---
title: Crop Adv
emoji: 🌍
colorFrom: blue
colorTo: yellow
sdk: gradio
sdk_version: 6.8.0
app_file: app.py
pinned: false
license: mit
---

# 🌾 Maharashtra Crop Recommendation System

An AI-powered crop advisory tool that recommends the most suitable crops for a Maharashtra farmer based on their **soil chemistry**, **local weather**, **district & season**, **water supply**, and **budget** — and explains *why* each crop is recommended in plain language.

---

## 🧠 How It Works: The Full Pipeline

### Step 1 — Two Models Run in Parallel

When you click **Find Best Crops For Me**, two independent models are run simultaneously:

#### Model 1 — Soil & Climate AI (Random Forest)
- Takes 7 numerical inputs: N, P, K, Temperature, Humidity, pH, Rainfall
- Runs them through a pre-trained **Random Forest Classifier** (`model1_npk.pkl`)
- Outputs a **probability score (0–1)** for every crop it knows (22 crops)
- Example: `{"rice": 0.42, "maize": 0.28, "cotton": 0.08, ...}`
- The label encoder (`model1_label_encoder.pkl`) maps numeric class indices → crop names
- **Training:** 2,200 samples × 22 crops, 99.6% cross-validation accuracy (5-fold stratified)

#### Model 2 — Regional History Scoring (Weighted Formula)
- Looks up the selected **District + Season** combination in `model2_full_scored.csv`
- Returns a pre-computed **suitability score** for each crop based on 26 years (1997–2022) of Maharashtra district-level crop production data
- **No ML here** — this is a transparent, rule-based scoring formula:

```
Suitability Score =
  0.35 × Crop Frequency %          (how often this crop is grown here)
+ 0.25 × Historically Grown Flag   (grown in ≥50% of years = 1, else 0)
+ 0.20 × Normalised Avg Yield      (yield relative to best crop in district)
+ 0.10 × Dominance Rate            (share of total district area)
+ 0.05 × Recent Years Grown        (grown post-2015 = recency bonus)
+ 0.05 × Yield Stability           (lower variance = higher score)
```

- Covers **35 districts × 4 seasons × 33 crops = 1,603 scored combinations**

---

### Step 2 — Scores Are Combined

```
Final Score = 0.60 × Model1_probability + 0.40 × Model2_suitability_score
```

- Model 2 scores are **normalised** to 0–1 before combining (divide by max score in that district/season)
- Crops are then **ranked highest → lowest** by Final Score
- Only crops present in `CROP_META` (hardcoded dictionary, see below) are kept

---

### Step 3 — Budget & Land Filtering

For each ranked crop:

| Condition | Result |
|---|---|
| `input_cost × land_ha ≤ budget` | `budget_status = "within"` — shown fully |
| `input_cost × land_ha > budget` | `affordable_ha = budget / input_cost` — shown with partial land warning |
| `affordable_ha < 0.1 ha` | Crop is **skipped entirely** |

---

### Step 4 — Risk Scoring (4-Dimensional)

Every crop gets a risk score from 0–100 across four dimensions, then averaged equally:

| Dimension | Weight | How It's Calculated |
|---|---|---|
| Weather Risk | 25% | Penalises extreme temp (<8°C or >42°C), extreme humidity, and rainfall mismatches vs crop water need |
| Market Risk | 25% | Based on hardcoded demand: High=20, Medium=40, Low=65 |
| Budget Risk | 25% | Ratio of `total_cost / budget_max` mapped to risk bands |
| Water Risk | 25% | Gap between crop's irrigation need vs farmer's available supply |

**Irrigation need levels** (hardcoded per crop):
- `"high"` — Rice, Sugarcane, Banana, Grapes, Jute, Apple
- `"medium"` — Cotton, Maize, Wheat, Onion, Tomato, etc.
- `"low"` — Chickpea, Bajra, Jowar, Pulses, Oilseeds, etc.

