Problem Statements Analysis

From what to build to how to build it — our platform helps you understand, plan, and solve SIH 2026 problem statements with complete insights into requirements, solution approaches, technology stacks, architecture, AI/ML, and implementation roadmaps.

Scroll down
01 / SIH 2026

PS: SIH26131

Early detection and management of crop diseases and pest infestations

Description: • Problem Description Farmers often recognise crop diseases or pest infestations only after visible damage has spread. Extension staff may cover large areas, while laboratory diagnosis and expert advice may not be immediately available. Weather, crop stage, variety, soil condition and local pest history influence risk, but these inputs are rarely combined into actionable farm-level alerts. Incorrect diagnosis may lead to delayed treatment, excessive or inappropriate pesticide use, increased cultivation cost,residue concerns and yield loss. The challenge is to provide timely, reliable and locally relevant detection,forecasting and management support. • Expected Solution / Outcome A farmer- and extension-worker-friendly crop-health system that supports image based symptom identification, pest-trap or sensor inputs, weather-based risk forecasting, geospatial hotspot mapping,expert validation and multilingual advisories. The system should recommend integrated pest and disease management actions, safe input usage,referral to extension or laboratories, and follow-up monitoring. It should learn from field confirmations and provide dashboards for agriculture officials.Expected outcomes include earlier detection, reduced crop loss, more targeted pesticide use, faster extension response, improved surveillance coverage and better planning of preventive interventions.
Organization: Government Of Maharashtra
Department: Maharashtra State Innovation Society, Department of Skills, Employment, Entrepreneurship and Innovation
Category: Software
Theme: Agriculture, FoodTech & Rural Developmenttd>
Youtube Link: Not provided
Dataset Link: Not provided
Contact info: Not provided

CropHealth-AI: Integrated Pest & Disease Early Detection, Forecasting, and Surveillance Platform for Maharashtra

1. Executive Snapshot

Summary

CropHealth-AI is an end-to-end, offline-first digital agricultural platform designed specifically for smallholder farmers, extension officers (Gram Sevaks/Krishi Sahayaks), and state agricultural administrators in Maharashtra [INFERENCE]. By fusing mobile edge computer vision for multi-symptom diagnosis, automated/manual pest-trap sensor logging, multi-factor micro-climate forecasting (combining weather, soil, crop stage, variety, and historical incidence), and geospatial hotspot analytics, the platform shifts crop protection from reactive chemical treatment to proactive Integrated Pest & Disease Management (IPM/IDM) [INFERENCE]. The system operates seamlessly in low-connectivity rural environments, routes ambiguous diagnoses to Krishi Vigyan Kendra (KVK) agronomists for human-in-the-loop validation, dispatches localized Marathi advisories, and continuously improves model accuracy via field confirmations [INFERENCE].

Metadata Table

Attribute Detail
PS Number SIH26131 [PS]
Title Early detection and management of crop diseases and pest infestations [PS]
Organization Government Of Maharashtra [PS]
Department Maharashtra State Innovation Society, Department of Skills, Employment, Entrepreneurship and Innovation [PS]
Category Software [PS]
Theme Agriculture, FoodTech & Rural Development [PS]
Deadline 30 September 2026 [PS]
Ideas Counter 2/500 [PS]

Fundamental System Objective

To deliver a field-ready, offline-capable digital surveillance and advisory ecosystem that combines multi-modal AI diagnosis, weather-agronomic forecasting, expert validation, and GIS hotspot mapping to reduce crop yield loss and pesticide overuse across rainfed and irrigated farming clusters in Maharashtra [INFERENCE].

2. Problem Deconstruction

The problem statement explicitly outlines key operational gaps and desired system capabilities [PS]. The atomic breakdown preserves all original requirements:

  • Requirement 2.a — Visual Disease & Pest Diagnosis: "supports image based symptom identification" [PS].
    • Interpretation: Farmers and extension workers can take leaf/stem/fruit photos on budget Android devices and receive instant diagnostic predictions [INFERENCE].
    • Implementation Implication: On-device Quantized Computer Vision (TFLite/ONNX) model paired with cloud fallback for low-confidence scans [INFERENCE]. Tag: [PS].
  • Requirement 2.b — Trap & Sensor Data Ingestion: "pest-trap or sensor inputs" [PS].
    • Interpretation: The platform must log insect counts from manual pheromone traps, sticky traps, or automated camera-enabled IoT traps [INFERENCE].
    • Implementation Implication: Structured input forms for extension staff and API endpoints for IoT trap telemetry [INFERENCE]. Tag: [PS].
  • Requirement 2.c — Multi-Factor Weather & Agronomic Risk Forecasting: "weather-based risk forecasting... Weather, crop stage, variety, soil condition and local pest history influence risk, but these inputs are rarely combined into actionable farm-level alerts." [PS].
    • Interpretation: Risk assessment cannot rely on images alone; it must synthesize 5 key variables: weather forecast, crop phenological stage, crop variety susceptibility, soil characteristics, and localized historical infestation trends [INFERENCE].
    • Implementation Implication: A multi-factor predictive risk engine that generates daily farm-level vulnerability scores before visual symptoms appear [INFERENCE]. Tag: [PS].
  • Requirement 2.d — Geospatial Hotspot Mapping: "geospatial hotspot mapping" [PS].
    • Interpretation: Pest and disease outbreaks must be aggregated geographically to identify cluster spreading at Gram Panchayat, Taluka, and District levels [INFERENCE].
    • Implementation Implication: Spatial clustering algorithms (e.g., DBSCAN/Getis-Ord Gi*) rendering real-time heatmaps on an interactive GIS administrative dashboard [INFERENCE]. Tag: [PS].
  • Requirement 2.e — Expert Human-in-the-Loop Validation: "expert validation" [PS].
    • Interpretation: Automated AI diagnosis can be uncertain or wrong; field cases requiring high accuracy or involving quarantine pests must be escalated to human agronomists [INFERENCE].
    • Implementation Implication: Triage queuing system for KVK scientists/extension officers to review flagged images, update diagnoses, and approve targeted advisories [INFERENCE]. Tag: [PS].
  • Requirement 2.f — Localized Multilingual Communication: "multilingual advisories" [PS].
    • Interpretation: Information must be delivered in local languages (primarily Marathi, with Hindi/English options) through audio/text to overcome literacy barriers [INFERENCE].
    • Implementation Implication: Voice-enabled localized advisory engine using Marathi text-to-speech (TTS) and localized structured advisories [INFERENCE]. Tag: [PS].
  • Requirement 2.g — Actionable IPM/IDM & Safe Input Guidance: "recommend integrated pest and disease management actions, safe input usage" [PS].
    • Interpretation: Recommendations must prioritize biological/cultural controls and provide exact chemical dosage and safety guidelines (dosage per acre, water ratio, PHI - Pre-Harvest Interval) to eliminate residue and overuse issues [INFERENCE].
    • Implementation Implication: Rule-based agronomic decision tree aligned with ICAR/CIB&RC package of practices and dosage calculators [INFERENCE]. Tag: [PS].
  • Requirement 2.h — Escalation & Diagnostic Referral: "referral to extension or laboratories" [PS].
    • Interpretation: Severe, unknown, or systemic outbreaks must trigger physical visits from extension workers or physical sample collection for laboratory testing [INFERENCE].
    • Implementation Implication: Digital ticket system generating referral passes with geotagged plot locations for state diagnostic labs [INFERENCE]. Tag: [PS].
  • Requirement 2.i — Follow-up Monitoring & Efficacy Feedback: "follow-up monitoring" [PS].
    • Interpretation: System must track whether recommended actions were taken and whether the pest/disease was successfully controlled after 7–14 days [INFERENCE].
    • Implementation Implication: Automated push/SMS reminders prompting farmers to upload follow-up pictures or update status [INFERENCE]. Tag: [PS].
  • Requirement 2.j — Active Field Confirmation & Continuous Model Learning: "It should learn from field confirmations" [PS].
    • Interpretation: Ground truth validations by extension officers and laboratory reports must be fed back into the training pipeline to prevent concept drift [INFERENCE].
    • Implementation Implication: Active learning pipeline and dataset versioning system for re-training AI models on verified local field images [INFERENCE]. Tag: [PS].
  • Requirement 2.k — Administrative Surveillance Dashboards: "dashboards for agriculture officials" [PS].
    • Interpretation: State/District officials need macro-level visibility for resource allocation (pesticide inventory, extension staffing, emergency containment) [INFERENCE].
    • Implementation Implication: Multi-tier web dashboard displaying real-time incidence rates, outbreak vectors, response times, and pesticide usage patterns [INFERENCE]. Tag: [PS].

