AI for Agriculture in Kenya: Crop Detection, Weather Prediction, and Food Security

Why Agriculture Is the Killer App for African AI

Agriculture employs 60-70% of sub-Saharan Africa’s workforce and contributes 15-35% of GDP across the continent. In Kenya specifically:

• 5.5 million smallholder farms (averaging 0.5-3 acres)
• 75% of agricultural output comes from smallholders
• 30-40% of crop yields are lost annually to pests, diseases, and weather
• $1B+ in crop losses per year from preventable diseases

These losses are not abstract. A maize farmer in western Kenya who loses 40% of their crop to fall armyworm isn’t just losing profit — they’re losing their family’s food security for the year.

AI offers interventions at every point in the agricultural value chain: before planting (which crop, where), during growth (disease detection, irrigation optimization), and after harvest (price prediction, supply chain logistics). This post covers the most impactful real-world applications.

Crop Disease Detection: The Flagship Use Case

The Problem

Maize (corn) is Kenya’s staple crop, grown by 4.5 million smallholder farmers. The two biggest threats:

• Maize Lethal Necrosis Disease (MLND): Viral disease, up to 100% yield loss
• Fall armyworm: Invasive pest, can destroy entire fields within weeks

Most smallholder farmers diagnose crop diseases by eye — and they often misdiagnose. By the time a disease is correctly identified, it’s often too late.

The ML Solution

Mobile-based crop disease detection using computer vision and transfer learning:

import tensorflow as tf
from tensorflow.keras.applications import MobileNetV3Small
from tensorflow.keras.layers import Dense, GlobalAveragePooling2D
from tensorflow.keras.models import Model

# Load a pre-trained model (trained on ImageNet)
base_model = MobileNetV3Small(
    input_shape=(224, 224, 3),
    include_top=False,
    weights="imagenet"
)

# Add disease-specific classification head
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(128, activation="relu")(x)
x = Dense(64, activation="relu")(x)
predictions = Dense(
    4, activation="softmax"  # healthy, MLND, fall armyworm, nutrient deficiency
)(x)

model = Model(inputs=base_model.input, outputs=predictions)

# Freeze base model layers initially
for layer in base_model.layers:
    layer.trainable = False

model.compile(
    optimizer="adam",
    loss="categorical_crossentropy",
    metrics=["accuracy"]
)

Real-World Dataset: The Kenya Maize Disease Dataset

Created by researchers at the University of Nairobi and CABI, this dataset contains:

• 15,000+ labeled maize leaf images
• 4 classes: Healthy, MLND, Fall Armyworm damage, Nitrogen deficiency
• Source: Photos taken by farmers on Tecno and Samsung budget phones
• Conditions: Varying lighting, angle, and background (not cleaned lab photos)

The dataset is available on HuggingFace: datasets/nairobi-agri/maize-kenya

Performance in the Field

After fine-tuning MobileNetV3Small on the Kenya dataset:

Metric	Value
Test accuracy	91.3%
Sensitivity (disease detection)	89.7%
Specificity (healthy identification)	93.2%
Inference time (budget phone)	145 ms
Model size (TFLite int8)	3.8 MB

Deployment Challenges

Challenge 1: Image quality variance Field photos are blurry, poorly lit, and taken at odd angles. The model must handle all of it. Solution: aggressive data augmentation during training.

data_augmentation = tf.keras.Sequential([
    tf.keras.layers.RandomFlip("horizontal_and_vertical"),
    tf.keras.layers.RandomRotation(0.3),
    tf.keras.layers.RandomZoom(0.2),
    tf.keras.layers.RandomBrightness(0.15),
    tf.keras.layers.RandomContrast(0.15),
])

Challenge 2: Class imbalance Fall armyworm damage is more common than MLND in most seasons. The dataset has 2x more armyworm images. Solution: weighted loss function and oversampling.

Challenge 3: Seasonal data drift Diseases look different in dry vs. wet seasons. The model trained on January-June images performs 5-8% worse on July-December images. Solution: continuous collection and periodic retraining.

