In today’s session of Satellite Data for Agricultural Economists: Theory and Practice, we focused on exam revision. We revisited past questions on machine learning and deep learning for satellite imagery, explored key concepts like patch size, batch size, and ethical considerations, and practiced applying evaluation metrics and binary maps for crops like tea and coffee. The exercises also covered temporal transfer risks, biodiversity impacts, and policy-relevant applications, helping students consolidate practical and theoretical knowledge before exams.

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.

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.

In today’s Satellite Data for Agricultural Economists session, we built spatial machine learning models using Tessera geospatial embeddings to map tea plantations in Kenya. Participants learned to download embeddings in Google Colab, prepare training data, train single- and multi-method ensemble models in R with sdm, create binary presence/absence maps, and calculate total tea area, showing how embeddings improve crop segmentation over traditional spectral bands.

Today’s session in our Satellite Data for Agricultural Economists course focused on building a spatial machine learning model using the new Google Embeddings dataset to map tea plantations in Kenya. Participants learned how to generate embeddings in Google Earth Engine, prepare training data, train multi-method ensemble models in R with the sdm package, create binary presence/absence maps, and estimate total tea area—highlighting how high-dimensional embeddings can improve crop segmentation compared to traditional spectral bands.

In today’s Satellite Data for Agricultural Economists session, we applied a pre-trained spatial machine learning model to map tea plantations in a different year—using Google Earth Engine composites, Area of Applicability diagnostics, and ensemble predictions to assess temporal transferability and quantify changes in tea area between 2020 and 2025.

Today’s session in our Satellite Data for Agricultural Economists course walked participants through building a spatial machine learning model in R to map tea plantations using Sentinel-2 imagery, training data, and ensemble methods for robust, high-resolution crop classification.

Teaching a hands-on module on Satellite Data for Agricultural Economists, covering Sentinel-2 preprocessing, cloud masking, canopy height integration, and building machine-learning training data in Google Earth Engine. Practical, coding-oriented, and focused on real agricultural applications.

Kicking off our Satellite Data for Agricultural Economists course with a hands-on introduction to R, geospatial data, and Google Earth Engine—guiding participants from setting up their first R project to extracting and visualizing real-world boundary data for applied agricultural analysis.

I update the sdm code on https://www.biogeoinformatics.org. I felt this was overdue since it was written in 2016 by Babak Naimi and many changes have been made to the package since then making it difficult for new users to rely on the tutorial.

Changing the font size of the legend.

Ensemble modeling in species distributions.

Visualizing srtm dem via rgee

Here I show simple code on how to get the proportion of pixel values in a raster that are above a given threshold.

Here I remind R trainers on terms that they should not take ligjtly when addressing new R users. At the same time, I encourage new R users to receive such statements positively. It is all to their own advantage.

A basic introduction to vegan package for ecologists in R.

A simple model of the potential suitable habitat of Phago loricatus of Western Africa.

Following the work at RUFORUM I made some efforts on ggscatter.

Changing layer values but retaining names does not affect predict function on sdm models.

Here I demonstrate how to make a simple 3D map in R.

Here I model Potential suitable habitats for the black rhino in both Kenya and Tanzania territories.

Demonstration of using ViewPipeSteps to print data at every step of piping in R.

Here I demonstrate how to reclassify raster object and calculate area under each class.

Call to improve sdm to accept sf, stars, and terra data output for model building.

A short exploration of the afrilearndata package

Here I demonstrate the power of mapsf as a mapping package in R.

Update of the previous plot with code

As an appreciation to exemplary mentor.

A short article which compares pixel values extracted within a buffer zone of species occurrence records to assess whether some species records are within 'very unclikely' habitats hence worth being left out while running sdm.

Running kmeans across raster stack in R.

Here I demonstrate how to mutate a column with some values as NA

Making citations within .Rmd file in R.

I demonstrate how to reclassify a raster layer and generate calculate the area under each class plus generate a plot.

Generating a legend for multiple column data can be challenging. Here I demonstrate how it can be done with simple code.

Tsavo is highly likely to be unsuitable habitat for Panthera leo by 2070 under RCP 8.5 climate scenario.

National Parks are confined areas and climate change is shifting suitable habitats. Will Tsavo National Park still be suitable for Panthera leo in the year 2070. Under RCP 8.5, the answer is sad NO. In fact, even Amboseli won't be suitable for the pride of Africa. I built the model using sdm package in R 4.1.1 with random forest, boosted regression trees and support vector machine under both bootstrapping and subsampling replications.

Demonstration of mapping from sf object, through dplyr and units to mapsf.

A very important insect that feeds on aphids and other crop-damaging pests. In the process, it also aids in pollination. Here we model its potential suitable habitat in South Africa using sdm package and an ensemble of rf, svm and glm algorithms in R 4.1.0.

Showing which schools are near which hospitals in constituency X. Popular geospatial analysis concept. Schools share the same color code as their nearest hospital. Constituency polygon, Schools points, and Hospitals points were randomly drawn in QGIS and imported to R for this demo purposes only. However, you can reach me for the random data and code used for your adoption freely. Happy coding in R.

For better planning within a constituency, the knowledge of which which health facility is neared to which schools is key. This can help to easily tell which health facilities are serving more schools and may need more staff or expansion of some sort.

