AI System for Biodiversity Monitoring

Camera traps, drones with multispectral cameras, hydrophones, and acoustic sensors generate terabytes of data daily. Without AI, processing this volume is impossible. Manual field observations cover limited areas — two or three specialists walk a few square kilometers per season. An AI system covers thousands of hectares in real time.

We design and deploy AI systems for automated biodiversity monitoring. Our team has completed projects for nature reserves, conservation organizations, and ecological researchers. We solve four core tasks: species identification from images and audio, individual counting on aerial photos, re-ID of specific animals for long-term tracking, and detection of habitat changes. Contact us for a consultation — we evaluate your project within two working days.

Biodiversity Monitoring AI: System Architecture

The system is built on a two-stage pipeline: a YOLO detector locates animals in the frame, and an EfficientNet classifier identifies the species. For re-ID of individuals, we use ArcFace embeddings compared against a gallery. Acoustic data is processed by BirdNET or PANN for bird and other animal classification.

All components are deployed in Docker containers. Results are stored in PostGIS for geospatial analysis. A Grafana dashboard displays an activity map and population dynamics.

AI Species Detector for Biodiversity Monitoring

import numpy as np
import cv2
import torch
from ultralytics import YOLO
from torchvision import models, transforms
from PIL import Image
from dataclasses import dataclass
from typing import Optional
import json

@dataclass
class WildlifeDetection:
    track_id: int
    species: str
    common_name_ru: str
    confidence: float
    bbox: list
    individual_id: Optional[str]   # re-ID if model trained
    behavior: Optional[str]        # resting / moving / feeding
    camera_id: str
    timestamp: float

class WildlifeMonitor:
    """
    Detection and classification of wild animals from camera trap images.
    Datasets:
    - iNaturalist (1.4M images, 5000+ species) — for pretraining
    - LILA BC (camera trap images) — leopard, snow leopard, etc.
    - Snapshot Serengeti (22M images, 48 species)
    - CaltechCameraTraps (20 species, North America)
    Two-stage pipeline: YOLO detection → EfficientNet species classification.
    """
    def __init__(self, detector_path: str,
                  classifier_path: str,
                  species_vocab_path: str,
                  reid_model_path: Optional[str] = None,
                  device: str = 'cuda'):
        self.detector = YOLO(detector_path)
        self.device = device

        # Classifier: EfficientNet-B3 fine-tuned on iNaturalist
        with open(species_vocab_path) as f:
            vocab_data = json.load(f)
        self.species_list = vocab_data['species']
        self.species_ru = vocab_data.get('species_ru', {})

        self.classifier = models.efficientnet_b3(pretrained=False)
        n_classes = len(self.species_list)
        self.classifier.classifier[-1] = torch.nn.Linear(
            self.classifier.classifier[-1].in_features, n_classes
        )
        state = torch.load(classifier_path, map_location=device)
        self.classifier.load_state_dict(state)
        self.classifier = self.classifier.to(device).eval()

        self.classify_transform = transforms.Compose([
            transforms.Resize((300, 300)),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406],
                                  [0.229, 0.224, 0.225])
        ])

        # Re-ID (optional): megapixel model for fur patterns
        self.reid_model = None
        if reid_model_path:
            self.reid_model = torch.load(reid_model_path,
                                          map_location=device).eval()
        self._individual_gallery: dict[str, np.ndarray] = {}

    def process_camera_trap_image(self, image: np.ndarray,
                                   camera_id: str,
                                   timestamp: float) -> list[WildlifeDetection]:
        """
        Process a camera trap image.
        Camera traps: night (IR) + day, various resolutions.
        """
        # Enhance night image
        enhanced = self._enhance_camera_trap(image)

        detections_raw = self.detector(
            enhanced, conf=0.25, verbose=False
        )
        results = []

        for box in detections_raw[0].boxes:
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            track_id = int(box.id.item()) if box.id is not None else -1

            # Crop for classification
            crop = enhanced[max(0,y1):y2, max(0,x1):x2]
            if crop.size == 0:
                continue

            species, conf = self._classify_species(crop)
            species_ru = self.species_ru.get(species, species)
            behavior = self._estimate_behavior(crop, box)

            # Re-ID
            individual_id = None
            if self.reid_model:
                individual_id = self._get_individual_id(crop, species)

            results.append(WildlifeDetection(
                track_id=track_id,
                species=species,
                common_name_ru=species_ru,
                confidence=round(conf, 3),
                bbox=[x1, y1, x2, y2],
                individual_id=individual_id,
                behavior=behavior,
                camera_id=camera_id,
                timestamp=timestamp
            ))

        return results

    @torch.no_grad()
    def _classify_species(self, crop: np.ndarray) -> tuple[str, float]:
        pil = Image.fromarray(cv2.cvtColor(crop, cv2.COLOR_BGR2RGB))
        tensor = self.classify_transform(pil).unsqueeze(0).to(self.device)
        logits = self.classifier(tensor)
        probs = torch.softmax(logits, dim=-1).squeeze()
        conf, idx = probs.max(dim=0)
        species = self.species_list[int(idx.item())]
        return species, float(conf.item())

    def _enhance_camera_trap(self, image: np.ndarray) -> np.ndarray:
        """Enhance IR camera trap images"""
        # Check for IR (low saturation)
        hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
        saturation = float(np.mean(hsv[:, :, 1]))

        if saturation < 20:  # IR image
            gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
            clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8, 8))
            enhanced_gray = clahe.apply(gray)
            return cv2.cvtColor(enhanced_gray, cv2.COLOR_GRAY2BGR)

