Over the past four decades, forests have experienced major disturbances, highlighting the need for Near Real-Time (NRT) monitoring. Traditional optical-based detection is cloud-sensitive, whereas Synthetic Aperture Radar (SAR)-based frameworks enable all-weather observation. Yet, SAR monitoring has mainly focused on humid tropical forests, with reduced performance in regions showing strong seasonal backscatter variation, such as tropical savannas. Detecting small-scale forest loss also remains difficult due to the spatial resolution loss from speckle filtering. This paper presents an unsupervised SAR-based disturbance detection method with NRT capabilities, using Bayesian inference. Building on an existing methodology, the approach processes singlepolarization Sentinel-1 SAR time series through Bayesian conjugate analysis. Forest disturbance is framed as a changepoint detection problem, where each new observation updates the probability of forest loss using prior information and a data model. The algorithm uses a hidden Markov chain to adapt recursively to seasonal variation and bypasses spatial filtering, preserving native data resolution and enhancing small-scale forest loss detection. Additionally, a methodology accounts for proximity to past disturbances. The method is tested on two 2020 reference datasets from the Brazilian Amazon and Cerrado savanna. The first covers small validation polygons (0.1–1 ha, excluding selective logging), totaling 2,650 ha in the Amazon and 450 ha in the Cerrado. The second includes larger clearings totaling 11,200 ha in the Amazon, and 12,700 ha in the Cerrado. A further comparison is conducted with operational NRT forest loss monitoring approaches. Results show substantial gains in detecting small-scale disturbances with reduced false alarms. In the Amazon, the method achieves an F1-score of 97.3% versus 93.1% for the current leading NRT approach. In the Cerrado, it reaches an F1-score of 97.4%, far exceeding the 33.3% of the optical-based method. For larger clearings, performance matches existing SAR approaches in the Amazon. While combined optical-SAR monitoring increases true positives, it also raises false alarm rates. In the Cerrado, the proposed method clearly outperforms optical monitoring, and in both regions it improves timeliness relative to individual operational approaches.

Les séries temporelles d’images SAR polarimétriques Sentinel-1 (S1) sont cruciales pour analyser les changements environnementaux. La détection de la déforestation, associée à un point de changement dans une série temporelle, doit surmonter les défis liés à la saisonnalité et au caractère aléatoire de la réflectivité des environnements forestiers. Cet article propose une méthode bayésienne appliquée à des séries temporelles de données SAR, exploitant les propriétés polarimétriques partielles des mesures S1 pour une meilleure détection. Les résultats réels montrent que le modèle surpasse les modèles bivariés standards, les méthodes à polarisation unique et une méthode existante basé sur l’imagerie optique.

The world’s forests are undergoing significant changes due to loss and degradation, emphasizing the need for Near Real-Time (NRT) monitoring to prevent further damage. Traditional monitoring methods using optical imagery are hindered by cloud coverage, while newer Synthetic Aperture Radar (SAR) systems, although operational in all weather conditions, face challenges such as sensitivity to soil moisture and the need for spatial filtering to reduce speckle effects. These limitations affect the detection of small-scale forest loss, especially in seasonally variable regions like dry forests and savannas. This paper presents a SAR-based forest disturbance detection method using Bayesian inference. Unlike traditional methods, this approach maintains the native resolution of the data by avoiding spatial filtering. Forest disturbance is modelled as a change-point detection problem within a non-filtered Sentinel-1 time series, where each new observation updates the probability of forest loss by leveraging prior information and a data model. This sequential adaptation ensures robustness against variations and trends, making it effective in monitoring disturbances across diverse forest types, including areas affected by seasonality. The proposed method was tested against other NRT monitoring systems for the year 2020, using small validation polygons (under 1 hectare) in the Brazilian Amazon and Cerrado savanna. Results demonstrate significant improvements in detecting small-scale disturbances and drastically reduced false alarm rates in both biomes. Notably, in the seasonality-sensitive Cerrado, our solution completely outperforms the leading and only existing optical technology.

