AWS Builder Center

Builder Center

KisanGuard: Harnessing AWS Power for AI-Driven Climate Intelligence in Agriculture

KisanGuard: Harnessing AWS Power for AI-Driven Climate Intelligence in Agriculture

KisanGuard is an AI-powered climate intelligence platform designed to help Indian farmers assess heat stress, crop drought risk, and irrigation needs. It combines weather data, satellite imagery, and machine learning with AWS cloud services to generate actionable insights. Through multilingual dashboards, SMS, and voice alerts, KisanGuard aims to make climate information accessible, helping farmers anticipate environmental risks and make informed decisions to protect their crops and livelihoods.

Manjiri

Three machine learning models, an AWS-powered architecture, and one question: how do we turn climate data into decisions that matter in the field?

Part one: The forecast is not the answer

A weather forecast can tell a farmer that tomorrow will be hot. It cannot always tell them what that heat means for their crop, whether irrigation needs to change, or when working in the field may become unsafe.
That distinction is where KisanGuard begins.
India's farmers operate in conditions where weather, water availability, crop growth, and livelihoods are closely connected. Heat stress and drought do not arrive as isolated numbers on a screen. They affect the decisions farmers make every day.
We wanted to explore whether machine learning could help connect those pieces of information. Not by replacing a farmer's experience, but by making environmental information more specific, understandable, and actionable.
KisanGuard became our attempt to build that connection.
plaintext
Environmental Data
       |
       v
Feature Engineering
       |
       v
Machine Learning Models
       |
       v
Risk Assessment + Trust Checks
       |
       v
Actionable Advisory
       |
       v
Dashboard / SMS / Voice Alerts

Part two: The data does not arrive as one picture

Our first challenge was bringing different environmental signals together.
The project repository describes a pipeline using weather observations, NASA soil-moisture data, satellite-derived vegetation and land-surface information, and solar radiation data. Each source describes a different part of the environment, and each comes with its own limitations.
Weather grids, satellite observations, and soil-moisture products do not necessarily share the same resolution or update frequency. Satellite observations can be missing, and a regional estimate cannot automatically describe every individual field.
Combining these sources therefore requires more than downloading datasets and joining columns. We need to align timestamps, handle missing observations, engineer useful features, and identify data that should not be trusted.
A model can only learn from the information it receives. If that information is incomplete or misleading, a sophisticated architecture will not fix the underlying problem.



Part three: Three models, three different questions

We organized KisanGuard around three machine learning models because agricultural risk is not a single prediction problem.
Model 1: WBGT heat-risk forecasting
Our first model focuses on Wet Bulb Globe Temperature (WBGT), an indicator used to assess environmental heat stress. The repository describes an XGBoost and LSTM ensemble intended to generate hourly forecasts over a 48-hour horizon.
The aim is to turn temperature, humidity, wind, solar radiation, and satellite-derived features into a clearer picture of changing heat conditions.
Model 2: Crop drought-heat intelligence
The second model focuses on crop stress. It combines environmental indicators to estimate drought-heat conditions and crop-specific yield risk. Growth stage matters: heat during flowering can have different implications from heat during vegetative growth.
The goal is to move from describing the weather to understanding what those conditions could mean for a crop.
Model 3: Smart irrigation
The third model explores irrigation recommendations using Proximal Policy Optimization, a reinforcement-learning method. The repository describes scenario-specific policies for conditions such as drought, extreme heat, monsoon weather, and water scarcity, alongside a Management Allowable Depletion safety framework.
The question is not simply how much water to apply. It is how irrigation decisions might respond to changing conditions while respecting crop-critical moisture limits.
These models address connected problems, but their outputs still need appropriate evaluation before they can be treated as reliable field guidance.
plaintext
Model Prediction
       |
       v
Data Quality Checks
       |
       v
Confidence + Consistency Checks
       |
       +--------------------+
       |                    |
       v                    v
   Acceptable          Needs Review
       |                    |
       v                    v
 Generate Advisory     Human Review
       |                    |
       +----------+---------+
                  |
                  v
             Alert Decision

Part four: A prediction is not automatically trustworthy

This is one of the most important design decisions in KisanGuard.
A model can return a number even when the underlying data is incomplete. A high confidence score does not guarantee that the prediction reflects reality.
Our documented architecture includes a trust layer designed to check prediction confidence, missing data, unusual values, and conflicting signals. Amazon Augmented AI (A2I) is included for human review when a result requires additional scrutiny.
The intention is to avoid treating every prediction as equally reliable. An uncertain result should be identified, not presented with the same authority as a well-supported one.
Evaluation matters just as much. A model that accurately reproduces a value calculated from its input data has not necessarily demonstrated that it can predict future field conditions. Chronological testing, baseline comparisons, leakage checks, and independent field measurements are essential to establishing that distinction.
We want KisanGuard to earn trust through evidence rather than impressive-looking metrics alone.


