learning models for the crop classification task. They were developed to facilitate the development of the solutions proposed by the Open Call winners of AI4Copernicus. All the pipelines are based on python and available as dockerized applications.

The list of the services are:

• TimeSen2Crop: This dataset, available to download and described more in detail at the link TimeSen2Crop, is a pixel based dataset made up of more than 1 million samples of Sentinel 2 time series associated to 16 crop types. The dataset contains the spectral signature of the samples during an agronomic year (September 2017 - August 2018).

• Harmonization of pre-processed Time Series of Sentinel-2 data: This pipeline processes an agronomic year of Sentinel-2 acquisition to extract 12 monthly composites based on the median of the spectral signature of the samples considered, allowing the definition of a harmonized time series reprojected on a standardized temporal grid.

• Long Short-Term Memory Neural Network for Sentinel-2: This architecture can either be trained using samples selected by the user or be trained directly on the TimeSen2Crop dataset to perform the crop type classification. The service include both the training phase and the inference.

• Pre-Trained Long Short-Term Memory Neural Network: A pre-trained version of the LSTM on the TimeSen2Crop dataset is available in .h5 format. This network can be used to classify the specified tile harmonized using the monthly composite approach.
• Deep Network for pixel-level classification of S2 patches: This service provides functionality for users to train a custom pixel-level classifier of Sentinel 2 patches. For more information, please refer to: Deep Network for pixel-level classification of S2 patches.

For more details, AI4Copernicus - Technical Documentation

Objectives and usage.

These assets and services are primarily aimed at supporting agricultural monitoring and crop classification tasks using remote sensing data, specifically Sentinel-2 satellite imagery. They cater to different stages of the process, from data acquisition and harmonization to model training and deployment, offering a comprehensive set of tools for agricultural analysis and research. In particular, TimeSen2Crop allows the training and evaluation of machine learning and deep learning models for the crop classification tasks. The Harmonization service is used to pre-process Sentinel-2 data for harmonization and standardization to prepare the data for further analysis at large scale. The LSTM and the deep network for pixel level classification, including the pre-trained model, offer pre-built and trainable models for crop type classification. These models can be fine-tuned or used as-is to perform crop classification tasks, saving users the effort of developing deep learning model from scratch. Custom deep learning models can be derived from the tools provided.

For effective utilization of the AI services and tools, we suggest that users possess expertise in the following areas: Python, Machine Learning and Deep Learning, Docker and Containerization.