        # Day image with excessive shadow
        lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
        clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
        lab[:, :, 0] = clahe.apply(lab[:, :, 0])
        return cv2.cvtColor(lab, cv2.COLOR_LAB2BGR)

    def _estimate_behavior(self, crop: np.ndarray,
                            box) -> str:
        """Simple behavior classification based on bbox aspect ratio"""
        x1, y1, x2, y2 = map(int, box.xyxy[0])
        w, h = x2 - x1, y2 - y1
        aspect = w / max(h, 1)
        if aspect > 2.0:
            return 'resting'  # lying horizontally
        elif aspect < 0.6:
            return 'alert'    # standing upright, head raised
        return 'moving'

    @torch.no_grad()
    def _get_individual_id(self, crop: np.ndarray,
                            species: str) -> Optional[str]:
        """Re-ID via embedding similarity (for leopards, tigers)"""
        pil = Image.fromarray(cv2.cvtColor(crop, cv2.COLOR_BGR2RGB))
        transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
        ])
        tensor = transform(pil).unsqueeze(0).to(self.device)
        embedding = self.reid_model(tensor).squeeze().cpu().numpy()
        embedding /= np.linalg.norm(embedding) + 1e-8

        # Search in gallery
        best_match = None
        best_sim = 0.75  # threshold

        gallery_key = f'{species}_'
        for ind_id, gallery_emb in self._individual_gallery.items():
            if not ind_id.startswith(gallery_key):
                continue
            sim = float(np.dot(embedding, gallery_emb))
            if sim > best_sim:
                best_sim = sim
                best_match = ind_id

        if best_match is None:
            # New individual
            new_id = f'{species}_{len(self._individual_gallery)+1:04d}'
            self._individual_gallery[new_id] = embedding
            return new_id
        return best_match


class AerialAnimalCounter:
    """
    Animal counting on aerial images (drone/aircraft).
    Application: monitoring ungulate herds, penguin counting,
    fur seal rookery inventories.
    SAHI is essential for 50+ MPix orthophotos.
    """
    from sahi import AutoDetectionModel
    from sahi.predict import get_sliced_prediction

    def __init__(self, model_path: str, device: str = 'cuda'):
        from sahi import AutoDetectionModel
        self.sahi_model = AutoDetectionModel.from_pretrained(
            model_type='ultralytics',
            model_path=model_path,
            confidence_threshold=0.35,
            device=device
        )

    def count_herd(self, aerial_image: np.ndarray,
                    species_hint: str = 'ungulate') -> dict:
        from sahi.predict import get_sliced_prediction
        result = get_sliced_prediction(
            aerial_image, self.sahi_model,
            slice_height=640, slice_width=640,
            overlap_height_ratio=0.2, overlap_width_ratio=0.2
        )
        count = len(result.object_prediction_list)
        density = count / (aerial_image.shape[0] * aerial_image.shape[1] / 1e6)

        return {
            'species_hint': species_hint,
            'count': count,
            'density_per_km2': round(density * 1e6, 1),  # pixels² → km²
            'detections': [
                {'bbox': [p.bbox.minx, p.bbox.miny, p.bbox.maxx, p.bbox.maxy],
                 'conf': p.score.value}
                for p in result.object_prediction_list
            ]
        }

Accuracy Metrics and Timeline

Below are indicative quality metrics for typical biodiversity monitoring tasks. Metrics depend on dataset quality and shooting conditions.

Acoustic Biodiversity Monitoring

Sound data reveals species that are difficult to detect visually: nocturnal birds, bats, aquatic mammals — all leave an acoustic trace. We deploy hydrophones and acoustic recorders (AudioMoth, SongMeter) and process recordings with BirdNET and PANN.

The acoustic analysis process includes three stages. First, the audio file is segmented into 3–5 second chunks. Then each chunk is converted to a mel-spectrogram and passed through the classifier. Finally, post-processing merges adjacent detections of the same species into events, filtering by a minimum confidence threshold of 0.5.

BirdNET achieves cmAP 72–78% on European and North American birds in BirdCLEF competition. For exotic or tropical species, fine-tuning on regional data is required. We help collect and label a training corpus.

Integration with Geographic Information Systems

Monitoring data is stored in PostGIS. Each observation has coordinates, a timestamp, and a source identifier. QGIS and ArcGIS connect directly via the PostGIS connector. For dashboards we use Grafana with the Geomap plugin — the observation grid updates every hour.

Population trends are calculated using a 30-day sliding window. An anomalous drop in species activity by more than 30% of the baseline automatically generates an alert. This allows ecologists to respond quickly to changes before a quarterly report.

What's Included

• Data audit: inventory of sensors, formats, and data volumes.
• Labeling and training: creating a dataset, fine-tuning detector and classifier.
• Processing pipeline: automated processing of camera trap images, orthophotos, or audio recordings.
• Map dashboard: visualization of species activity on a map with filters.
• Documentation and team training: operation instructions and API specifications.
• Technical support: 3 months of maintenance after launch.

Time Estimates

Contact us for a consultation — we will evaluate your data and propose the optimal architecture. Reach out for a preliminary audit of your material. We have helped 15+ organizations automate biodiversity monitoring: nature reserves, environmental NGOs, industrial enterprises conducting environmental impact assessments. We guarantee transfer of source code and documentation after project completion. Our clients value that the system runs offline: all models are deployed on your own infrastructure; data never leaves the perimeter.

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.