Part five: AWS is the connection between the models and the farmer

A set of machine learning models is not yet a usable agricultural platform.
KisanGuard's architecture uses AWS services to connect data ingestion, model inference, validation, advisory generation, and alert delivery. Services described in the repository include Amazon S3, AWS Lambda, AWS Step Functions, Amazon SageMaker, Amazon Bedrock, Amazon Polly, Amazon SNS, and Amazon DynamoDB.
The intended workflow brings model outputs together, checks their reliability, generates a contextual advisory, and delivers information through appropriate channels.
Amazon Bedrock is included for generating localized agricultural guidance, while Amazon Polly supports converting text into speech. SMS, voice calls, push notifications, and WhatsApp voice notes are part of the documented communication design.
This architecture matters because useful information should not stop at an API response. It needs a path to the person who might use it.

Part six: Language is part of the engineering

A technically correct advisory is not useful if the intended recipient cannot understand it.
KisanGuard is designed to support Hindi, English, Marathi, Punjabi, Telugu, and Bengali. The aim is to communicate recommendations in familiar language and through channels that suit different users.
A dashboard can present detailed information. A voice message can communicate a short, urgent advisory. An SMS can deliver a concise alert.
These channels serve different needs, and they introduce their own challenges: messages must remain grounded in the model outputs, communicate uncertainty appropriately, and avoid overwhelming users with unnecessary notifications.
Accessibility is not a finishing touch. It is part of making the system useful.

Part seven: What we are learning

KisanGuard has taught us to think beyond whether a model runs or a dashboard looks complete.
Are the data sources suitable for the question? Does the evaluation test genuinely future periods? What happens when a satellite observation is missing? Can the system explain why an alert was generated? Is the recommendation safe to act on? How will we establish that the system works outside a development environment?
These questions influence the entire project, from data preparation and model design to cloud workflows and user experience.
  • The next step is not simply adding more features. It is strengthening evaluation, validating predictions against independent observations, and testing whether the resulting guidance is understandable and useful to farmers.
AWS service
Intended responsibility
Amazon S3
Store datasets and model artifacts
AWS Lambda
Run event-driven processing tasks
AWS Step Functions
Coordinate multi-stage workflows
Amazon SageMaker
Support machine learning workflows
Amazon Bedrock
Generate contextual agricultural advisories
Amazon Polly
Convert advisory text into speech

What comes next

KisanGuard is an effort to connect climate data, machine learning, cloud infrastructure, and practical agricultural decisions.
Our goal is to help farmers understand potential risks earlier and make more informed choices about heat, crop stress, and irrigation. Achieving that goal requires careful engineering, honest communication about limitations, and evidence from real-world testing.
We began with a question about how technology could serve farming communities. The work continues with a more demanding one:
Can we make the information reliable enough, accessible enough, and useful enough to matter when a farmer needs it?
That is what we are building KisanGuard to find out.

Explore the project
Research papers we refer to 1.  Impacts of rising temperatures and farm management practices on global yields of 18 crops  — Nature Food (2020). Relevant to KisanGuard’s heat-stress forecasting and crop-yield risk assessment*s.*2.  More accurate specification of water supply shows its importance for global crop production  — Nature Food (2022). Relevant to combining soil moisture, temperature, satellite observations, and crop productivity for drought and climate-risk analysis3.  Irrigation optimization with a deep reinforcement learning model: Case study on a site in Portugal  — Agricultural Water Management (2022). Directly relevant to KisanGuard’s reinforcement-learning-based irrigation recommendations and climate-aware water management

# amazon-sagemaker# amazon-polly# aws-amplify# amazon-cloudfront# amazon-ec2
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article