Monitoring agriculture from satellite remote sensing data, such as multispectral images, has become a powerful tool since it has demonstrated a great potential for providing timely and accurate knowledge of crops. Detecting anomalies in time series of multispectral remote sensing images for crop monitoring is generally performed using a large sample of historical data at a pixel level. Conversely, this article presents a framework for anomaly detection (AD), localization, and classification that exploits the temporal information contained in a given season at a parcel level to detect and localize outliers using hidden Markov models (HMMs). Specifically, the AD part is based on the learning of HMM parameters associated with unlabeled normal data that are used in a second step to detect abnormal crop parcels referred to as anomalies. The learned HMM can also be used in time segments to temporally localize the anomalies affecting the crop parcels. The detected and localized anomalies are finally classified using a supervised classifier, e.g., based on support vector machines. The proposed framework is applicable to images partially covered by clouds and can handle a set of crop parcels acquired in the same season bypassing problems due to crop rotations. Numerical experiments are conducted on synthetic and real data, where the real data correspond to vegetation indices extracted from several multitemporal Sentinel-2 images of rapeseed crops. The proposed approach is compared to standard AD methods yielding better detection rates with the advantage of allowing anomalies to be localized and characterized.

This paper studies the detection of anomalous crop development at the parcel-level based on an unsupervised outlier detection technique. The experimental validation is conducted on rapeseed and wheat parcels located in Beauce (France). The proposed methodology consists of four sequential steps: (1) preprocessing of synthetic aperture radar (SAR) and multispectral images acquired using Sentinel-1 and Sentinel-2 satellites, (2) extraction of SAR and multispectral pixel-level features, (3) computation of parcel-level features using zonal statistics and (4) outlier detection. The different types of anomalies that can affect the studied crops are analyzed and described. The different factors that can influence the outlier detection results are investigated with a particular attention devoted to the synergy between Sentinel-1 and Sentinel-2 data. Overall, the best performance is obtained when using jointly a selection of Sentinel-1 and Sentinel-2 features with the isolation forest algorithm. The selected features are co-polarized (VV) and cross-polarized (VH) backscattering coefficients for Sentinel-1 and five Vegetation Indexes for Sentinel-2 (among us, the Normalized Difference Vegetation Index and two variants of the Normalized Difference Water). When using these features with an outlier ratio of 10%, the percentage of detected true positives (i.e., crop anomalies) is equal to 94.1% for rapeseed parcels and 95.5% for wheat parcels.

This paper tackles the problem of missing data imputation for noisy and non-Gaussian data. A classical imputation method, the Expectation Maximization (EM) algorithm for Gaussian mixture models, has shown interesting properties when compared to other popular approaches such as those based on $k$ -nearest neighbors or on multiple imputations by chained equations. However, Gaussian mixture models are known to be non-robust to heterogeneous data, which can lead to poor estimation performance when the data is contaminated by outliers or have non-Gaussian distributions. To overcome this issue, a new EM algorithm is investigated for mixtures of elliptical distributions with the property of handling potential missing data. This paper shows that this problem reduces to the estimation of a mixture of angular Gaussian distributions under generic assumptions ( i.e. , each sample is drawn from a mixture of elliptical distributions, which is possibly different for one sample to another). In that case, the complete-data likelihood associated with mixtures of elliptical distributions is well adapted to the EM framework with missing data thanks to its conditional distribution, which is shown to be a multivariate $t$ -distribution. Experimental results on synthetic data demonstrate that the proposed algorithm is robust to outliers and can be used with non-Gaussian data. Furthermore, experiments conducted on real-world datasets show that this algorithm is very competitive when compared to other classical imputation methods.

This paper investigates tree species classification using the Sentinel-2 multispectral satellite image time series (SITS). Despite its importance for many applications and users, such mapping is often unavailable or outdated. The value of using SITS to classify tree species on a large scale has been demonstrated in numerous studies. However, many methods proposed in the literature still rely on a standard machine learning algorithm, usually the random forest (RF) algorithm. Our analysis shows that the use of deep learning (DL) models can lead to a significant improvement in classification results, especially in an imbalanced context where the RF algorithm tends to predict the majority class. In our case study in central France with 10 tree species, we obtained an overall accuracy (OA) of around 95% and an F1-macro score of around 80% using three different benchmark DL architectures (fully connected, convolutional, and attention-based networks). In contrast, using the RF algorithm, the OA and F1 scores obtained were 92% and 60%, indicating that the minority classes are poorly classified. Our results also show that DL models are robust to imbalanced data, although small improvements can be obtained by specifically addressing this issue. Validation on independent in situ data shows that all models struggle to predict in areas not well covered by training data, but even in this situation, the RF algorithm is largely outperformed by deep learning models for minority classes. The proposed framework can be easily implemented as a strong baseline, even with a limited amount of reference data.

This article studies the monitoring of oak dieback in forests of the Centre-Val de Loire region (France), where drought-induced dieback has become a major concern due to climate change. The main objective of the study is to evaluate the applicability of multispectral satellite time series for operational monitoring of forest dieback. Using in situ data collected from 2017 to 2022 on approximately 2700 oak plots, a multiyear mapping of the analyzed region was performed using the random forest algorithm and Sentinel-2 images. Our results show that it is possible to detect oak dieback accurately (average overall accuracy = 80% and average balanced accuracy = 79%). A spatial cross-validation analysis also evaluates the performance of the model on regions that were never encountered during training, across all years, resulting in a slight decrease in accuracy ( $\sim$ 5%). The study also highlights the importance of measuring the stability and performance of the classification model over time, in addition to standard cross-validation metrics. A feature analysis shows that the shortwave infrared part of the spectrum is the most important for mapping forest dieback, while the red-edge portion of the spectrum can increase the stability of the model over time. Overall, both in situ data and model predictions showed evidence of forest decline in many areas of the study region. Our results suggest that large areas of forest can decline over short periods of time, highlighting the interest of satellite data to provide timely and accurate information on forest status.

Remote sensing satellites allow large-scale and fast detections of forest loss. Operational forest loss detection systems have been mainly developed over tropical forests, however, it is increasingly important to have access to accurate and up-to-date information on temperate forests. In this study, we adapted a Sentinel-1 based near real-time tropical forest loss detection method, based on the radar change ratio, to detect French temperate forests clear-cuts. Using ancillary data, annual and sub-monthly clear-cuts were assessed for broadleaf and conifer forests, for various tree species, over public and private forests. Using 967 validation plots, the maps exhibited recall and precision of 80.9% and 99.4% respectively. The clear-cuts area shows remarkable stability over time from 2020. We found seven times more clear-cuts in private forests than in public forests, although the surface area of private forests is only three times that of public forests. It was also demonstrated that only 1.6% out of 4,530 dieback reference plots, and 6.2% of maps of forest bark beetle attacks, were confused with clear-cuts before clear-cuts actually occurred, which makes our maps complementary with forest dieback maps. Collectively, the findings of this study could have significant implications for the implementation of a radar-satellite-based system designed for the real-time detection of large-scale clear-cuts in European temperate forests. The data are available via https://ee-sufosatclearcuts.projects.earthengine.app/view/sufosat-clearcuts.