/

Fall armyworm reached 70+ countries after arriving in Africa in 2016; traditional field scouting in Mali was too slow to catch infestations before yield loss. Omdena partnered with OKO Finance to build a satellite-based pest detection pipeline on Sentinel-2 and Landsat-8 imagery. Delivered: a ResNet50 model at 94% classification accuracy, a four-class damage assessment system, and a FastAPI inference API with Streamlit dashboard.
| Outcome | Detail |
|---|---|
| Classification accuracy | 94% (ResNet50 deep learning model) |
| Damage area prediction | 0.39 RMSE (field-level area estimate) |
| Satellite images analyzed | 2,000+ Sentinel-2 and Landsat-8 images |
| Infestation categories | 4 severity levels: Zero, Low, Medium, High |
| Project duration | 10 weeks, June–August 2021 |
| Active collaborators | 30 |
| Final deliverable | FastAPI inference API + Streamlit dashboard + Docker container |
The fall armyworm (Spodoptera frugiperda) arrived in Africa in 2016 and spread to more than 70 countries within five years. In Mali and neighboring West African nations, it attacks maize—the staple crop for millions of smallholder farmers—sometimes devastating entire fields before detection. Early identification of infestation severity is essential for reducing yield losses and enabling timely, targeted interventions.
Traditional detection relies on trained scouts inspecting fields on foot. The process is labor-intensive, expensive, and too slow. By the time an infestation was confirmed and intervention materials reached affected farmers, the damage was often irreversible. For OKO Finance, accurate and timely damage data was foundational to pricing risk and processing claims fairly.
Omdena partnered with OKO Finance, a West African crop microinsurance company that designs insurance products for smallholder farmers across Mali and neighboring countries. The client needed a data-driven method to understand where and how severely fall armyworm was striking before policyholders filed claims, a capability that did not exist in its operational toolkit.
Thirty collaborators from the Omdena network contributed to the ten-week project, which ran from June to August 2021. The team combined expertise in remote sensing, computer vision, geospatial analysis, and machine learning to build a proof-of-concept detection pipeline grounded in publicly available satellite imagery and open-source field data.

The team assembled training data from six sources: Sentinel-2 and Landsat-8 for multispectral imagery, FAO fall armyworm scouting records for infestation labels, Radiant ML Hub’s 2018 African landcover dataset, administrative boundaries for geolocation filtering, and LULC grids to retain only high-vegetation areas.
Eight vegetation indices were extracted from the red, green, blue, and near-infrared bands: NDVI, GNDVI, RDVI, CVI, GCI, SAVI, TVI, and ARVI. SMOTE (synthetic minority oversampling technique) addressed the class imbalance. K-means clustering removed FAO geolocation outliers by filtering records more than 3 kilometers from administrative boundaries, retaining only village and hamlet locations with confirmed vegetation coverage.

The team evaluated five ML classifiers—KNN, random forest, SVM, logistic regression, and decision tree—alongside ResNet50, EfficientNet, InceptionV3, and U-Net. KNN with n=2 neighbors reached 56% test accuracy. ResNet50 required two data collection rounds: an initial run on 5,000 images yielded 47%; a second round using 1,040 larger images tiled into 26,000 patches pushed accuracy to 94%, with 0.39 RMSE on damage area estimation.
The final classification system assigned each field location to one of four severity categories: Zero infestation (no damage), Low (0–20% field area affected), Medium (20–50% affected), and High (50–100% affected). The ResNet50 model received three stacked vegetation-index channels—GCI, NDVI, and SAVI—as input and produced both a class label and a continuous area estimate in a single forward pass.
The complete pipeline was deployed as a FastAPI service accessible via REST. A Streamlit web application let users enter a location and date, then view vegetation indices and an infestation severity prediction returned as a JSON response. The full system was containerized with Docker for independent deployment.

Before this project, the client had no scalable, automated method for detecting fall armyworm infestations or assessing crop damage levels without deploying field scouts. The delivered system provided a first operational proof of concept: a pipeline that takes a geographic coordinate and a date, queries publicly available satellite data, and returns a damage severity estimate, with no ground-level access required.
The more consequential finding emerged from data validation rather than modeling: the FAO scouting dataset, widely used as ground truth in fall armyworm research, had a data-quality problem severe enough to compromise any model trained on it without filtering, corroborated by expert review from PlantVillage at Penn State University, and relevant to any team working on satellite-based pest monitoring in West Africa.
FAO geolocations pointed to urban areas in more than 60% of cases, rendering those records unusable as training labels. Filtering by administrative boundaries and LULC vegetation grids reduced the dataset but improved reliability; without that step, every model trained would have been learning from noise.
The highest-infestation category represented only 5% of the dataset, and every model struggled to detect it. SMOTE improved the class balance, but synthesized samples cannot substitute for real high-severity observations. Future deployments will require a deliberate strategy to collect geolocated, photographically confirmed high-infestation records.
ResNet50 reached 94% accuracy in controlled evaluation, but that figure reflects a filtered and partially synthesized dataset, not live unvalidated satellite imagery. Moving to production will require verified ground-truth photographs paired with satellite acquisitions before accuracy can be claimed at operational scale.
The team built and delivered a complete sequential pipeline: satellite image collection, automated vegetation-index extraction across eight spectral indices, model training across machine learning and deep learning architectures, four-class damage severity classification, and a containerized inference API, end-to-end with no manual steps between coordinate input and severity prediction.
ResNet50 combined infestation classification and damage area estimation in a single forward pass, a dual-output architecture no simpler model could replicate. Stacking GCI, NDVI, and SAVI as three input channels extracted feature representations that scalar vegetation-index values alone could not capture, demonstrating the value of deep convolutional architectures for agricultural remote sensing.
By verifying FAO scouting coordinates against administrative boundaries and LULC vegetation maps, and confirming the problem with PlantVillage domain experts, the team documented that the most widely used fall armyworm ground-truth dataset in West Africa was majority-unreliable, alerting future teams to apply the same filtering steps before training.
Omdena partnered with OKO Finance over ten weeks in 2021 to build a satellite-based fall armyworm detection pipeline for southern Mali. Thirty collaborators analyzed 2,000+ Sentinel-2 and Landsat-8 images, trained machine learning and deep learning models for four-class damage classification, and delivered a FastAPI inference API, a Streamlit dashboard, and a Docker container as the final output.