Conference Agenda

The first two years of this project focused on retrieval of biophysical variables of vegetation, estimation of bare soil properties, and monitoring pest and diseases. The remote sensing data used include non-imaging hyperspectral data and image data from satellites such as Sentinel-2, GF-1 and PRISMA. The field experiments were carried out at Maccarese farm in Central Italy and farms in the Quzhou district in China.

For vegetation, field campaigns were primarily performed contemporary to the Sentinel-2 or PRISMA acquisitions on the Maccarese (IT) farm along the winter wheat and maize growing season. A Decision Trees Ensembles (DTE) model has been developed by training a database of plant canopy spectral and directional reflectance generated by using PROSAIL-PRO. A reduction of dimensionality performed by selecting a subset of PRISMA bands (i.e. excluding water vapor absorption bands and noisy bands) were applied prior to the training phase. DTE model has been selected among the tested algorithms as appears a robust method in dealing with noise, less prone to overfitting and less sensitive to the difference between simulated and real data. LAI, LCC, FCOVER and FAPAR were retrieved on the Maccarese test site with a RMSE value of 1.15 m² m⁻², 11.45 μg cm⁻², 0.16% and 0.16%, respectively. RMSE values are always lower than the ones derived by Sentinel-2 with SNAP. Further steps will consider the retrieval of additional biophysical variables as carotenoid, nitrogen and leaf mass area.

For topsoil characterization the test site is located in the Quzhou district and more than 100 soil samples were collected and analyzed in the lab to assess SOC. PRISMA data preprocessing was similar to the one for vegetation with the inclusion of data pre-treatment i.e. Savitzy–Golay filter (SG), Standard Normal Variate (SNV) and Continuum Removal (CR). PLSR and Cubist algorithms were applied only to bare soil pixel by using a 10-fold cross validation. Results show a significant accuracy for SOC showing a RPD higher than 2. The result for Quzhou reflects the small field size with respect to the PRISMA GSD [30m]. To minimize the field size effects, an optimized pan sharpening procedure by using Sentinel-2 bands and the 5m/pixel PRISMA PAN, is under development.

For what concerns crop threats among which pests & diseases, Italian and Chinese research groups are working on the development of a high resolution crop mapping and a crop early warning to be combined with the pest/diseases monitoring model to accurately estimate its possible effect in terms of production loss and food security. To this purpose, the activity carried out in the framework of Dragon 5 will take into account results coming from other initiatives in the context of the GEO work programme (CROP_PEST_MONITORING and AFRICULTURES). High resolution (10m) crop mapping based on Sentinel-2 image series methodology has been developed and it will be integrated with pest/diseases monitoring model in order to be able to forecast/estimate the potential affected area.

For desert locust presence risk forecasting, we have proposed a dynamic prediction method of desert locust presence risk at Somalia-Ethiopia-Kenya. First, we obtained six static environmental factors, including DEM, land cover, soil sand content, clay content, silt content, and coarse fragment content. Then the time lag variables of four dynamic environmental factors, namely rainfall, soil moisture, NDVI, and LST, were then extracted based on multi-source time-series data. Finally, we have constructed a dynamic remote sensing-based risk forecast model of desert locust presence combining a multivariate time lag sliding window technique and machine learning algorithms. Monthly prediction experiments from February to December 2020 were then conducted, extracting high, medium and low risk areas of desert locust occurrence in the study area. Results demonstrated that the overall accuracy was 77.46%, and the model enables daily dynamic forecasting of desert locust risk up to 16 days in advance, providing early warning and decision support for preventive ground control measures for the desert locust.

The project, therefore, at the Mid Term event, is well aligned to the timetable described at the time of the submission and most of the activities have been started. Moreover, the PhD candidate, selected within the YS program, is working on the project using common data.

The recognition of intra-urban heat islands depends on the capacity to identify, at varying temporal and spatial scales, of variations in urban form and urban fabric. In this paper Earth Observation is used to define Land Surface Temperature, recognize intra-urban heat islands and identify the impact of urban form/fabric characteristics [land cover (greenery), the impervious fraction, the building density and height and the type of construction materials) to surface thermal dynamics. Particular attention is given to the relation between high/medium/low density areas and LST on the one hand, and on the cooling potential of greenery (in the above areas) on the other, in view of defining urban patterns to be taken into consideration once adaptation plans to urban heat are developed. Results are examined against spatial resolution, time of day and season, whereas special attention is given to the impact of climate extremes, namely heat waves, to any recognized patterns as well as to the presence of intra-urban heat islands. Landscape metrics are also used in support of the extraction of urban patterns and the understanding of their links to surface thermal dynamics. Finally, results are rolled out as a climate service for Climate friendly and smart cities.

Urbanization is continuing at an unprecedented rate in many cities across globe. Rapid urbanization poses significant social and environmental challenges, including sprawling informal settlements, increased pollution, urban heat island, loss of biodiversity and ecosystem services, and making cities more vulnerable to disasters. Therefore, timely and accurate information on urban changing patterns is of crucial importance to support sustainable and resilient urban planning and monitoring of the UN 2030 Urban Sustainable Development Goal (SDG). The overall objective of this project is to develop innovative, robust and globally applicable methods, based on Earth observation (EO) big data and AI, for urban land cover mapping and urbanization monitoring.