Yield Prediction with Satellite and Weather Data

The Approach

Rather than relying on farmer-reported yields (often inaccurate), modern yield prediction uses:

1. Satellite imagery: Sentinel-2 (10m resolution, 5-day revisit) for NDVI (vegetation health index)
2. Weather data: CHIRPS rainfall data, ERA5 temperature data
3. Soil data: Africa Soil Information Service (AfSIS) soil property maps
4. Historical yields: County-level agricultural census data

import pandas as pd
import numpy as np
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.model_selection import train_test_split

# Feature engineering
def create_features(df):
    features = pd.DataFrame()
    
    # Vegetation features
    features["ndvi_mean"] = df.groupby("field_id")["ndvi"].mean()
    features["ndvi_std"] = df.groupby("field_id")["ndvi"].std()
    features["ndvi_trend"] = df.groupby("field_id")["ndvi"].apply(
        lambda x: np.polyfit(range(len(x)), x, 1)[0]
    )
    
    # Weather features
    features["total_rainfall"] = df.groupby("field_id")["rainfall_mm"].sum()
    features["rainfall_days"] = df.groupby("field_id")["rain_day"].sum()
    features["mean_temp"] = df.groupby("field_id")["temperature_c"].mean()
    
    # Soil features
    features["soil_organic_carbon"] = df.groupby("field_id")["soc_pct"].first()
    
    return features

# Train gradient boosting model
X = create_features(raw_data)
y = raw_data.groupby("field_id")["yield_ton_ha"].first()

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

model = GradientBoostingRegressor(
    n_estimators=200,
    max_depth=5,
    learning_rate=0.1,
    subsample=0.8,
    random_state=42,
)
model.fit(X_train, y_train)

print(f"R² score: {model.score(X_test, y_test):.3f}")
# Typical result: R² = 0.73-0.78 for county-level predictions

Startups Doing This Well

Apollo Agriculture (Kenya) is the standout. They use ML to provide smallholder farmers with:

• Credit scoring: Using satellite data, historical yields, and phone metadata to assess creditworthiness (no traditional credit history needed)
• Input recommendations: Optimizing seed varieties, fertilizer blends, and planting dates per field
• Insurance products: Index-based insurance triggered by satellite-verified drought or flood events

Apollo Agriculture serves 200,000+ farmers across Kenya and is expanding to Zambia and Nigeria. Their default rate is lower than traditional agricultural lenders — proof that ML-based credit scoring works in contexts without formal credit infrastructure.

Aerobotics (South Africa) provides drone and satellite-based crop health monitoring for larger commercial farms across Africa. Their AI platform detects pest infestations and irrigation issues at the individual tree level for orchards.

Weather Prediction and Climate Adaptation

The Gap

Global weather models (ECMWF, GFS) provide forecasts at 9-30 km resolution. A single degree of latitude at the equator is 111 km. For a farmer managing 2 acres in a microclimate zone, a 30 km resolution forecast is useless.

Localized Downscaling with ML

Researchers at the Kenya Meteorological Department have developed ML-based downscaling models that take global forecasts and produce localized predictions:

import xgboost as xgb

# Input: global model forecasts + local topography features
features = [
    "gfs_temperature_2m", "gfs_precipitation",
    "gfs_humidity", "gfs_wind_speed",
    "elevation", "distance_to_lake", "land_cover_type",
    "slope_aspect", "season_day"
]

X = df[features]
y_precip = df["local_precipitation_mm"]

model = xgb.XGBRegressor(
    n_estimators=300,
    max_depth=6,
    learning_rate=0.05,
    subsample=0.7,
    colsample_bytree=0.8,
)
model.fit(X, y_precip)

# Results: 5x improvement over raw GFS precipitation
# Original GFS: MAE = 8.2 mm/day
# Downscaled: MAE = 1.7 mm/day

This localized data feeds into decision support tools like Planting Date Advisors — a web app (also available as SMS for non-smartphone users) that tells farmers the optimal planting window based on predicted onset of rains.

