In today’s session of Satellite Data for Agricultural Economists: Theory and Practice, we compared tea field segmentation models in Kericho, Kenya, using MaxEnt, SDM ensemble, Google Embeddings, deep learning, and GeoTessera. Models were evaluated against independent point data with accuracy, precision, recall, F1, and Kappa. We visualized metrics and maps, then created a consensus map labeling a pixel as tea if at least three models agreed, reducing model-specific errors and improving reliability.

In this session, we demonstrated how to segment tea fields in Kenya using high-resolution Sentinel-2 imagery and manually digitized labels in Google Colab. Using torchgeo and torchseg, we loaded and visualized the raster and vector data, generated training patches on the fly, and trained a U-Net model for semantic segmentation. Training metrics such as loss, accuracy, and IoU were tracked, and the best model was saved. The trained model was then used to predict tea plantations across the full study area and evaluated against ground-truth points. This workflow provides a scalable pipeline for combining geospatial data with deep learning for agricultural mapping.

Trabajo 1 de la clase de valorizacion

Trabajo 1 de la clase de la semana

In today’s Satellite Data for Agricultural Economists session, we used Google Earth Engine to create and export labeled polygons for tea and non-tea areas in Kenya. Participants learned to generate cloud-free Sentinel-2 composites, apply cloud masks, visualize regions of interest, digitize training polygons for deep learning, and export datasets to Google Drive for further analysis, laying the foundation for spatial ML and deep learning in agriculture.

semana 5

Trabajo 1 de la clase de la semana 5

Trabajo 1 de la semana 5

Trabajo semana 5