AI4 E2O.GREEN AI assets are introduced as a tool for increasing agricultural productivity, an application of a deep learning model designed for object classification for observing grassland in golf and urban environments. The models have been trained on our custom-labelled dataset containing different grassland conditions as observed in Sentinel, drone, in situ and Planet Earthnet  imagery.

All the tools developed for golfing will have the superior generalization possibilities to be executed in different AOIs and use cases ranging from urban green irrigation to football, gated community, alpine pasture and motorway green space scenarios.

Set 1 of the assets is focused on boosting models vegetation index and evapotranspiration monitoring based on boosting algorithm as described in the text below:

Vegetation index change detection models are  first layer of alert for grassland irrigation interventions and  involve two key components. The first component is change detection, which aims to identify significant changes in grassland irrigation patterns. The second component involves supervised classification analysis of the change map generated from the change detection process.

For the implementation of the workflow on backend side and visualization on the frontend side, the following workflow is considered:

• Data Acquisition: Utilizing EO data brokers to acquire relevant Sentinel images for the irrigation intervention model.
• Preprocessing: Applying the Sentinel-2 preprocessing pipeline provided by the AI4Copernicus bootstrapping services. This pipeline will process the acquired Sentinel-2 products, generate a product with a common resolution for all bands in GeoTiff format, and apply land/sea and cloud masks.
• Change Detection: Employing the custom MAD change detection algorithm and Sentinel-2 Change Detection pipeline (for testing purposes) to compute the changes. Using a pair of S2-L2A products (NDMI) as input and applying the Change Vector Analysis approach.
• Generating NDMI Change Map: Obtaining the change map as the output of the Change Detection pipeline. This map will highlight significant changes in grassland irrigation patterns.
• Supervised Classification: Utilization of the custom-developed XGBoost model and custom pixel-level classifier service provided by the AI4Copernicus bootstrapping services (for testing purposes). Training of the classifier using Sentinel-2 patches to classify irrigation conditions, if possible, for mentioned use case (this model will be compared to the XGBoost model).
• Visualization: Developing visualizations to present the classification results in the frontend interface. This includes interactive maps to provide users with a clear understanding of the detected changes and classified features.

Set 2 of the assets is focused computer vision models combined with AR enabled mobile App as described in the text below:

The CV models connect VHR EO data, in situ imagery to computer vision models to enable virtual content such as computer vision predictions of turf disease and state, animations and annotations to be placed on top of a real world objects.

In the golf turf context the ML/CV models presented are introduced in order to resolve daily turf maintenance pertaining to grassland state and disease identification.

Full CV and web app design and implementation is presented  – Visual recognition presentation attached to this page