Using ESA Sentinel-1 SAR, Sentinel-2 MSI and Chinese GaoFen-1 images, the EO-AI4Urban team has developed varous deep learning-based methods for urban mapping and change detection during the past two year. For urban mapping, a novel Domain Adaptation (DA) approach using semi-supervised learning has been developed for urban extraction. The DA approach jointly exploits Sentinel-1 SAR and Sentinel-2 MSI data to improve across-region generalization for built-up area mapping [1]. For urban change detection, several novel methods have been developed including a dual-stream U-Net [2] and a Siamese Difference Dual-Task network with Multi-Modal Consistency Regularization [3]. Further, a high-resolution feature difference attention network (HDANet) is proposed to detect changes using the Siamese network structure. To tackle the loss of the spatial features of buildings caused by the multiple successive down-sampling operations using fully convolutional networks (FCNs), a multi-resolution parallel structure is introduced in HDANet, and the image information with different resolutions is comprehensively employed Using Landsat time series, a machine learning method was developed to map built-up areas and to analyze changes during 2000 to 2020 [4]. In addition, techniques for the characterization of urban changes exploiting sequences of SAR and/or optical data, with focus on Chinese cities are further developed. The temporal fitting of the coherence sequence using Sentinel-1 SAR data has been explored. This method has shown potential to discriminate and label accordingly the changes in time for intra-urban areas [5]. Another approach has been developed to focus on an automatic labeling of these temporal sequences based on a statistical analysis of the changes in time, clustered into different behaviors (e. g., stable clusters, fast changes, etc.). This approach has been tested in Chinese cities and achieved significant results [6]. In consideration of the difficulties to find an object of interest in large-scale scenes, a deep learning-based visual grounding method in remote sensing images has been developed to automatically locate a target object from large-scale scenes by a language description [7]. A robust cloud detection approach has also been developed [8]. Finally, using Landsat time series, a machine learning method was developed to map built-up areas and to analyze changes during 2000 to 2020 [9].

Experiments conducted on a test set comprised of sixty representative sites across the world showed that the proposed DA approach achieves strong improvements upon fully supervised learning from Sentinel-1 SAR data, Sentinel-2 MSI data and their input-level fusion. The fusion DA offers great potential to be adapted to produce easily updateable human settlements maps at a global scale [1]. Using the Onera Satellite Change Detection (OSCD) dataset, the results showed that the dual-stream U-Net outperformed other U-Net-based approaches together with SAR or optical data and feature level fusion of SAR and optical data [2]. Using bi-temporal SAR and MSI image pairs as input, the Siamese Difference Dual-Task network with Multi-Modal Consistency Regularization have been tested in the 60 sites of the SpaceNet7 dataset. The method achieved higher F1 score than that of several supervised models when applied to the sites located outside of the source domain [3]. Using the LEVIR-CD dataset, the results confirmed that HDANet can improve the differential feature representation for building change detection [4].

Using Landsat time series, the results show that machine learning method could extract built-up areas effectively. To analyze urbanization in 13 cities in the Beijing–Tianjin–Hebei region, SDG indicator 11.3.1, the ratio of land consumption rate to population growth rate (LCRPGR) is calculated and the results show that the LCRPGR in Beijing–Tianjin–Hebei region fluctuated from 1.142 in 2000–2005 to 0.946 in 2005–2010, 2.232 in 2010–2015 and 1.538 in 2015–2020. Apart from the megacities of Beijing and Tianjin, the LCRPGR values were greater than 2 in all cities in the region after 2010, indicating inefficient urban land use [9].

References:

[1] Hafner, S., Y. Ban and A. Nascetti, 2022. Unsupervised Domain Adaptation for Global Urban Extraction Using Sentinel-1 and Sentinel-2 Data. Remote Sensing of Environment (in press).

[2] Hafner, S., A. Nascetti, H. Azizpour and Y. Ban, 2022. Sentinel-1 and Sentinel- 2 Data Fusion for Urban Change Detection Using a Dual Stream U-Net. IEEE Geoscience and Remote Sensing Letters, vol. 19, pp. 1-5.

[3] Hafner, S., Y. Ban and A. Nascetti, 2022. Multi-Modal Consistency Regular- ization Using Sentinel-1/2 Data for Urban Change Detection. Submitted to ISPRS Journal of Photogrammetry and Remote Sensing.

[4] Wang, X., J. Du, K. Tan et al., 2022. A High-Resolution Feature Difference Attention Network for Change Detection. To be submitted.

[5] Che, M. and P. Gamba, Temporal and Spatial Change Pattern Recognition by Means of Sentinel-1 SAR Time-Series. IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, 2020, pp. 160-163.

[6] Che, M., A. Vizziello, P. Gamba, "Spatio-temporal Change Mapping with Coherence Time-Series", submitted to IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.

[7] Sun, Y., S. Feng, X. Li, Y. Ye, J. Kang, X. Huang. 2022. Visual Grounding in Remote Sensing Images. Proceedings of the ACM International Conference on Multimedia, 2022.

[8] Li, X., X. Yang, X. Li, S. Lu, Y. Ye, Y. Ban. 2022. GCDB-UNet: A novel robust cloud detection approach for remote sensing images. Knowledge-Based Systems, 238: 107890.

[9] Zhou, M., Lu, L., Guo, H., Weng, Q., Cao, S., Zhang, S., & Li, Q. (2021). Urban Sprawl and Changes in Land-Use Efficiency in the Beijing–Tianjin–Hebei Region, China from 2000 to 2020: A Spatiotemporal Analysis Using Earth Observation Data. Remote Sensing, 13(15).