**Irrigation supply levels** (farmer input):
- Assured (Canal / River) = 3
- Borewell / Pump = 2
- Seasonal / Rain-fed = 1
- No Irrigation (Dryland) = 0

**Risk label thresholds:**
- Score < 35 → 🟢 **Low**
- Score 35–64 → 🟡 **Medium**
- Score ≥ 65 → 🔴 **High**

---

### Step 5 — Confidence Labelling

Each result is tagged with an AI confidence level based on **Model 1's raw probability**:

| Model 1 Score | Label | Meaning |
|---|---|---|
| ≥ 35% | High confidence | Soil & climate strongly match |
| 12–34% | Moderate confidence | Partial soil match, regional history confirms |
| < 12% | Based on regional history | Driven primarily by Model 2 |

---

### Step 6 — "Why We Recommend This" Explanation

Every card generates **5 plain-language bullet points** dynamically from real data — nothing is templated or hardcoded:

| Bullet | Data Source |
|---|---|
| Soil fit | `m1_score` from Model 1 probability |
| Season fit | `season_match` boolean (best_season == selected season) |
| Water fit | Irrigation gap calculation (need − available) |
| Risk summary | `risk` label + highest `risk_breakdown` dimension |

---

## 📥 Inputs

### Location & Season

| Input | Type | Options |
|---|---|---|
| District | Dropdown | 35 Maharashtra districts |
| Season | Dropdown | Kharif (Jun–Sep), Rabi (Oct–Mar), Summer (Apr–Jun), Whole Year |

### Farm Details

| Input | Type | Range |
|---|---|---|
| Land (Hectares) | Number | 0.1 – 100 |
| Total Budget (₹) | Number | ₹1,000+ |
| Water source | Radio | Assured / Borewell / Seasonal / Dryland |

### Soil Test Results (from Krishi Vigyan Kendra or soil lab)

| Input | Range | Unit |
|---|---|---|
| Nitrogen (N) | 0 – 300 | mg/kg |
| Phosphorus (P) | 0 – 300 | mg/kg |
| Potassium (K) | 0 – 300 | mg/kg |
| Soil pH | 3.5 – 9.5 | — |

### Local Weather

| Input | Range | Unit |
|---|---|---|
| Avg Temperature | 5 – 50 | °C |
| Avg Humidity | 10 – 100 | % |
| Annual Rainfall | 20 – 3000 | mm/year |

---

## 📤 Output: What Each Card Shows

| Element | Source | Notes |
|---|---|---|
| Crop name + medal | Combined score ranking | 🥇🥈🥉 for top 3 |
| Best season | `CROP_META` hardcoded | e.g. "Kharif", "Rabi", "Whole Year" |
| Days to harvest | `CROP_META` hardcoded | Approximate growing period |
| Budget badge | Budget filter logic | ✅ Fits / ⚠️ Partial |
| Risk badge | Risk score calculation | Low / Medium / High |
| "Why we recommend this" | Live data from both models | 4 dynamic bullet points |
| Irrigation verdict | Irrigation gap formula | ✅ / ⚠️ / ❌ |
| Cost overview | `CROP_META` × land_ha | Input cost per ha and total |
| Market demand | `CROP_META` hardcoded | High / Medium / Low |
| Risk breakdown bars | 4-dimensional risk formula | 0–100 per dimension |
| Model confidence | Model 1 probability | High / Moderate / Regional |
| Score breakdown | Both model scores | Soil %, Regional %, Combined % |

---

## 🗄️ What Is Hardcoded (`CROP_META`)

The following per-crop values are **static estimates** based on Maharashtra Agriculture Department averages. They are not fetched from any live source:

| Field | Description | Example (Rice) |
|---|---|---|
| `input_cost` | Estimated cost to grow 1 hectare (₹) | ₹30,000 |
| `yield_qtl` | Expected yield in quintals per hectare | 25 qtl/ha |
| `demand` | Market demand category | "High" |
| `days` | Days from sowing to harvest | 120 days |
| `best_season` | Ideal growing season | "Kharif" |

**Full crop list in CROP_META (50 crops):**
Rice, Maize, Chickpea, Kidneybeans, Pigeonpeas, Mothbeans, Mungbean, Blackgram, Lentil, Pomegranate, Banana, Mango, Grapes, Watermelon, Muskmelon, Apple, Orange, Papaya, Coconut, Cotton, Jute, Coffee, Arhar/Tur, Bajra, Castor Seed, Gram, Groundnut, Jowar, Linseed, Moong (Green Gram), Niger Seed, Onion, Other Cereals, Other Kharif Pulses, Other Rabi Pulses, Other Summer Pulses, Ragi, Rapeseed & Mustard, Safflower, Sesamum, Small Millets, Soyabean, Sugarcane, Sunflower, Tobacco, Tomato, Urad, Wheat, Other Oilseeds, Cotton(Lint)

> ⚠️ **Note:** `input_cost`, `yield_qtl`, and `demand` are estimates. Actual values vary by year, variety, and farming practices.

