Can a satellite really see the growth stages of coffee in Brazil?

Coffee consumption is rising, but growers face drought and temperature swings. To curb losses, they try new farming techniques. Yet, crop variability makes coffee mapping hard. This study proposes a novel method to identify four coffee growth phases: sowing, production, pruning, and renewal. The team used a dense time series of Sentinel‑2 images combined with Landsat data, updated roughly every three days. By analysing spectral bands, indices and texture, they trained two ensemble models: Random Forest and Extreme Gradient Boosting. Random Forest achieved over 95 % sensitivity and specificity, showing strong ability to spot coffee pixels. Phase classification reached 77 % accuracy for renewal and 93‑95 % for other phases. The approach is scalable and helps farmers monitor climate‑smart practices, boost yields and cut environmental impact. Images are captured almost every three days, providing near‑continuous coverage of Brazil. Harmonization aligns spatial (10 m) and temporal resolution, enabling rapid vegetation change analysis. Visible (B2‑B4), near‑infrared (B5‑B7) and short‑wave infrared (B8‑B12) bands are used to compute spectral indices. Main indices include NDVI, EVI, NDWI and the green energy ratio, all linked to plant health. Texture metrics from GLCM add canopy structure details. These features feed two ensemble learning models: Random Forest and XGBoost. Random Forest builds on decision trees, while XGBoost sequentially optimises gradients to reduce errors. Both models are trained on manually outlined coffee polygons from various producing regions. Validation showed a confusion matrix with error rates below 5 % for coffee detection. Detection accuracy exceeds 95 % for both sensitivity and specificity, confirming method effectiveness. Phase accuracies range from 77 % (stumping) to 95 % (production). These results show that combining high‑frequency data with advanced algorithms supports more precise agronomic decisions. The framework is modular, allowing use on diverse crops and varied climates. Its flexibility makes it suitable for transitioning to more climate‑resilient production systems.

The project uses the Harmonized Landsat Sentinel‑2 (HLS) time series, which merges optical data from Landsat‑8/9 and Sentinel‑2. Images are acquired roughly every three days, providing near‑continuous coverage of Brazil’s territory. Harmonization aligns spatial (10 m) and temporal resolution, enabling rapid vegetation change analysis. Visible (B2‑B4), near‑infrared (B5‑B7) and short‑wave infrared (B8‑B12) bands are used to compute spectral indices. Main indices include NDVI, EVI, NDWI and the green energy ratio, all linked to plant health. Texture metrics from GLCM add canopy structure details. These features feed two ensemble learning models: Random Forest and Extreme Gradient Boosting (XGBoost). Random Forest builds on decision trees, while XGBoost sequentially optimises gradients to reduce errors. Both models are trained on manually outlined coffee polygons from various producing regions. Validation showed a confusion matrix with error rates below 5 % for coffee detection. Detection accuracy exceeds 95 % for both sensitivity and specificity, confirming method effectiveness. Phase accuracies range from 77 % (stumping) to 95 % (production). These results show that combining high‑frequency data with advanced algorithms supports more precise agronomic decisions. The framework is modular, allowing use on diverse crops and varied climates. Its flexibility makes it suitable for transitioning to more climate‑resilient production systems.