Abstract
Forests play an important role in the Earth systems for carbon sequestration and climate change mitigation, yet they have been increasingly disturbed by deforestation and forest degradation at an unprecedented pace. The Brazilian Amazon, for instance, experienced a 140% rise in deforestation from 2012 to 2020, with a record loss of 13,200 km2 between August 2020 and July 2021. Alarmingly, 87% of 2019 deforestation alerts occurred on private properties, with 61% in legally restricted areas. Existing deforestation monitoring systems, such as PRODES and the Global Forest Change dataset, use about 30m resolution satellite imagery, which is insufficient for operational validation at fine scales. The Deforestation Alert System by Imazon leverages high-resolution PlanetScope data (3-4m) but faces challenges due to fewer spectral bands and variations in reflectance values across different satellite sensors and dates. As a result, current validation is based mainly on manual inspection which is highly time-consuming. To address the challenges and automate the validation process, this work develops a system based on deep learning – known for its ability to capture complex texture patterns in high-resolution images – to inspect and confirm new deforestation sites. Specifically, the system takes inputs from potential deforestation sites suggested by coarse-resolution products and uses a pair of PlanetScope images before and after the change at each site to determine new deforestation (excluding existing deforestation). Our results demonstrate that the new system achieves robust and high-quality accuracy under various test conditions.
Introduction
Monitoring the Amazon is a race against time. While systems like PRODES and Global Forest Change provide essential data, they rely on 30-meter resolution imagery (like Landsat). At this scale, it is difficult to validate specific deforestation events or detect small-scale illegal clearing. High-resolution data from providers like PlanetScope (3-4m resolution) offers the necessary detail, but it introduces new challenges: high-res images have fewer spectral bands and significant variations in reflectance depending on the sensor and date. Currently, organizations like Imazon rely on experts to manually inspect these images to confirm deforestation alerts—a process that is accurate but slow and difficult to scale. This study seeks to bridge that gap by automating the validation process using deep learning.
Methods
The proposed system utilizes a U-Net architecture, a convolutional neural network (CNN) designed for fast and precise image segmentation. Unlike traditional pixel-based classification, U-Net captures spatial context and texture, which are critical for interpreting high-resolution imagery where individual pixel values can be noisy.
The model performs change detection by taking a "stack" of two images as input: one from time $T_0$ (before) and one from $T_1$ (after). Each image includes four spectral bands (RGB + Near-Infrared) and a calculated NDVI (Normalized Difference Vegetation Index) layer. The U-Net's encoder-decoder structure allows it to capture multi-scale contextual features for identification and precise localization for delineating boundaries. To address class imbalance—where deforested patches are small relative to the total forest area—the researchers employed a Dice Loss function, which prioritizes the overlap between predicted and actual deforested areas.
Datasets
The study focused on a region near the Amazon River (55° to 52.5°W). High-resolution Planet NICFI imagery (3-4m resolution) provided the primary data source. Ground truth labels were derived from deforestation alerts and polygons from MapBiomas, which aggregates data from monitoring systems like DETER, SAD, and GLAD. The dataset comprised 1,912 image pairs, partitioned into 80% training, 10% validation, and 10% testing sets.
Results and Discussion
The U-Net model was the clear winner across all metrics, achieving an F1-score of 0.8251, compared to 0.5505 for Random Forest and 0.5833 for FCN. It also demonstrated a high recall of 0.9226, indicating it is highly effective at catching actual deforestation events, with an overall accuracy of 0.9720.
Qualitatively, the U-Net produces continuous, clean polygons that closely match human-mapped "truth." In contrast, pixel-wise methods like Random Forest suffer from "salt-and-pepper" noise because they lack the spatial context necessary to distinguish between actual clearing and sensor-induced reflectance variations.
A key insight from the study is that at high resolutions, texture is more informative than raw spectral signatures. While NDVI is a standard proxy for vegetation, setting a global threshold is difficult due to varying sensor calibrations. The deep learning approach overcomes this by learning to recognize structural changes in the forest canopy rather than relying on absolute index values.
Conclusion
This work demonstrates a scalable, automated alternative to manual deforestation validation. While the current model shows robust performance, future work will focus on increasing sample diversity across different forest types and climates. Refining the model's sensitivity to distinguish between total clear-cutting and sparse tree removal remains a priority to support global environmental governance and forest conservation efforts.
Summary updated