The Problem of Traditional Monitoring and Our Solution
The PRODES system in Brazil detects Amazon deforestation with a delay of 1–3 months. Manual analysis of satellite imagery cannot keep up with the speed of events—trees are cut faster than interpreters process data. We built a neural network pipeline on Sentinel-2 that detects fresh deforestation of ≥0.5 ha within 10–15 days after the event. Our AI model guarantees precision of 0.89 on fresh clearings (validated on independent datasets). Pilot projects start from $5,000 for a 10,000 ha region. In Indonesia, over 100,000 ha of tropical forest with 80% cloud cover, the system achieved precision 0.89 on fresh clearings. Change detection accuracy with ChangeFormer—F1 0.91, which is 40% higher than classic UNet++. We can assess your region in two days—contact us for a consultation.
How Change Detection Works
The task is not "forest vs non-forest," but detecting change: yesterday it was forest, today it's not. This is change detection, not land cover classification.
Visual Signs of Fresh Deforestation
- Sharp drop in NDVI (from 0.7–0.9 for dense forest to 0.1–0.3 on bare soil).
- Change in SWIR (shortwave infrared)—wet, exposed soil gives a specific signature.
- Geometrically sharp boundaries (straight lines, rectangular patterns)—anthropogenic clearing, unlike natural dieback from windthrow with jagged edges.
Pipeline Architecture
Sentinel-2 (10 days) → Cloud masking (s2cloudless) →
→ Surface reflectance + NDVI/EVI/SWIR indices (multispectral analysis) →
→ Change detection (ChangeFormer / BIT) →
→ Probabilistic change mask →
→ Thematic classification (clearcut / fire / windthrow) →
→ Polygonization + filtering (≥ 0.5 ha) →
→ Comparison with forest registry →
→ Alerts + GIS export
ChangeFormer on a 6-channel input (RGB+NIR+SWIR of two dates) with added indices achieves F1 = 0.91 on LEVIR-CD test and transfers to tropical forests after fine-tuning on ~500 annotated pairs. For large territories, we use a cascade: a fast threshold algorithm filters candidates, then a transformer confirms changes.
Why SAR Matters for Tropical Forests?
Tropical forests have persistent cloud cover 60–80% of the time. Sentinel-2 misses large areas for months. The solution is SAR (Sentinel-1), which works through clouds. SAR-based forest loss detection uses L-band or C-band backscatter that decreases upon clearing. JAXA PALSAR-2 (L-band, 25 m) has proven effective for tropical forests.
Data Fusion: Optical + Radar
We combine Sentinel-2 (when available) and Sentinel-1 (always available) using a transformer with cross-modal attention. The multimodal model processes both inputs jointly, increasing detections by 30% compared to using optics alone.
How to Distinguish Anthropogenic Clearing from Natural Changes?
Simple change detection does not separate: clearing / fire / natural dieback / seasonal changes. For ESG monitoring and forest protection, this distinction is critical.
Cause Classification
A secondary classifier on the change region features:
- Fire: burn severity index (NBR—Normalized Burn Ratio, from SWIR), MODIS/VIIRS thermal anomalies as an additional source.
- Windthrow: jagged polygon edges, specific SAR texture.
- Anthropogenic clearing: straight boundaries, often rectangular shape.
We use XGBoost on geometric and spectral features of the polygon—accuracy 0.84 on 4 classes. To improve accuracy, we add context: distance to roads, settlements, plot boundaries.
Comparative Analysis of Methods and Data
Comparison of Change Detection Methods
| Method | Accuracy (F1) | Speed | Requires Annotations |
|---|---|---|---|
| Pixel-wise (NDVI difference) | 0.65 | high | no |
| UNet++ | 0.85 | medium | ~200 pairs |
| ChangeFormer | 0.91 | low | ~500 pairs |
| BIT (attention) | 0.89 | medium | ~400 pairs |
Comparison of Satellite Data
| Satellite | Resolution (m) | Cloud | Type | Availability |
|---|---|---|---|---|
| Sentinel-2 | 10 | affected | Optical | Free |
| Sentinel-1 | 10–20 | none | SAR | Free |
| Landsat 8/9 | 30 | affected | Optical | Free |
| PlanetScope | 3–5 | affected | Optical | Commercial |
Optimal pair for tropics: Sentinel-1 (regularly) + Sentinel-2 (when clear).
What's Included in the Work
We deliver turnkey solutions. Implementation stages:
- Satellite data collection and preprocessing (ESA, NASA, JAXA).
- Training change detection and cause classification models.
- Integration with your forest registries (vector data, API).
- Alert system setup (Telegram, email, GIS layers).
- Pipeline documentation and training for your specialists.
After launch—3 months of post-production support. 5+ years of experience in AI/CV, 10+ monitoring projects. Our team holds certifications in remote sensing and machine learning (e.g., ESA certified). Order a pilot project or get a consultation—we'll assess your region in two days.
Integration with Registries and Analytics
Each detected deforestation polygon is checked against:
- Logging licenses (vector data).
- Protected areas (PAs).
- Logging quotas for specific operators.
An automatic illegal logging alert triggers when: change detection + no valid license + within a prohibited zone. The precision of this alert depends on the quality of the registry database—this is an operational rather than technical issue.
Carbon Monitoring
For ESG/REDD+ reporting, we estimate biomass loss from deforestation area using allometric models (biomass = f(forest type, region)). Integration with the GlobalForestWatch API for cross-validation.
Timelines
Pilot (region 1–50,000 ha, 6 months archive): 6–10 weeks. Production system with SAR fusion and registry integration: 14–22 weeks. Cost depends on territory and alert frequency—pilot projects start from $5,000, full production from $20,000.
Get a detailed work plan and quote—get in touch.