3. Stakeholders & User Needs

Stakeholder Goal Current Pain System Need
Smallholder Farmer (Kharif/Rabi growers in Maharashtra) Maximize crop yield, reduce input cost, detect pests/diseases before widespread loss [INFERENCE]. Detects symptoms too late; receives generic advice; struggles with complex chemical calculations; literacy/language barriers [RESEARCHED] [1.10, 2.5]. Offline mobile app with camera diagnosis, Marathi voice/text advisories, exact dosage calculator, and risk forewarning [INFERENCE].
Agricultural Extension Staff (Gram Sevak, Krishi Sahayak) Efficiently cover large geographical areas (often 5–10 villages per officer) [RESEARCHED] [1.10]. Overwhelmed by travel distances; lacks diagnostic tools for rare diseases; manual report logging [RESEARCHED] [1.10]. Geotagged field reporting, trap logging module, automated diagnostic assistance, and offline sync [INFERENCE].
Agronomist / KVK Scientist Validate complex crop health cases and issue scientifically sound advisories [INFERENCE]. Flooded with unorganized farmer calls; lacks structured image and micro-climate context [INFERENCE]. Web-based expert triage workspace with structured image history, weather logs, and one-click advisory dispatch [INFERENCE].
District/State Agriculture Official (Dept. of Agriculture, Maharashtra) Early containment of outbreaks (e.g., Pink Bollworm in Cotton, Fall Armyworm in Maize, Rust in Soybean) [RESEARCHED] [1.5, 1.10]. Fragmented manual reports, delay in identifying cluster hotspots, lack of real-time field intelligence [RESEARCHED] [1.4, 1.10]. Real-time GIS hotspot map, trend forecasting analytics, response monitoring, and automated pest alert dispatches [INFERENCE].

4. Requirements Traceability

Proposed System Feature Direct Requirement Justification Core Functional Mechanism
Edge Vision Diagnostic Engine Req 2.a (Image-based symptom identification) [PS] On-device MobileNetV3/YOLOv8-nano classifier running offline, falling back to server-side Swin Transformer for low confidence [INFERENCE].
Pest Trap & Sensor Logging Module Req 2.b (Pest-trap or sensor inputs) [PS] Digital trap counter UI for manual sticky/pheromone trap counts + REST API for automated IoT camera trap telemetry [INFERENCE].
Agronomic Multi-Factor Risk Engine Req 2.c (Weather-based risk forecasting with 5 parameters) [PS] Rule-based & XGBoost model combining IMD/Open-Meteo weather, crop phenology, variety vulnerability, soil matrix, and 3-year historical pest logs [INFERENCE].
PostGIS Hotspot Mapping Engine Req 2.d (Geospatial hotspot mapping) [PS] Spatio-temporal cluster detection (DBSCAN) plotting incidence intensity on interactive admin map [INFERENCE].
KVK Expert Verification Portal Req 2.e (Expert validation) [PS] Human-in-the-loop web queue for low-confidence AI scans (<85% confidence) or priority quarantine pests [INFERENCE].
Marathi Multilingual Advisory System Req 2.f (Multilingual advisories) [PS] Text-to-speech (e.g., AI4Bharat Indic-TTS / Google TTS) producing audio advisories in Marathi dialect [INFERENCE].
Safe Input & Dosage Calculator Req 2.g (Integrated pest management & safe input usage) [PS] CIB&RC-compliant chemical dosage calculator based on field area, pump volume, and mandatory Pre-Harvest Intervals (PHI) [INFERENCE].
Lab & Extension Referral Flow Req 2.h (Referral to extension/labs) [PS] Automated ticket generation with QR-coded digital pass sent to nearest soil/plant pathology lab [INFERENCE].
Follow-Up Loop & Active Learning Req 2.i (Follow-up monitoring) & Req 2.j (Learn from field confirmations) [PS] Scheduled SMS/push prompts for Day 7/14 re-scans; confirmed ground-truth data auto-saved to ML retraining pipeline [INFERENCE].
State Level Analytics Dashboard Req 2.k (Dashboards for agriculture officials) [PS] Web GUI providing Taluka-level outbreak alerts, extension coverage metrics, and pesticide usage trends [INFERENCE].

5. Constraints & Environment Realities

  1. Connectivity Limitations: Over 60% of rural farmlands in Marathwada and Vidarbha experience frequent 2G/3G degradation or complete signal loss [INFERENCE]. System must execute image diagnosis and display local advisory offline [INFERENCE].
  2. Hardware Constraints: Farmers use low-end Android smartphones (2GB–3GB RAM, budget processors, low-resolution cameras) [INFERENCE]. Mobile application binary size must be under 30MB, with model footprint under 12MB [INFERENCE].
  3. Language & Digital Literacy: The primary language across rural Maharashtra is Marathi [KNOWN]. Many smallholder farmers struggle with text-heavy UI [INFERENCE]. UI must be icon-driven with voice input/output in Marathi [INFERENCE].
  4. Agro-Climatic Diversity: Maharashtra features 9 distinct agro-climatic zones ranging from heavy-rainfall Konkan coastal belt to drought-prone Rain Shadow areas in Madhya Maharashtra and Eastern Vidarbha [KNOWN]. Risk models must be tuned to zone-specific microclimates [INFERENCE].
  5. Seasonal Crop Dynamics: Major Kharif crops (Cotton, Soybean, Paddy, Tur) and Rabi crops (Gram, Jowar) have tight cultivation windows where delay in advisory (>48 hours) leads to total crop loss [RESEARCHED] [1.10].

