Post-earthquake damage assessment is a critical step in disaster response and recovery. This paper introduces an innovative approach leveraging the EfficientNet-B0 neural network combined with wavelet-based data fusion (early and intermediate fusion strategies) to classify earthquake-related damage to existing buildings. The classification framework distinguishes between three damage levels: slight, moderate, and severe, in line with the standards defined by the Italian AeDES manual. The core innovation lies in the creation of a robust and expertly validated database, derived from a restructured and enhanced version of the PHI-Net dataset and a newly curated dataset based on the Macroseismic Photographic Database (DFM) from the Italian National Institute of Geophysics and Volcanology (INGV). This last dataset incorporates samples from various documented Italian earthquakes. The proposed method aims at supporting field operators by automating the damage classification step, thus streamlining the compilation of AeDES forms. Results demonstrate that the proposed method achieves a higher accuracy and reliability, compared to other state-of-the-art strategies, offering a significant contribution to automated seismic damage assessment.

The present paper aims at addressing the issue of how the use of multispectral and hyperspectral satellite imagery can help identify the presence of diseases and toxins within the agri-food sector in order to provide technical and scientific support to Food Security and Safety policies in Africa. This topic is one among those addressed within the AFRI4Cast project. A project, funded by European Space Agency, that in its entirety will provide national-, regional-, parcel-, pixel-specific in season production estimates of rust disease outbreak probability.

Reliable and efficient crop field segmentation is a fundamental pre-requisite for statistical analyses of agricultural practices. Traditional methodologies such as the Canny-Watershed (CW) algorithm require expert tuning of parameters for optimal results. This paper introduces an innovative approach for crop field segmentation in high-resolution satellite images, leveraging the use of multi-temporal Canny edge detection to train convolutional neural networks (CNNs) and fully automate the segmentation process. The Canny filter, applied to Sentinel-2 multi-temporal data, provides refined input for training ResUnet models, facilitating the generation of a generalized training dataset. ResUnet allows the model to learn complex features from diverse data, encapsulating seasonal changes. The dataset was specifically designed to enable the model to make accurate predictions from a single image, significantly outperforming traditional Canny filter predictions. In addition, the ResUnet may be applied to multiple images, generating output masks that, when overlayed, produce better results with respect to the multi-temporal Canny approach, demonstrating superior ability in recognizing real field boundaries while reducing false detections.To enhance generalizability, ResUnet is trained on a varied global dataset capturing a wide range of agricultural conditions and seasonal variations. This generalized model is tested across different regions and seasons, and preliminary results indicate that the proposed approach offers operational efficiency and accuracy in automating crop field segmentation.

Live Fuel Moisture Content (LFMC) is a fundamental variable of fire meteorology, fire behavior models and fire danger indices. The possibility of creating remote sensing LFMC products by directly training machine learning algorithms onto field measurements is severely limited by the sparse geographic distribution of such datasets, which are mostly concentrated in USA, Mediterranean Europe and Australia. Therefore, the physical foundation provided by a Radiative Transfer Model (RTM) such as PROSAIL remains an irreplaceable component of any LFMC product designed for global applicability. However, radiative transfer model inversion usually requires a lot of time and computing power. A Look-Up Table (LUT) approach saves time by running the model forward only during LUT creation, but still requires each row of the LUT to be compared with each observation when searching for the optimal solution. In this paper, we trained a random forest regressor on LUTs generated using PROSAIL and the Jasinski geometric model, aiming to exploit the efficiency of ML regressors to speed up calculation time while still maintaining the foundation of a physically-based approach that enables global applicability. The regressor was trained specifically to invert the LFMC, and was tested using Globe-LFMC v2 field-collected LFMC timeseries as a ground truth. The inversion, while returning results comparable in accuracy with the ones obtained using conventional methods, is now performed virtually instantly.

The incidence of wildfires and megafires with their disastrous consequences is increasing all over the planet, both in terms of burned area surface and in terms of power released by the fires. For this reason, the strategic importance of wildfire prevention is greater than ever, and providing the decision makers with powerful and state-of-the-art tools is an utmost priority. To this end, satellite observations offer a privileged platform to cover large spatial scales with a high time frequency. However, the most popular fire danger products tend to cover very large spatial scales at a coarse resolution, and inherently lack the capability to provide a level of detail which is very useful at the local scale. At the same time, using different products at different spatial scales would be impractical, and would increase the workload and training requirements of the personnel. To this end, this paper proposes a scalable-resolution fire danger index, named Daily Fire Danger Index, based on Sentinel-2 L2A and Sentinel-3 Synergy products. This novel index exploits both weather and satellite data to estimate all the main fire weather variables, and is calibrated using the historical records of wildfire occurrence in the area of interest.

The aim of this paper is to classify irrigated crop fields in Kenya during the year 2021 using Machine-Learning (ML) techniques and Sentinel-2 time-series data, identifying the best performing classifier and combinations of spectral indices in terms of accuracy scores.To observe changes due to irrigation and monitoring vegetation status, remote sensing data can provide valuable information. By combining different spectral bands, high-resolution multispectral Sentinel-2 imagery can provide various vegetation indices, sensitive to certain aspects of crop status and moisture content.To distinguish irrigated fields using these indices, ML methods for multivariate time-series classification can be applied. By training a ML model on time-series data consisting of six different vegetation indices, two different classifiers were tested: the Time Series Forest (TSF) and the Weasel-Muse algorithm (one of the most promising according to Ruiz et al., 2020).In conclusion, both algorithms demonstrated to be reliable for multi-index Sentinel-2 time-series classification, and the Normalized Multiband Drought Index and the Modified Normalized Water Index were revealed as good indicators to detect irrigation.

This paper presents the results of a collaboration between Italian and Chinese research groups carried out under the context of the GEO work programme and AfricultuReS H2020 project. The paper encompasses three main aspects: (a) the description of the results achieved on the high resolution crop mapping carried out in some African countries in the framework of the AfricultuReS project; (b) the description of the crop early warning service delivered in the framework of the AfricultuReS project; and (c) the application of the desert locust disaster monitoring model to the case of Somali. Using a multi-source data approach, the factors that have an important influence on the desert locust occurrence and spread process were extracted. The connection between the three points mentioned above must be sought in the fact that the combination of an accurate mapping of agricultural areas accompanied by techniques for estimating any threats to them allows to accurately estimate the possible effect in terms of production loss and food security.

Agricultural areas are naturally affected by significant variations within relatively short time intervals, in accordance with the growing season. These dynamics could, in principle, be exploited to classify different types of crops. Thus, this study aims to investigate methodologies and results of crop type classification making use of phenological information extracted from high spatial resolution satellite imagery. Vegetation indices (VI) retrieved from Sentinel-2 imagery are evaluated to track the year-round vegetation behavior. Starting from a multi-temporal image series of the same scene, the phenological profiles can be extracted and introduced into a supervised classification process to detect crop fields, discriminating among different species. Following this, we propose a cross-correlation based model that, using a priori information from ground training data, searches for the best matching phenology. When compared to machine learning models for crop classification, the one proposed in this study can provide useful information about phenology that can be stored and used for better monitoring spatio-temporal variations of crops species through the future years and guiding agricultural management accordingly. Our case studies are the regions of Bothaville and Harrismith, located in South Africa, and the region of Jendouba in Tunisia. The results for the Bothaville region show 89.19% of user accuracy on the main crop type classification (maize crops). For beans and sorghum, the confusion matrix attests 93% and 74% of accuracy respectively, even if their statistics are less significant due to the limited number of available ground data for secondary crops.