Forests worldwide have undergone significant transformations due to forest loss, highlighting the critical need for real-time forest monitoring to prevent further vegetation loss and facilitate prompt interventions. Traditionally, forest loss monitoring relied on optical imagery, which is obstructed by its susceptibility to cloud coverage, especially in tropical regions. In recent times, Synthetic Aperture Radar (SAR)-based systems have emerged to enable all-weather operability. However, SAR-based approaches encounter challenges, such as the alterations in backscatter caused by factors like soil moisture variations. Moreover, accurately detecting small-scale disturbances remains problematic for SAR systems, partly due to the spatial filtering techniques employed to mitigate the effects of speckle. Additionally, monitoring forest loss in regions characterized by pronounced seasonality in backscatter signals, such as dry forests and savannas, poses limitations, resulting in substantial under-monitoring of these extensive carbon sinks. This study introduces an unsupervised SAR-based method for detecting forest loss, employing Bayesian inference through an infinite state Markov chain.

In this paper, we propose an unsupervised statistical approach for near real-time monitoring of forest loss, leveraging Bayesian inference. We address the identification of forest loss as a change-point detection problem within non-filtered Sentinel-1 single polarization time series data. Each new observation contributes to the probability of deforestation occurrence, utilizing prior knowledge and a data model. Our method offers the advantage of detecting small-scale deforestation without resorting to spatial filtering techniques, thus preserving the native spatial resolution of the Sentinel-1 measurements. To assess its effectiveness, we conducted comparative evaluations against existing operational deforestation monitoring systems. The validation campaign revealed that our method exhibits enhanced detection performance with low false alarm rates with respect to existing systems across diverse landscapes, including dense forest regions such as the Brazilian Amazon, as well as seasonality-dependent areas like the Cerrado, which is strongly under-monitored by existing technology. This robustness stems from the sequential adaptive process inherent in our approach, which enables effective monitoring even in the presence of backscatter variations.

Forest loss is a global issue that requires real-time surveillance to prevent further vegetation loss. This study presents an unsupervised SAR-based technique that leverages Bayesian inference and infinite state Markov chains to identify forest loss, overcoming the limitations of current methods. Our approach significantly improves accuracy and reduces false alarm rates compared to existing Near Real-Time (NRT) forest loss monitoring systems and enlarges the conditions of operability.

The world’s forests have undergone substantial changes in the last decades. In the tropics, 17% of moist forests disappeared between 1990 and 2019, through deforestation and forest degradation [7]. These changes contribute greatly to biodiversity loss through habitat destruction, soil erosion, terrestrial water cycle disturbances, and anthropogenic CO2 emissions. Continuous monitoring of global deforestation is a fundamental tool to support preservation actions and to stop further destruction of vegetation. Several forest disturbance detection systems have already been developed, mainly based on space-borne optical remote sensing [4] which is severely limited by cloud coverage in the tropics. Contrarily to optical imagery, SAR products have the great potential of being insensitive to the presence of clouds. In recent years, several SAR-based systems have been developed and are now operational in different dense forest areas across the tropics [2], [3], [5], [6]. Despite the extensive coverage and temporal density of acquisitions, C-band SAR data like Sentinel-1 are not ideal for deforestation monitoring since the returned backscatter can be altered by variations in soil moisture and others. In this work, we investigate a new method to monitor forest loss in a near real-time manner exploiting the principle of Bayesian inference. In particular, forest loss is treated as a change-point detection problem within a univariate time series (i.e. Sentinel-1 single polarization), in which each new observation contributes to the probability of having or not deforestation in a Bayesian-like manner [1]. Detection delay and false alarm reduction have been investigated through the extension of the algorithm to the multivariate case of dual-polarization Sentinel-1 acquisitions. Given the synchronous nature of VV, VH acquisitions, such a modification allows an increase in the equivalent number of looks on a pixel on the ground, hence augmenting the level of confidence of an issued alert. A validation campaign has been conducted to assess the performance of the method. The test sites are located in French Guiana and Brazil where deforestation takes place constantly and near real-time monitoring is fundamental for law enforcement practices. Additionally, a comparison with a well-known deforestation monitoring technique, namely Maximum Likelihood Ratio Test, has been performed to further evaluate the proposed method. Conclusively, the potential of extending the current method to asynchronous data sources such as Sentinel-2 optical data is addressed.