---

## 🗂️ Files in This Space

| File | Description |
|---|---|
| `app.py` | Main Gradio application — all logic lives here |
| `model1_npk.pkl` | Trained Random Forest classifier (soil & climate, 22 crops) |
| `model1_label_encoder.pkl` | Label encoder — maps class indices → crop names |
| `model2_full_scored.csv` | Pre-computed regional suitability scores (35 districts × 4 seasons × 33 crops) |
| `requirements.txt` | Python dependencies |
| `README.md` | This file |

---

## 📦 Requirements

```
gradio>=4.0.0
joblib
numpy
pandas
scikit-learn
```

> ⚠️ `scikit-learn` must be listed even though it is never directly imported — `joblib` needs it to deserialise the `.pkl` model files.

---

## 🏗️ System Architecture

```
Farmer Inputs (soil + weather + location + farm details)
        │
        ├──► Model 1 (Random Forest)
        │    model1_npk.pkl + model1_label_encoder.pkl
        │    → probability score per crop (22 crops)
        │
        ├──► Model 2 (Regional Scoring Table)
        │    model2_full_scored.csv
        │    → suitability score per crop (33 crops, filtered by district + season)
        │
        └──► Combined Score (60% Model 1 + 40% Model 2)
                    │
                    ▼
             Ranked crop list
                    │
                    ├──► Budget filter (skip if < 0.1 ha affordable)
                    ├──► Risk scoring (weather + market + budget + water)
                    ├──► Confidence labelling (based on Model 1 score)
                    └──► "Why" explanation (dynamically generated from real data)
                                │
                                ▼
                       HTML output cards (mobile-first UI)
```

---

## 🌱 Crops by Model Coverage

**Model 1 — Soil/Climate (22 crops):**
Apple, Banana, Blackgram, Chickpea, Coconut, Coffee, Cotton, Grapes, Jute, Kidneybeans, Lentil, Maize, Mango, Mothbeans, Mungbean, Muskmelon, Orange, Papaya, Pigeonpeas, Pomegranate, Rice, Watermelon

**Model 2 — Regional (33 crops):**
Arhar/Tur, Bajra, Banana, Castor Seed, Cotton, Gram, Grapes, Groundnut, Jowar, Linseed, Maize, Mango, Moong (Green Gram), Niger Seed, Onion, Other Cereals, Other Kharif Pulses, Other Rabi Pulses, Other Summer Pulses, Ragi, Rapeseed & Mustard, Rice, Safflower, Sesamum, Small Millets, Soyabean, Sugarcane, Sunflower, Tobacco, Tomato, Urad, Wheat, Other Oilseeds

**Crops in both models** get a fully blended score. Crops in only one model get a 0 from the missing model but are still shown if the other model scores them highly.

---

## 📊 Data Coverage

| | Details |
|---|---|
| State | Maharashtra, India |
| Districts | 35 |
| Seasons | Kharif (Jun–Sep), Rabi (Oct–Mar), Summer (Apr–Jun), Whole Year |
| Historical data range | 1997 – 2022 (26 years) |
| Regional data source | Maharashtra district-level crop production statistics |
| Soil model training data | 2,200 samples across 22 crops (100 per crop) |
| Soil model accuracy | 99.6% cross-validation (5-fold stratified) |

---

## 🏷️ Suitability & Risk Labels

### Regional Suitability (Model 2 raw score)

| Label | Score Range | Meaning |
|---|---|---|
| 🟢 Highly Suitable | ≥ 75% | Strong regional farming history |
| 🟡 Moderately Suitable | 50–74% | Good candidate, moderate history |
| 🟠 Low Suitability | 30–49% | Possible but limited evidence |
| 🔴 Not Recommended | < 30% | Little or no regional history |

### Overall Risk (combined score)

| Label | Score Range | Action |
|---|---|---|
| 🟢 Low | < 35 | Safe choice for this season |
| 🟡 Medium | 35–64 | Proceed with planning for main risk factor |
| 🔴 High | ≥ 65 | Consider reducing land allocation |

---

## 👨‍💻 Built For

Maharashtra farmers and agricultural advisors. Designed to be used on mobile phones in the field — large text, simple language, clear YES/NO verdicts, with detailed breakdown available on demand.

**Hackathon criteria addressed:**
- ✅ Integration of multiple data inputs (soil, weather, water, location, budget)
- ✅ Crop suitability prediction via dual-model AI pipeline
- ✅ Risk-aware scoring with 4-dimensional uncertainty handling
- ✅ Clear, simple recommendation outputs with plain-language explanations
- ✅ Farmer-friendly, mobile-responsive interface
- ✅ Scalable architecture (local model files, no external API dependencies)