6. Research Questions

  1. What are the official Economic Threshold Levels (ETL) and weather triggers for primary Maharashtra crop pests (e.g., Pink Bollworm in Cotton, Spodoptera litura in Soybean, Stem Borer in Rice)? [RESEARCHED] [1.5, 1.10]
  2. How can AgriStack's Farmer ID and Crop Sown Registry be integrated without custom API access during the hackathon phase? [RESEARCHED] [2.1, 2.3]
  3. How can image quantization impact detection accuracy on blurred leaf images taken under harsh sunlight field conditions? [INFERENCE]
  4. What existing public domain datasets cover Indian crop varieties and regional pest manifestations? [RESEARCHED] [2.5]
  5. How can spatial clustering prevent privacy leakage while accurately identifying village-level disease hotspots? [INFERENCE]

7. Data Requirements & Sources

Data Type Purpose Ideal Real-World Source Hackathon Substitute Access Method Availability Privacy Sensitivity Confidence Tag
Crop Leaf Images Train & test multi-class disease classifier [INFERENCE]. ICAR-NCIPM field repository / Plantix DB [RESEARCHED] [2.5, 3.4]. PlantVillage dataset + Kaggle Indian Paddy/Cotton/Soybean disease datasets [RESEARCHED]. Open HTTP Download / Script [KNOWN]. High [KNOWN]. Public / Non-sensitive [KNOWN]. [RESEARCHED]
Weather Data (Temp, Humidity, Rainfall, Wind) Feed micro-climate pest forecasting engine [INFERENCE]. IMD Gramin Krishi Mausam Sewa (GKMS) / AWS Stations [RESEARCHED] [4.2]. Open-Meteo API / ERA5 Reanalysis historical weather data [RESEARCHED]. REST API / CSV ingestion [KNOWN]. High [KNOWN]. Non-sensitive [KNOWN]. [RESEARCHED]
Crop Phenology & Soil Matrix Calculate crop stage vulnerability and soil moisture stress [INFERENCE]. AgriStack Crop Sown Registry + ICAR Soil Survey [RESEARCHED] [2.1, 2.3]. Synthetic agronomic lookup tables based on sowing date and ICAR crop calendars [ASSUMPTION]. Static JSON / Local DB [ASSUMPTION]. Simulated [ASSUMPTION]. Non-sensitive [INFERENCE]. [ASSUMPTION]
Pest Trap Telemetry Quantify regional pest population density [INFERENCE]. Automated IoT Pheromone Trap feeds / CROPSAP manual trap logs [RESEARCHED] [1.1, 1.10]. Synthetic trap log generator simulating daily bollworm counts per trap [ASSUMPTION]. Mock WebSocket / JSON API [ASSUMPTION]. Simulated [ASSUMPTION]. Non-sensitive [INFERENCE]. [ASSUMPTION]
Historical Pest Outbreaks Spatial baselines for historical infestation risk [INFERENCE]. Maharashtra Dept of Ag CROPSAP 15-year GIS database [RESEARCHED] [1.7]. Synthesized district-level historical disease incidence CSVs [ASSUMPTION]. Static DB tables [ASSUMPTION]. Simulated [ASSUMPTION]. Non-sensitive [INFERENCE]. [ASSUMPTION]
Farmer Plot Coordinates Map geospatial hotspots [INFERENCE]. AgriStack Geo-referenced plot registry [RESEARCHED] [2.1, 2.3]. Simulated geotagged coordinates centered around Marathwada/Vidarbha districts [ASSUMPTION]. GeoJSON / PostGIS spatial query [KNOWN]. Simulated [ASSUMPTION]. Sensitive (requires 100m blurring) [INFERENCE]. [INFERENCE]

Data Categorization Summary

  • A. Explicitly Mentioned: Images of crop symptoms, trap/sensor counts, weather readings, crop stage, crop variety, soil conditions, local pest history [PS].
  • B. Data Required by Inference: CIB&RC approved pesticide formulation database, Pre-Harvest Intervals (PHI), KVK expert location mapping, audio Marathi translation tokens [INFERENCE].
  • C. Data to collect: Ground-truth confirmation tags from KVK experts, follow-up efficacy reports [PS].
  • D. Synthetic Data: Mock trap counts, simulated AgriStack plot boundaries, synthetic weather alerts for demo [ASSUMPTION].
  • E. Data Needing Validation: Live API schema for Maharashtra state AgriStack server integration [NEEDS VALIDATION].

8. Existing Solutions & Gap Analysis

Existing System Description Strengths Weaknesses Relevance Unaddressed Gap Source Confidence Tag
Plantix Mobile App Popular AI crop disease diagnostic app [RESEARCHED] [2.5]. High image classification accuracy; large global dataset [RESEARCHED] [2.5]. Requires active internet connection; does not integrate weather-soil risk forecasting; generic commercial product ads [RESEARCHED] [2.5]. High benchmark for image UX [INFERENCE]. Does not combine weather, soil, variety, and trap counts into multi-factor forewarning [INFERENCE]. Mayank Digital Labs [RESEARCHED] [2.5] [RESEARCHED]
CROPSAP / NPSS (Govt of India / MH) Maharashtra Crop Pest Surveillance & Advisory / National Pest Surveillance System [RESEARCHED] [1.1, 3.9]. Official government reach; extension worker network integration; structured pest reports [RESEARCHED] [1.1, 1.10, 3.9]. Historically reliant on manual field scout reports; delayed advisory generation cycle (often 3-5 days); UI is complex for direct farmer use [RESEARCHED] [1.10]. Direct organizational overlap (Govt of Maharashtra) [RESEARCHED] [1.1]. Lacks real-time mobile edge AI image inference and instant multi-factor predictive forecasting for individual farmers [INFERENCE]. ResearchGate / Samagra [RESEARCHED] [1.2, 1.4, 1.10] [RESEARCHED]
IMD Meghdoot App Agromet advisories based on weather forecasts [RESEARCHED] [4.2]. Highly accurate block-level weather forecasts from IMD [RESEARCHED] [4.2, 4.4]. Broad regional text advisories; no image-based disease diagnosis; no farm-level crop phenology tracking [INFERENCE]. Weather data benchmark [INFERENCE]. Lacks farm-specific image diagnosis, trap count logging, and localized field confirmation learning loop [INFERENCE]. Journal of Agrometeorology [RESEARCHED] [4.7] [RESEARCHED]

9. Candidate Solution Approaches

Approach 1: Client-Only Computer Vision Mobile Application

  • Mechanism: Offline mobile app (Flutter + On-device TFLite model) classifying disease from leaf photos [INFERENCE].
  • Required Data: Image dataset only [INFERENCE].
  • Architecture: Mobile app with embedded ML weights [INFERENCE].
  • Complexity: Low [INFERENCE].
  • Cost: Minimal [INFERENCE].
  • Explainability: Low (black-box CNN) [INFERENCE].
  • Scalability: High client-side scalability [INFERENCE].
  • Hackathon Feasibility: Very High [INFERENCE].
  • Production Feasibility: Low (fails to meet PS requirements for weather forecasting, trap inputs, and GIS mapping) [INFERENCE].
  • Risks: High false-positive rate under bad lighting; complete lack of context (weather/stage) [INFERENCE].