It is fun mapping the distribution of this magnificent bird in Pakistan. Nest building, egg incubation, and feeding hatchlings are done by females. The pictured male defends territory, can be aggressive even to his own reflection!

Getting started with Bayesian Statistics. Finding the approach to be interesting and real.

Potential suitable habitat of dikdik (Esuruo in Turkana Language). It is surprising that the species is available in northwestern Kenya, the border with Uganda but since no occurrence records from that region, the habitat is generally modelled unsuitable. Typical wrong model.

Potentially suitable habitat of Panthera leo in Kenya. sdm package version 1.0.89, in R 4.1.0. Plots are replicates of random forest method under bootstrapping (top three) and subsampling (bottom three).

Pollinators like sunbirds are crucial for the survival and regeneration of flowering plants globally. We may need to relocate those native to Socotra island to the mainland as future climates diminish their suitable habitat within their native lonely island. Rescue the species before it is gone.

Filtering through DAFNEE (https://dafnee.isem-evolution.fr/) for your journal of choice made easier.

Some fun time

Simple map of Kenya Counties by Area. Having histogram as a legend with mapsf package (https://riatelab.github.io/mapsf/articles/web_only/how_to_create_inset_maps.html)

This is to further test whether there is a 'strange' mismatch when publishing raster, polygon and points overlaid using mapsf

Took me some little sweat to have vectors overlaid on raster with mapsf. Distribution of sunbird (Cinnyris coccinigastrus) from GBIF in West Africa.

Some fun moment with mapsf

Simple map of Uganda. Variable mapped are random numbers.

This is a demo of mapping some random variables per county in Kenya. The skill/technique can be transferred to actual values like maize yield or the number of elephants. Again all glory to R programming language, mapsf and raster packages, plus my awesome ubuntu based system.

A simple map of the Kenyan counties. Data downloaded from GADM using raster package's getData function and mapped using mapsf package.

The BIG FIVE (Leopard, Lion, Buffalo, Elephant, and Rhino) are very important to Kenya's tourism economy and ecological functions within the African savannah of Kenya. How much of their suitable habitat fall within protected areas is therefore an important information for their conservation. This is a work in progress to establish the proportion of suitable habitats within protected areas for the BIG FIVE in Kenya.

This is one insect that distributes animal dung across savannah thus potentially influencing soil nutrient dynamics. Its potential suitable habitat is therefore of importance for range management. It would be nice to correlate the density of dung beetles and soil fertility levels in an area.

mapsf development version has the mf_background() function that allows for setting image background of choice and add maps on top. Cool.

..and the color shades are elegant.

The power of mapsf package in R still amazes me.

Central and parts of western Kenya are cooler as evident from worldclim tmin data. I used counties shapefile from GADM to extract mean temperature for each county.

Potentially suitable habitats and standard deviation around those predictions in Africa. Occurrence points from GBIF are also indicated.

How much suitable habitat for critically endangered Vulpes bengalensis do we still have in India? How certain/uncertain are we about the predictions? Find the answers, at least some, from the interactive map.

The suitable habitat of the critically endangered Priodontes maximus in South America is quite narrow and fragmented. The extensible sdm package in R is helping us to visualize this and a call for robust conservation strategies is inevitable.

This is a simple distribution model of wildebeest (Connochaetes taurinus) in Eastern and Southern Africa. Even though there appears to be a more suitable habitat for the species in Southern Africa than in Eastern, the uncertainty around the prediction is equally higher in Southern Africa.

Map showing the potential suitable habitat for South Andean Deer (Hippocamelus bisulcus) within native Chilean border.

Potentially suitable habitats of Cannabis sativa and occurrence records (red dots) in Northern Americas.

Honeycomb plot on the relationship between Ozone levels and temperature.

Population and the median income of Martinique Island

Distribution of life expectancy across continents between 1952 and 2007

200 shallow wells in semi-arid lands of northern Kenya (Turkana County) where the nomads could get water for their livestock.

Trying to help nomadic pastoralists in semi-arid lands of northwestern Kenya (Turkana County) to know WHERE water points are located. This can be very important information to them during extreme drought events.

A map of 150 points WHERE nomadic pastoralists in northern-western Kenya (Turkana County) could find water. Positional attributes are accessible by clicking a point.

Improving the knowledge of nomadic pastoralists in Northern Kenya on WHERE 'oases' are found within their semi-arid environment. Click points to get attributes. This is a project in progress...

Getting more points, hitting 40 points. Click a point to get some attributes.

Knowledge of WHERE water points are located in an arid environment is very crucial. In this project, I map wells and dams in Northern Kenya semi-arid environment especially Turkana County.

How could I forget to better label the y axis.

Rotating the equation of the abline to better align beside the line itself.

I love ggplot2 products, but I thinks this is also awesome with Base R plot.

Simple plot of top 1000 observations.

Stretching the limits of using DTM and DSM lidar layers for vegetation lifeforms classification and other assessments.

NDVI for the month of December 2020 in Loima Sub-County, Turkana County in Kenya. Possible projection/prediction of 'greenness' indices could help pastoralists design their seasonal migration with cattle in an informed manner.

Generating diamonds plot against price using assorted packages within tidyverse.