Smart Irrigation

Kenya is a water-scarce country. Agriculture accounts for 75% of freshwater use, most of it inefficiently applied. Smart irrigation systems use ML to optimize water usage:

The Sensor ML Pipeline

1. Soil moisture sensors (cost: $15-30 each) deployed across the field send hourly readings
2. Weather forecast API provides predicted evapotranspiration rates
3. ML model predicts optimal irrigation schedule

class IrrigationOptimizer:
    def __init__(self):
        self.model = self.load_model()
        self.soil_moisture_thresholds = {
            "maize": {"min_vwc": 0.25, "max_vwc": 0.45},  # Volumetric Water Content
            "beans": {"min_vwc": 0.20, "max_vwc": 0.40},
            "kale": {"min_vwc": 0.30, "max_vwc": 0.50},
        }
    
    def predict_irrigation(self, crop_type, soil_vwc, forecast_et, days_since_rain):
        """Returns recommended irrigation volume in mm and time of day"""
        crop_params = self.soil_moisture_thresholds[crop_type]
        
        # Predict soil moisture 24h from now
        predicted_vwc = soil_vwc - forecast_et * 0.01  # Simplified
        # Account for crop type and growth stage
        adjustment = self.model.predict([[
            crop_type_encoded, soil_vwc, forecast_et, days_since_rain
        ]])
        
        if predicted_vwc < crop_params["min_vwc"]:
            irrigation_mm = (crop_params["max_vwc"] - predicted_vwc) * 20
            return {
                "irrigate": True,
                "volume_mm": round(irrigation_mm, 1),
                "best_time": "06:00" if forecast_et["morning"] < 3 else "18:00",
                "savings_vs_schedule": "32%"
            }
        return {"irrigate": False}

Real-world impact: A pilot project in Kiambu County found that ML-optimized irrigation reduced water usage by 32% while maintaining or improving yields.

Challenges and Lessons Learned

Dataset Quality

The biggest bottleneck for agricultural AI in Kenya is not algorithms — it’s data. Field photos taken by farmers are noisy. Yield reports are often inaccurate. Satellite data is weather-dependent (cloud cover blocks optical sensors for weeks during rainy seasons).

Lesson: Start with synthetic data or publicly available satellite data. Collect field data incrementally. A model trained on 1,000 real farmer photos often outperforms one trained on 10,000 curated lab photos.

Connectivity

Most agricultural AI tools need to work offline. A farmer in rural Busia County may have no internet for days at a time. All inference must be on-device (see post 4 in this series on TFLite). Model updates can sync when the farmer visits a town with connectivity.

Adoption

Building a good model is not enough. The farmer needs to trust it. The most successful deployments in Kenya have used:

• SMS interfaces (lowest barrier, works on any phone)
• Voice in local languages (Kiswahili, Dholuo, Kikuyu)
• Local champions (trusted farmers who verify AI recommendations against their own knowledge)

Business Model

Smallholder farmers cannot afford $5/month apps. Successful startups use:

• Freemium model: Free disease detection, paid yield prediction
• Bundle with inputs: Free AI + purchase of seeds/fertilizer from the same platform
• B2B2C: Sell to cooperatives and agribusinesses who distribute to farmers for free

The Road Ahead

Agricultural AI in Kenya is not a future possibility — it’s a present reality. Thousands of farmers already use AI-powered disease detection. Apollo Agriculture’s ML-based credit scoring has disbursed $50M+ in loans. The Kenya Meteorological Department’s downscaled forecasts reach 1M+ farmers via SMS.

The next frontiers:

• Multi-modal models: Combining satellite, weather, soil, and phone-camera data into unified advisory systems
• Crop-specific foundation models: Pre-trained vision models fine-tuned for cassava, beans, sorghum, and other African staples
• Farmer-facing LLMs: Question-answering systems in local languages that advise on pests, planting, and market prices

The ultimate metric of success won’t be model accuracy or paper citations. It will be whether Kenyan farmers lose less of their harvest, spend less on water and inputs, and sleep more soundly knowing their food supply is secure.

Final post in this series: Building and deploying AI in constrained environments — low-bandwidth and edge solutions for the real world.