Approach 2: Server-Centric Rule-Based Risk Forewarning Portal

  • Mechanism: Web backend calculating pest risk purely from weather thresholds and crop stage tables, dispatching SMS alerts [INFERENCE].
  • Required Data: Weather forecasts, crop calendar tables [INFERENCE].
  • Architecture: Relational DB + Scheduled Cron jobs + SMS Gateway [INFERENCE].
  • Complexity: Medium [INFERENCE].
  • Cost: Low [INFERENCE].
  • Explainability: High (transparent IF-THEN rules) [INFERENCE].
  • Scalability: High [INFERENCE].
  • Hackathon Feasibility: High [INFERENCE].
  • Production Feasibility: Medium (lacks visual symptom identification and ground-truth confirmation) [INFERENCE].
  • Risks: High false-alarm rate without visual/trap field verification [INFERENCE].

Approach 3: Hybrid Multi-Modal Surveillance & Advisory Platform (Recommended)

  • Mechanism: An integrated architecture combining:
    1. Offline Mobile Edge AI for initial image-based symptom classification [INFERENCE].
    2. Agronomic Risk Engine synthesizing weather, phenology, variety, soil, and trap sensor data using hybrid decision rules + XGBoost [INFERENCE].
    3. PostGIS-based Geospatial Hotspot Engine for administrative surveillance [INFERENCE].
    4. Human-in-the-Loop Expert Triage Workflow for low-confidence or high-impact diagnoses [INFERENCE].
    5. Continuous Feedback Loop retraining edge models from confirmed field cases [INFERENCE].
  • Required Data: Multi-modal (Images, Weather APIs, Trap Telemetry, Soil Maps, Historical DB) [INFERENCE].
  • Architecture: Mobile Client (Offline First) + Cloud Backend (FastAPI, PostGIS, PyTorch Server, Celery Workers) + Web GIS Dashboard [INFERENCE].
  • Complexity: High (modular, well-defined components) [INFERENCE].
  • Cost: Moderate [INFERENCE].
  • Explainability: High (AI output combined with explicit agronomic weather rule explanation) [INFERENCE].
  • Scalability: High (decoupled microservices) [INFERENCE].
  • Hackathon Feasibility: Achievable via modular division across a 4-6 student team [INFERENCE].
  • Production Feasibility: High (fully satisfies all explicit requirements of SIH26131) [INFERENCE].
  • Risks: Coordination complexity between offline sync and cloud services [INFERENCE].

10. Recommended Approach & Why

Selection: Approach 3 (Hybrid Multi-Modal Surveillance & Advisory Platform) [INFERENCE].

Why Alternatives Were Rejected

  • Approach 1 Rejection: Pure computer vision apps fail when symptoms are early or non-visual (e.g., root rot, stem borer before deadheart appears). They completely ignore explicit PS requirements for weather forecasting, trap sensor integration, and official GIS dashboards [PS, INFERENCE].
  • Approach 2 Rejection: Pure weather-rule systems miss actual localized pest occurrences and lack visual confirmation tools. They generate excessive false alerts, causing alert fatigue among farmers [INFERENCE].

System Superiority

Approach 3 fulfills every explicit requirement of problem statement SIH26131 [PS]. It provides immediate value offline to the farmer via edge AI, validates edge predictions against local micro-climate risk profiles, routes doubtful cases to experts, and aggregates field data into actionable intelligence for government officials [INFERENCE].

11. Technical Architecture

Layered Architecture Diagram

+-----------------------------------------------------------------------------------+
|                               1. USER INTERACTION LAYER                            |
|  +-------------------------------------+  +------------------------------------+  |
|  |   Farmer / Extension Mobile App     |  |   Agriculture Official & Expert    |  |
|  |   (Flutter - Offline First, PWA)    |  |   Web Portal (React.js + OpenLayers)   |  |
|  +-------------------------------------+  +------------------------------------+  |
+-----------------------------------------------------------------------------------+
                                         |  ^
                       Sync API (JSON)   |  | Push / WS Alerts
                                         v  |
+-----------------------------------------------------------------------------------+
|                               2. EDGE COMPUTING LAYER                             |
|  +-----------------------------------------------------------------------------+  |
|  |  On-Device Quantized Model (TFLite / ONNX Runtime)                         |  |
|  |  - Leaf Symptom Classifier (MobileNetV3-Small / YOLOv8-nano)                |  |
|  |  - Local SQLite DB (Offline cache, pending sync queue)                     |  |
|  +-----------------------------------------------------------------------------+  |
+-----------------------------------------------------------------------------------+
                                         |
                                  HTTPS / TLS 1.3
                                         v
+-----------------------------------------------------------------------------------+
|                               3. API & GATEWAY LAYER                              |
|  +-----------------------------------------------------------------------------+  |
|  |  API Gateway (FastAPI / NGINX Router)                                      |  |
|  |  - OAuth2 / JWT Auth  | Rate Limiting  | Multi-lingual Response Serializer  |  |
|  +-----------------------------------------------------------------------------+  |
+-----------------------------------------------------------------------------------+
                                         |
         +-------------------------------+-------------------------------+
         |                               |                               |
         v                               v                               v
+------------------+   +----------------------------------+   +---------------------+
| 4. AI & FORECAST |   |   5. GEOSPATIAL & HOTSPOT ENGINE |   | 6. EXPERT TRIAGE    |
|    SERVICES      |   |   (PostGIS + Celery)             |   |    & WORKFLOW       |
|                  |   |                                  |   |                     |
| - High-Res Swin  |   | - Spatial Clustering (DBSCAN)    |   | - KVK Expert Queue  |
|   Vision Server  |   | - Outbreak Heatmap Aggregator    |   | - Diagnostic Lab    |
| - Weather Risk   |   | - Tehsil/District Boundary Map   |   |   Ticket Dispatch   |
|   XGBoost Engine |   |                                  |   | - Active Learning   |
| - Voice TTS/STT  |   |                                  |   |   Dataset Staging   |
+------------------+   +----------------------------------+   +---------------------+
         |                               |                               |
         +-------------------------------+-------------------------------+
                                         |
                                         v
+-----------------------------------------------------------------------------------+
|                               7. DATA STORAGE LAYER                               |
|  +---------------------------+  +---------------------------+  +---------------+  |
|  | PostgreSQL + PostGIS      |  | Redis Cache               |  | Object Store  |  |
|  | (Primary Relational & Spatial) | (Weather & Active Sessions) | (MinIO / S3)  |  |
|  +---------------------------+  +---------------------------+  +---------------+  |
+-----------------------------------------------------------------------------------+

Detailed Data Flow Narrative

  1. Field Capture & Edge Inference: A farmer or Krishi Sahayak selects crop type, crop stage, and takes a leaf/pest photo. The local TFLite model runs inference on the mobile GPU/NPU in under 300ms [INFERENCE].
  2. Confidence Check & Local Storage:
    • If confidence $\ge 85%$ and the diagnosis aligns with local weather risk thresholds, the app displays the Marathi advisory immediately (offline) [INFERENCE].
    • If confidence $< 85%$ or the pest is a high-risk quarantine species, the app marks the item for "Expert Verification" [INFERENCE].
  3. Background Synchronization: Once internet connectivity (2G/3G/Wi-Fi) is detected, the app syncs geotagged image payloads, trap counts, and farm logs to the API Gateway via encrypted compressed JSON [INFERENCE].
  4. Multi-Factor Risk Assessment: The backend fetches weather forecast data (Open-Meteo), retrieves the crop stage and variety vulnerability, and runs the XGBoost risk forecaster [INFERENCE].
  5. Geospatial Hotspot Processing: A background Celery worker executes DBSCAN spatial clustering over the incoming geotagged reports. If a cluster of $>5$ similar pest cases is detected within a 2km radius inside 48 hours, a "Hotspot Warning" is automatically raised [INFERENCE].
  6. Expert Triage & Laboratory Escalation: Low-confidence scans land in the KVK Agronomist Web Portal. The scientist reviews high-resolution images alongside historical plot weather data, updates the diagnosis, and clicks "Approve Advisory" [INFERENCE]. If necessary, a diagnostic lab ticket with a QR code is issued to the farmer [INFERENCE].
  7. Farmer Advisory & Continuous Learning: The confirmed diagnosis updates the farmer's mobile app via push notification/SMS in Marathi audio/text [INFERENCE]. The validated image and metadata are committed to the ML Retraining Staging Bucket [INFERENCE].

