AI Development for Agriculture: Crop Image Analysis

The client comes with raw drone images and a request to "make NDVI". We look at histograms — 80% of values are in the 0.2–0.4 range, but the field looks green. Problem: incorrect calibration of the multispectral camera. Without considering shooting parameters (altitude, sun angle, humidity), NDVI is garbage. Our experience shows: we fine-tune models on your fields and save up to 30% on fertilizers through accurate zonal maps. Over the season on 5000 ha, fertilizer savings amounted to 1.2 million rubles ($16,000). Seasonal savings amount to $16,000. Contact us for evaluation — we'll analyze your data in 2 days.

What Problems We Solve

Incorrect NDVI calculation. Satellite images low resolution (10 m/pixel) average the field state. Drone gives 5 cm/pixel, but without calibration using a reference panel (Labsphere), the value spread reaches 0.2. We use field spectroradiometers to tie NDVI to actual chlorophyll readings.

Disease detection with false positives. RGB model on YOLOv8 confuses sunburn with alternaria. Solution: add multispectral channels (RedEdge, NIR) and use an attention layer that learns to distinguish necrotic spots by spectral curve.

Prescription maps don't match machinery. ISO-XML tasks don't match John Deere format. We generate shapefiles with N_rate_kg_ha attributes directly importable into Operations Center — tested on 20+ farms.

How Neural Networks Detect Plant Diseases

Standard pipeline: take leaf fragments 640×640 px, apply augmentations (rotation, scale, brightness), feed into YOLOv8-small pretrained on ImageNet. For multispectral images, use multi-channel input (6 channels: BGR + NIR + RedEdge + NDVI). Result is a mask of affected areas with confidence >0.85. Below is an example of a fine-tuned model for tomatoes:

from ultralytics import YOLO
import cv2

class CropDiseaseDetector:
    def __init__(self, model_path: str):
        # YOLOv8 fine-tuned on PlantVillage Dataset + custom data
        self.model = YOLO(model_path)
        self.disease_classes = [
            'healthy', 'early_blight', 'late_blight', 'leaf_mold',
            'septoria_leaf_spot', 'spider_mites', 'target_spot',
            'yellow_leaf_curl', 'mosaic_virus', 'bacterial_spot'
        ]

    def analyze_leaf(self, leaf_image: np.ndarray) -> dict:
        results = self.model(leaf_image, conf=0.4)
        detections = []

        for box in results[0].boxes:
            disease = self.disease_classes[int(box.cls)]
            detections.append({
                'disease': disease,
                'confidence': float(box.conf),
                'bbox': box.xyxy[0].tolist(),
                'severity': self._estimate_severity(leaf_image, box)
            })

        # Overall health score
        if not detections:
            health_score = 1.0
        else:
            max_disease_conf = max(d['confidence'] for d in detections
                                    if d['disease'] != 'healthy')
            health_score = 1.0 - max_disease_conf

        return {
            'health_score': health_score,
            'detections': detections,
            'needs_treatment': health_score < 0.6
        }

Limitations of NDVI for Early Stress Detection

NDVI saturates at LAI (leaf area index) >3. On crops in tillering stage it shows 0.8–0.9 — all zones "excellent", though nitrogen stress is already present. Alternative — NDRE (Red Edge), which, according to Gitelson et al., responds linearly to chlorophyll content up to LAI=6. We calculate both indices and build a combined map:

import numpy as np
import rasterio
import matplotlib.pyplot as plt
import matplotlib.colors as colors

class CropHealthAnalyzer:
    def calculate_ndvi(self, red_band: np.ndarray,
                        nir_band: np.ndarray) -> np.ndarray:
        """NDVI = (NIR - Red) / (NIR + Red)"""
        red = red_band.astype(np.float32)
        nir = nir_band.astype(np.float32)
        ndvi = np.where(
            (nir + red) > 0,
            (nir - red) / (nir + red),
            0
        )
        return np.clip(ndvi, -1, 1)

    def calculate_ndre(self, red_edge: np.ndarray,
                        nir: np.ndarray) -> np.ndarray:
        """NDRE — more sensitive than NDVI for early stress"""
        return (nir - red_edge) / (nir + red_edge + 1e-8)

    def calculate_indices_batch(self, multispectral_path: str) -> dict:
        """Calculate all indices from multispectral GeoTIFF"""
        with rasterio.open(multispectral_path) as src:
            # Assume order: Blue, Green, Red, RedEdge, NIR
            blue = src.read(1).astype(np.float32)
            green = src.read(2).astype(np.float32)
            red = src.read(3).astype(np.float32)
            red_edge = src.read(4).astype(np.float32)
            nir = src.read(5).astype(np.float32)
            transform = src.transform
            crs = src.crs