12. Technology Stack

Layer Selected Technology Why Alternative Why Rejected
Mobile Frontend Flutter (Dart) Single codebase for Android/PWA; excellent offline sqlite/hive support; native TFLite bindings [INFERENCE]. React Native Slower image processing bridge; larger binary footprint [INFERENCE].
Edge AI Runtime TensorFlow Lite / ONNX Runtime Low latency on budget ARM chips; model quantization support (<10MB) [INFERENCE]. PyTorch Mobile Less optimized for low-end Android hardware [INFERENCE].
Backend Framework Python FastAPI Native async performance; seamless integration with PyTorch, XGBoost, and GIS libraries [INFERENCE]. Node.js (Express) WEAK ML ecosystem integration; extra IPC latency [INFERENCE].
Database & GIS PostgreSQL 16 + PostGIS 3.4 Industry standard spatial queries (ST_ClusterDBSCAN, ST_DWithin); bulletproof transaction consistency [KNOWN]. MongoDB + GeoNear Lacks robust spatial query operations and ACID compliance for lab ticket workflows [INFERENCE].
Cloud AI Server PyTorch + TorchVision (Swin Transformer / YOLOv8) SOTA accuracy for complex fine-grained agricultural disease classification [INFERENCE]. TensorFlow/Keras Slower model iteration and lower availability of fine-grained pre-trained vision transformer backbones [INFERENCE].
Caching & Queue Redis + Celery High-throughput asynchronous background job execution (hotspot calculation, SMS dispatch) [INFERENCE]. RabbitMQ Redis serves dual purpose as cache and message broker, reducing infra complexity [INFERENCE].
Object Storage MinIO (Self-hosted S3 API) S3-compatible open-source object storage for image datasets and model checkpoints [INFERENCE]. AWS S3 Requires paid cloud subscription during development phase [INFERENCE].
Multilingual Voice AI4Bharat Indic-TTS / Google Cloud Speech API Superior Marathi text-to-speech pronunciation quality for rural agricultural terms [RESEARCHED]. eSpeak / native Android TTS Robotic tone; fails on local Maharashtrian agricultural vocabulary [INFERENCE].

13. AI/ML Justification

1. Leaf Symptom Classifier

  • Purpose: Instant visual identification of disease/pest symptoms from leaf/stem/fruit photographs [PS].
  • Model Type: Dual-tier vision system: Quantized MobileNetV3-Small on Edge; Swin Transformer Small on Server [INFERENCE].
  • Input: RGB Image ($224 \times 224$ edge / $384 \times 384$ cloud) [INFERENCE].
  • Output: Class probabilities across top 30 crop-disease pairs + Bounding boxes for symptoms [INFERENCE].
  • Data Required: 25,000+ annotated images (PlantVillage, Kaggle Indian Agriculture subsets) [RESEARCHED].
  • Why AI: Manual diagnostic knowledge takes years of agronomic training; visual AI grants instant expertise to rural farmers [INFERENCE].
  • Evaluation Metric: Top-1 Accuracy ($\ge 90%$), Macro F1-Score ($\ge 0.88$) [INFERENCE].
  • Failure Mode: Blurry images, low light, or complex mixed infections. Mitigation: Blur detector prompts re-take; low confidence triggers KVK expert triage [INFERENCE].
  • Human Oversight: High. Scans $<85%$ confidence are auto-escalated to human agronomists [INFERENCE].

2. Multi-Factor Agronomic Risk Forecasting Engine

  • Purpose: Forecast pest outbreak likelihood 3–7 days before visible symptoms spread [PS].
  • Model Type: Hybrid Rule-Based Decision Logic + Gradient Boosted Decision Trees (XGBoost) [INFERENCE].
  • Input: 7-day weather forecast (Temp, RH, Rainfall), Crop Phenology Stage (days post-sowing), Variety Susceptibility Index (0–1), Soil Moisture Stress Index, Historical Tehsil Infestation Index [PS, INFERENCE].
  • Output: Outbreak Risk Level (Low / Medium / High / Severe) + Key Driving Weather Triggers [INFERENCE].
  • Data Required: Historical weather-pest correlation datasets (ICAR-NCIPM rulebooks) [RESEARCHED] [3.8, 4.7].
  • Why AI/ML: Biological pest development exhibits non-linear interactions between humidity, temperature, and crop stage that simple threshold rules cannot fully capture [INFERENCE].
  • Evaluation Metric: Receiver Operating Characteristic - Area Under Curve (ROC-AUC $\ge 0.85$) [INFERENCE].
  • Failure Mode: Inaccurate weather forecast inputs. Mitigation: Fall back to deterministic ICAR "thumb rules" if forecast variance is high [RESEARCHED] [4.7].
  • Human Oversight: District agronomists receive forewarning summaries prior to mass farmer broadcast [INFERENCE].

3. Non-AI System Components

  • Dosage & Water Volume Calculator: Pure deterministic math based on CIB&RC label claims ($\text{Dose/Acre} \times \text{Plot Area}$) [RESEARCHED] [3.8].
  • Geospatial Hotspot Clustering: Deterministic spatial query (ST_ClusterDBSCAN) [INFERENCE].
  • Referral Ticket Routing: Database relational lookup by proximity to nearest registered lab/KVK [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

Database Schema (Core Tables)

-- Schema Overview
CREATE TABLE farmers (
    farmer_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    agristack_id VARCHAR(50) UNIQUE,
    full_name VARCHAR(100) NOT NULL,
    phone_number VARCHAR(15) NOT NULL,
    village_code VARCHAR(20) NOT NULL,
    tehsil VARCHAR(50) NOT NULL,
    district VARCHAR(50) NOT NULL,
    preferred_language VARCHAR(10) DEFAULT 'mr',
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

CREATE TABLE farm_plots (
    plot_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    farmer_id UUID REFERENCES farmers(farmer_id),
    crop_type VARCHAR(50) NOT NULL,
    crop_variety VARCHAR(50),
    sowing_date DATE NOT NULL,
    area_acres NUMERIC(5,2) NOT NULL,
    soil_type VARCHAR(30),
    location GEOMETRY(Point, 4326) NOT NULL,
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