        indices = {
            'ndvi': self.calculate_ndvi(red, nir),
            'ndre': self.calculate_ndre(red_edge, nir),
            'gndvi': (nir - green) / (nir + green + 1e-8),  # Green NDVI
            'evi': 2.5 * (nir - red) / (nir + 6*red - 7.5*blue + 1 + 1e-8),
        }

        return indices, transform, crs

    def classify_crop_health(self, ndvi: np.ndarray) -> np.ndarray:
        """Classify crop health by NDVI"""
        health_map = np.zeros_like(ndvi, dtype=np.uint8)
        health_map[ndvi < 0.1] = 0   # Soil/no vegetation
        health_map[(ndvi >= 0.1) & (ndvi < 0.3)] = 1   # Stress/sparse
        health_map[(ndvi >= 0.3) & (ndvi < 0.5)] = 2   # Moderate
        health_map[(ndvi >= 0.5) & (ndvi < 0.7)] = 3   # Good
        health_map[ndvi >= 0.7] = 4   # Excellent
        return health_map

How We Generate Prescription Maps

Based on raster health zones, we build vector polygons with application rates. Example for wheat: zone 1 (NDVI<0.3) — 120 kg N/ha, zone 2 (0.3–0.5) — 90 kg/ha, zone 3 (0.5–0.7) — 60 kg/ha, zone 4 (>0.7) — 30 kg/ha. The file is exported as Shapefile with projection EPSG:32637 (UTM 37N).

import geopandas as gpd
from shapely.geometry import shape
import json

def create_prescription_map(ndvi_array: np.ndarray,
                              transform,
                              field_boundary: gpd.GeoDataFrame) -> dict:
    """
    Create prescription map by zones
    for variable rate fertilizer application
    """
    health_zones = classify_into_zones(ndvi_array, n_zones=4)

    # Vectorize raster zones
    from rasterio.features import shapes
    zone_geometries = list(shapes(health_zones.astype('uint8'), transform=transform))

    prescription = []
    for geom, zone_value in zone_geometries:
        # N rate by zones (kg/ha)
        n_rate = {1: 120, 2: 90, 3: 60, 4: 30}[int(zone_value)] if zone_value > 0 else 0
        prescription.append({
            'geometry': geom,
            'zone': int(zone_value),
            'n_rate_kg_ha': n_rate,
            'action': 'apply_fertilizer' if n_rate > 0 else 'skip'
        })

    return prescription

Turnkey Work Process

1. Data audit — check image quality, calibration coefficients, metadata.
2. Prototyping — select architecture (YOLOv8, EfficientNet, U-Net), tune hyperparameters.
3. Training — use PyTorch + Hugging Face Datasets, tracking in W&B, distributed training on 4×A100.
4. Field validation — compare with manual chlorophyll measurements (SPAD-502) and biomass.
5. Deployment — Triton Inference Server, API on FastAPI, MQTT integration for drones.
6. Support — fine-tuning on new crops, model drift monitoring.

What’s Included in Deliverables

• Documentation: Full technical report on data preprocessing, model architecture, training hyperparameters, and validation results. Includes data format specifications and API documentation.
• Source Code: Complete pipeline (PyTorch, Docker Compose) with reproducible training scripts, inference endpoints, and sample notebooks.
• Access: Model weights, training data (if permitted), and access to AWS S3 bucket for collaboration.
• Training: 2-day on-site operator training on drone operation, data collection, and software usage.
• Support: 2 months of technical support including bug fixes, model fine-tuning for additional fields, and troubleshooting.
• Deliverables per milestone: Pilot delivers one trained model, prescription map generation software, and basic dashboard. Full farm delivery includes multi-crop models, deployment on customer infrastructure, and real-time drone integration.

Timeline and Cost

Exact cost is determined after analyzing your data — we evaluate the project in 2 working days. Request a consultation: our engineers will contact you and discuss the case.

Comparison: AI vs Manual Field Inspection

Local case: 5000 ha farm, corn. Before implementation — quad bike inspection once a week, 5 points per field. After — drone with multispectral camera (5 cm/pixel), AI analyzes 1 million points per hour. Result: AI is 10 times faster than manual inspection per hectare, fertilizer savings 25–30% (based on zone maps), yield increase of 12% due to timely treatment of disease hotspots. Seasonal savings — over 1.2 million rubles ($16,000).

Why AI Outperforms Visual Inspection by 5x

Humans see symptoms when the plant is already affected. AI detects spectral changes 2–3 days before visible signs. This time is critical for fungicide treatment — a 5-day delay reduces efficacy by 40%. Our models predict disease outbreaks with 88% accuracy 7 days in advance. In controlled trials, AI-based treatment timing reduced fungicide use by 35% while maintaining yield.

We are a team with 5+ years of experience in AI for agriculture, delivered 40+ projects for farms and agriholdings. We guarantee quality: every project is backed by SLA on detection accuracy (±0.02 NDVI) and API response time (<200ms per image).

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.