CREATE TABLE crop_diagnoses (
    diagnosis_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    plot_id UUID REFERENCES farm_plots(plot_id),
    image_url TEXT,
    edge_predicted_disease VARCHAR(100),
    edge_confidence NUMERIC(4,3),
    final_disease VARCHAR(100),
    verified_by_expert BOOLEAN DEFAULT FALSE,
    expert_id UUID,
    status VARCHAR(20) DEFAULT 'PENDING', -- PENDING, VERIFIED, ESCALATED
    location GEOMETRY(Point, 4326) NOT NULL,
    captured_at TIMESTAMP WITH TIME ZONE NOT NULL,
    synced_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

CREATE TABLE trap_logs (
    trap_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    plot_id UUID REFERENCES farm_plots(plot_id),
    trap_type VARCHAR(30) NOT NULL, -- PHEROMONE, STICKY, IOT_CAMERA
    pest_count INT NOT NULL,
    logged_at TIMESTAMP WITH TIME ZONE NOT NULL
);

CREATE TABLE advisories (
    advisory_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    diagnosis_id UUID REFERENCES crop_diagnoses(diagnosis_id),
    disease_name VARCHAR(100) NOT NULL,
    chemical_recommendation TEXT,
    biological_recommendation TEXT,
    dosage_per_acre VARCHAR(100),
    phi_days INT,
    audio_url_marathi TEXT,
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

Offline & Synchronization Architecture

[ Mobile SQLite DB ]  <--->  [ Sync Engine (WatermelonDB/Hive) ]
                                            |
                              (Network Detector: Ping Server)
                                            |
                   +------------------------+------------------------+
                   | (Online)                                        | (Offline)
                   v                                                 v
   [ Compression & Payload Queue ]                    [ Append to Local Storage Queue ]
   - Images compressed to WebP (50KB)                 - Store encrypted JSON + Image
   - Vector Clock / Timestamp appended                - Show offline cached advisory
                   |
                   v
   [ POST /api/v1/sync/batch ]
                   |
                   v
   [ Server Conflict Resolver ]
   - Server timestamp wins for status updates
   - Client log appended for new field scans
  • Bandwidth Reduction Strategy: Photos captured on mobile are downsampled to $1024 \times 1024$, converted to WebP format at $80%$ quality (average file size reduces from 4MB to ~60KB) before transmission [INFERENCE].
  • Conflict Handling: Immutable event-sourcing pattern for diagnosis logs; updates from expert validation override client edge predictions using server vector clocks [INFERENCE].

15. Security, Privacy & Governance

  • Personal & Location Data Protection: Farmer plot locations are precise to sub-meter accuracy in private tables for targeted advisories [INFERENCE]. On public GIS administrative dashboards, plot coordinates are obfuscated using random Gaussian noise (jittered within a 500-meter radius) to protect farmer identity and land value [INFERENCE].
  • Authentication & Authorization:
    • Farmers log in via OTP backed by AgriStack Farmer ID or phone number [INFERENCE].
    • Role-Based Access Control (RBAC): Roles defined for FARMER, EXTENSION_WORKER, KVK_EXPERT, and STATE_ADMIN [INFERENCE].
  • Data Encryption: All data in transit uses TLS 1.3 encryption [KNOWN]. SQLite databases on mobile devices and PostgreSQL tables in the cloud are encrypted at rest using AES-256 [KNOWN].
  • Regulatory Validation & Pesticide Governance: Chemical recommendations strictly match Central Insecticides Board and Registration Committee (CIB&RC) label claims [RESEARCHED] [3.8]. The system automatically suppresses restricted chemical suggestions for organic-certified plots or within mandatory Pre-Harvest Intervals (PHI) [INFERENCE].
  • Disclaimer & Liability: Every advisory carries a standard state-approved legal disclaimer: "Advisories are generated based on AI predictions and micro-climate models. For severe outbreaks, verify with local Krishi Sahayak before mass spraying." [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope (MUST BUILD & SHOULD BUILD)

  • Mobile App (Flutter):
    • Offline leaf image capture & TFLite inference for 10 major pests/diseases affecting Soybean, Cotton, Paddy, and Chickpea [INFERENCE].
    • Manual Pheromone Trap logging UI [PS, INFERENCE].
    • Marathi UI with Text-to-Speech audio output for advisories [PS, INFERENCE].
    • Local SQLite offline caching & background sync manager [INFERENCE].
    • Chemical dosage calculator based on field size [PS, INFERENCE].
  • Backend & Cloud Services (FastAPI + PostgreSQL/PostGIS):
    • REST APIs for authentication, diagnosis sync, and risk calculation [INFERENCE].
    • Weather forecasting risk engine integrating Open-Meteo API [PS, INFERENCE].
    • DBSCAN spatial clustering worker generating disease hotspots [PS, INFERENCE].
  • Web Dashboards (React + OpenLayers):
    • KVK Expert Portal: Triage queue for low-confidence image verification [PS, INFERENCE].
    • Official GIS Dashboard: Interactive Maharashtra map displaying real-time disease intensity, hotspot clusters, and extension ticket status [PS, INFERENCE].

Described Only / Out of Scope (NICE TO HAVE & DO NOT BUILD)

  • Hardware IoT Pheromone Trap: Physical hardware trap fabrication (simulated via API during demo) [ASSUMPTION].
  • Live Live Integration with State AgriStack Production Servers: Direct API integration with locked government databases (simulated using AgriStack schema) [NEEDS VALIDATION].
  • Drone Autonomous Spraying Dispatch: Third-party drone API integration [INFERENCE].

17. Live Demo Simulation Plan

To present a seamless, end-to-end demonstration to SIH judges without relying on live field outbreaks or restricted state networks, the team will run a multi-device live simulation:

[ Device 1: Android Phone (Farmer) ]
  - Turn ON Airplane Mode (Offline)
  - Scan physical leaf sample with early Pink Bollworm damage
  - TFLite detects symptom -> Displays local Marathi advisory + CIB&RC dosage
  - Log 12 moths in manual Pheromone Trap form
  - Turn OFF Airplane Mode -> App syncs background queue

[ Device 2: Admin Laptop (KVK Expert Portal) ]
  - Simulated ambiguous image triggers "Pending Expert Triage" alert
  - Expert views high-res image, plot weather history, updates diagnosis to "Spodoptera litura"
  - Clicks "Send Advisory" -> Push notification arrives instantly on Device 1

[ Device 3: Large Display (Official GIS Map) ]
  - Trigger "Simulate Monsoon Pest Surge" script (injects 50 geotagged field reports in Latur district)
  - PostGIS DBSCAN worker calculates hotspot within 3 seconds
  - Hotspot polygon turns RED on Latur Tehsil map
  - Admin clicks "Dispatch Emergency Advisory" -> Mass SMS dispatch logged

Never present simulated data as live government APIs. Explicitly state: "Using simulated AgriStack records and Open-Meteo weather feeds formatted to official state data schemas." [INFERENCE].

18. Implementation Roadmap

Target Deadline: 30 September 2026 [PS]

Phase 1: Foundation & Data Preparation (Weeks 1 - 3) [By June 20, 2026]
- Benchmark PlantVillage & Indian Crop Disease datasets (Cotton, Soybean, Paddy, Chickpea).
- Train PyTorch Swin Transformer & quantize MobileNetV3 TFLite models.
- Set up PostgreSQL + PostGIS database schema and FastAPI scaffolding.

Phase 2: Mobile App & Risk Engine Core (Weeks 4 - 7) [By July 20, 2026]
- Build Flutter offline UI with embedded TFLite inference engine.
- Implement WatermelonDB/SQLite offline sync manager.
- Develop multi-factor XGBoost/Rule risk engine combining Open-Meteo & crop stages.

Phase 3: Web Dashboards & Spatial Engine (Weeks 8 - 10) [By August 15, 2026]
- Build React.js KVK Expert Triage Portal and GIS Map Dashboard with OpenLayers.
- Implement DBSCAN spatial clustering backend workers in Celery/Redis.
- Integrate AI4Bharat / Google Marathi Text-to-Speech audio advisories.

Phase 4: Integration, Simulation & Testing (Weeks 11 - 13) [By September 10, 2026]
- Execute end-to-end integration tests (Offline edge capture -> Sync -> Expert validation -> GIS Hotspot).
- Perform field testing with agricultural students/extension staff in local conditions.
- Run false-positive and model performance stress tests under poor lighting.

Phase 5: Refinement & Demo Polish (Weeks 14 - 15) [By September 30, 2026]
- Lock MVP scope, generate demo dataset scripts, and rehearse live judge presentation story.

19. Testing & Validation Plan

Test Category Target Metric / Scenario Validation Methodology
Edge AI Performance Latency $<400\text{ms}$ on Android device; Model size $<12\text{MB}$ [INFERENCE]. Benchmark TFLite inference runtime on budget Android phone (2GB RAM) [INFERENCE].
Classification Accuracy Top-1 Accuracy $\ge 90%$, F1-Score $\ge 0.88$ across 10 crop classes [INFERENCE]. Hold-out test set evaluation on 2,500 real field images captured in field lighting [INFERENCE].
False-Positive Resilience $<5%$ false alerts on healthy leaves or non-disease spots [INFERENCE]. Test suite of 500 healthy leaf images and background foliage noise [INFERENCE].
Offline Sync Reliability $100%$ eventual consistency without duplicate records after reconnecting [INFERENCE]. Disconnect network mid-capture, create 20 logs, re-enable network, verify PostGIS DB sync [INFERENCE].
GIS Hotspot Clustering Correct spatial polygon generation within $<3$ seconds for 1,000 geotagged points [INFERENCE]. Benchmark ST_ClusterDBSCAN execution time on PostgreSQL [INFERENCE].
Usability (SUS Score) System Usability Scale (SUS) score $>80$ among rural non-tech users [INFERENCE]. User testing with 10 non-tech users evaluating Marathi voice/icon navigation [INFERENCE].

20. Risk Register

Risk Likelihood Impact Severity Mitigation Strategy
Poor Field Camera Quality / Blurry Leaf Images High High High Implement an automated real-time image quality estimator (Laplacian variance blur detector) in the app that prompts the user to retake the photo before running inference [INFERENCE].
Lack of Live Data Connection in Farmlands High Medium High Offline-first architecture storing models, crop diagnostic trees, and local database on-device; background sync executes automatically when connectivity returns [INFERENCE].
High False-Positive Rate from Pure Vision AI Medium High High Multi-factor validation: Vision model output is cross-referenced with weather risk profiles and crop phenology; low-confidence cases auto-route to KVK experts [INFERENCE].
Farmer Distrust in AI Advisory Medium High Medium Provide transparent explanation (e.g., "High risk due to 88% humidity over last 3 days"), Marathi voice translation, and expert verification badge [INFERENCE].
AgriStack API Access Restrictions during Hackathon High Low Low Build an open schema wrapper conforming to published AgriStack standards; run demo using simulated AgriStack IDs [ASSUMPTION].

21. Cost & Scalability

Hackathon Phase (Cost: ~₹0 / Free Tier)

  • Mobile App: Built using Flutter (Open Source) [KNOWN].
  • Edge Inference: Runs on user smartphone hardware [INFERENCE].
  • Backend Server: Hosted on local developer laptops or free-tier cloud instances (e.g., Render / Supabase / GCP Free Tier) [KNOWN].
  • Weather Data: Open-Meteo Free API Tier (10,000 calls/day) [RESEARCHED].
  • Map Tiles: OpenStreetMap / CartoDB free tiles [KNOWN].

Production Deployment (State-Wide Scale: 10 Million Farmers)

  • Cloud Infrastructure: Kubernetes (GKE / AWS EKS) with auto-scaling API pods [INFERENCE].
  • Database: Managed PostgreSQL (AWS Aurora / GCP Cloud SQL with PostGIS extension) with read replicas [INFERENCE].
  • Object Storage: AWS S3 / Cloudflare R2 for petabyte-scale image archival [INFERENCE].
  • SMS & Voice Gateway: Telecom integration (e.g., CDAC MSDG / Kaleyra) for localized Marathi SMS and IVR voice calls [RESEARCHED].
  • Cost Efficiency: Edge AI offloads $90%$ of inference compute costs from cloud servers to client smartphones, keeping cloud operational costs under ₹2 per farmer annually [INFERENCE].

22. Differentiation

Parameter Standard / Generic Submissions CropHealth-AI (Our Proposed Solution)
Diagnostic Scope Pure leaf-photo CNN classifier app [INFERENCE]. Multi-Modal Diagnostics: Integrates edge vision, trap sensor counts, micro-climate weather risk, soil, and crop variety susceptibility [PS, INFERENCE].
Risk Management Reactive (detects disease only after visible spots appear) [PS, INFERENCE]. Proactive Forewarning: Predicts pest outbreaks 3–7 days before physical symptoms manifest using multi-factor weather models [PS, INFERENCE].
Connectivity Design Cloud-dependent (fails without internet) [RESEARCHED] [2.5]. Offline-First Edge Architecture: Quantized TFLite engine processes scans directly on low-cost phones without internet [INFERENCE].
Expert Integration Automated AI output presented as infallible [INFERENCE]. Human-in-the-Loop Triage: Auto-escalates ambiguous cases to KVK agronomists, learning continuously from verified field confirmations [PS, INFERENCE].
Government Utility Single-user app without administrative view [INFERENCE]. State-Wide GIS Command Center: PostGIS spatial clustering maps disease hotspots for targeted pesticide/extension deployment [PS, INFERENCE].

23. Future Scope

  1. AgriDrone Fleet Integration: Automated dispatch of geotagged hotspot coordinates to local drone service hubs for targeted ultra-low volume pesticide spraying [INFERENCE].
  2. Autonomous IoT Trap Mesh: Solar-powered optical pheromone traps sending automated hourly insect counts via LoRA/NB-IoT networks [INFERENCE].
  3. AgriStack Deep Native Integration: Direct single-sign-on (SSO) and automated crop registry sync once government production APIs are opened to third-party developers [RESEARCHED] [2.1, 2.3].
  4. Spectral Satellite Stress Monitoring: Integration of Sentinel-2 multi-temporal vegetation indices (NDVI, NDWI) to detect large-scale crop stress before field scouting [RESEARCHED] [2.2].

24. Assumptions & Open Items Log

Consolidated Evidence Tags

Assumptions [ASSUMPTION]

  1. [ASSUMPTION] Synthetic agronomic tables based on ICAR crop calendars are sufficient to simulate crop phenology during the hackathon demo.
  2. [ASSUMPTION] Mock trap count generators accurately reflect regional pest population surges for demo purposes.
  3. [ASSUMPTION] Historical district-level disease incidence CSVs can stand in for CROPSAP's internal 15-year historical GIS database during development.
  4. [ASSUMPTION] Hardware IoT traps are simulated via REST API JSON payloads during the live hackathon presentation.
  5. [ASSUMPTION] Simulated AgriStack Farmer IDs and plot boundaries conform to published IndEA 2.0 governance standards.

Items Needing Validation [NEEDS VALIDATION]

  1. [NEEDS VALIDATION] Production API endpoints, authentication keys, and rate limits for official Maharashtra State AgriStack / Digital Agriculture Mission servers.
  2. [NEEDS VALIDATION] Access policy and data sharing agreement protocols with KVK scientists for real-time human-in-the-loop expert triage queuing.
  3. [NEEDS VALIDATION] Exact CIB&RC chemical pesticide formulation update feeds for automated synchronization with the dosage calculator backend.

Final Recommendation

Build CropHealth-AI using Approach 3 (Hybrid Multi-Modal Surveillance Platform). It is technically feasible, satisfies every requirement of PS SIH26131, solves real offline rural field constraints, and delivers immediate value to farmers, extension officers, and state officials [INFERENCE].

Recommended Stack

  • Mobile: Flutter + TensorFlow Lite (Quantized MobileNetV3) + SQLite [INFERENCE].
  • Backend: Python FastAPI + PyTorch (Swin Transformer) + XGBoost + Celery + Redis [INFERENCE].
  • Database: PostgreSQL 16 + PostGIS 3.4 [INFERENCE].
  • Web Dashboards: React.js + OpenLayers + Tailwind CSS [INFERENCE].
  • External APIs: Open-Meteo Weather API + AI4Bharat Marathi TTS [RESEARCHED].

Biggest Dependency

Availability of well-annotated, high-quality leaf disease image datasets covering Indian crop varieties (Cotton, Soybean, Paddy, Chickpea) under varied natural field lighting conditions [RESEARCHED] [2.5].

Biggest Risk

High false-positive rates from computer vision classifiers when leaf images are taken under harsh sunlight, motion blur, or dirty camera lenses [RESEARCHED] [2.5]. Mitigation: On-device Laplacian blur filter + multi-factor weather verification + human agronomist escalation [INFERENCE].

Biggest Technical Challenge

Implementing seamless offline-first bidirectional database synchronization between local mobile SQLite stores and cloud PostGIS servers without data corruption or duplicate records [INFERENCE].

First 5 Actions

  1. Download and curate PlantVillage and Kaggle Indian crop disease image subsets for Cotton, Soybean, Paddy, and Chickpea [RESEARCHED].
  2. Train baseline PyTorch vision classifier and quantize to TFLite format (<12MB) [INFERENCE].
  3. Set up PostgreSQL + PostGIS local database container with core tables (farmers, farm_plots, crop_diagnoses, advisories) [INFERENCE].
  4. Build Flutter mobile skeleton app with camera integration and embedded TFLite inference engine [INFERENCE].
  5. Create mock Open-Meteo weather risk rule script for key Maharashtra pests (e.g., Pink Bollworm humidity thresholds) [RESEARCHED] [1.5, 4.7].

Validation Before Coding

Verify that the quantized TFLite model achieves $<400\text{ms}$ inference latency and $>85%$ top-1 accuracy on a low-spec 2GB RAM budget Android test device [INFERENCE].

Buildability Score /100

92 / 100 [INFERENCE] (High score due to modular stack architecture, availability of public open-source computer vision datasets, open weather APIs, and standard Flutter/FastAPI/PostGIS toolchains).

Confidence

High [INFERENCE] (The proposed solution strictly adheres to the source of truth JSON requirements, leverages well-documented software architectures, and addresses real rural deployment constraints).

Research Sources

  1. CROPSAP Dashboard & Project Documentation
    • URL: http://98.70.59.108/ [1.1]
    • Source Type: Official Government / System Portal [1.1]
    • Claim Supported: Crop Pest Surveillance & Advisory Project (CROPSAP) operates in Maharashtra for state-wide pest tracking [1.1, 1.2, 1.11].
    • Reliability: High [1.1].
  2. Rethinking Pest Surveillance in Maharashtra (CROPSAP 2.0)
  3. AgriStack: India's Digital Public Infrastructure for Agriculture
    • URL: https://agristack.gov.in/ [2.1]
    • Source Type: Official Government Portal (Ministry of Agriculture & Farmers Welfare) [2.1, 2.8]
    • Claim Supported: AgriStack provides federated Farmer Registry, Geo-Referenced Village Maps, and Crop Sown Registry across states including Maharashtra [2.1, 2.3, 2.8].
    • Reliability: High [2.1].
  4. AI Crop Disease Detection for Indian Farmers
  5. Insecticide and Fungicide Calculators (IFC) - ICAR-NCIPM
    • URL: https://icar.org.in/en/node/10817 [3.8]
    • Source Type: Official Academic / Research Institute (ICAR) [3.8]
    • Claim Supported: ICAR-NCIPM provides validated rulebooks and calculators for pesticide dosages, label claims, and application technologies across 12 major crops [3.8].
    • Reliability: High [3.8].
  6. National Pest Surveillance System (NPSS) Awareness & Rollout
    • URL: https://www.pib.gov.in/PressReleasePage.aspx?PRID=2233738 [3.9]
    • Source Type: Press Information Bureau (PIB) / Government Release [3.9]
    • Claim Supported: NPSS is an AI/ML-driven national pest surveillance framework deployed across states under DPPQ&S and ICAR [3.9].
    • Reliability: High [3.9].
  7. Development of DSS for Forewarning of Pests and Diseases
    • URL: https://journal.agrimetassociation.org/index.php/jam/article/view/3144 [4.7]
    • Source Type: Peer-reviewed Academic Publication (Journal of Agrometeorology) [4.7]
    • Claim Supported: Weather-based thumb rules and heuristic decision models forecast crop pest/disease incidence with 33% to 100% accuracy depending on season and crop [4.7].
    • Reliability: High [4.7].
02 / What our

DEVELOPERS SAY

" 30M+ Google Gemini tokens went into researching, analyzing, and understanding all 233 SIH problem statements to make them easier to explore and solve. "
Satyam Kumar
Satyam Kumar
CEO , Full Stack Developer
" The platform makes SIH problem statements much easier to understand. Having everything from the problem details to the procedure in one place. "
Rishabh Ranjan
Rishabh Ranjan
COO , Content Writer
" Oh, you got ads? Well, we need to make money too! But, you can use our platform to help you win SIH and take home the prize money. "
Himanshu Kumar
Himanshu Kumar
CMO , Product Reviewer
03 / Are You

READY TO
BUILD?

Get in touch if you need help bringing your ideas to life, or simply have an interesting idea you'd like to talk through. I'm always open to a conversation and exploring new possibilities together.

LinkedIn
[email protected]