↓ 

While a lot of the recent machine learning methods have focused on reproducing global weather reanalysis and forecast models, a discrepensasy exists between the global models and local weather stations. This project aims at downscaling global weather models to local scales, and even arbitrary locations, by combining global reanalysis data with local weather station data.



To tackle this task, we first focus on forecasting the weather at fixed locations. We developed a series of deep learning models that take as input reanalysis data (ERA5/HRRR) and local weather station data (MADIS), and output a forecast for the next 24 hours at the location of the weather station.



The results show that the models are able to accurately forecast the weather at the location of the weather station.
Our current focus lies in extending the model to arbitrary locations, by integrating the global reanalysis data with local weather station data, as well as satellite imagery.

↓ 

By combining historical MODIS satellite images with modern Sentinel-1 images, we produce reliable return period estimates.


The proposed framework takes advantage of the ability of radar to "see" water even through clouds, one of the major challenges of flood mapping.

By crating a model able to fuse MODIS 8-days composite images with Sentinel-1 derived fraction of flooded area, we infer a 20 years long historical time series over Bangladesh based on the 20 years long MODIS time series.

The data to train, test and validate the model is comprised of every 32x32 500 meters resolution chip where MODIS 8days composite and Sentinel-1 fully overlap. Sentinel-1 is available since 2017.

The proposed model takes advantage of both the spatial and temporal information contained in MODIS images.

For each time step, the n previous time steps are considered to make a prediction on the fraction of flooded area, leveraging the temporal dynamics of floods.

Each image first goes through a Convolutional Neural Network (CNN), extracting the spatial information and compiling the multi bands information into a single value for each pixel.

This generates a series of values for each pixel, which is then passed through a Long Short Term Memory (LSTM) network designed to, through a series of gates, predict the probability of flood to be present.

Finally, the output of the LSTM is combined with the MODIS composite image at time t and passed through a final CNN, generating a prediction for the fraction of flooded area for each pixel in the image.




For the testing, a single year (2018) is completely removed from the dataset.

The result of the developed fusion model is shown here for a single day (Mai 5 2020) on all available chips for this day, and as a time series for all of Bangladesh for all dates where chips are present, as examples of inference on the testing dataset.

The var R² of 0.66 shows a good agreement between the observed and modelled data.

The time series shows that the low water values during the dry season are well modeled, and the peaks during the monsoon and irrigation season correctly reproduced.



This result permits us to run inference on the historical data in order to produce an estimate for 2001-2020.


The yearly maximum of the inferred historical data is compared to another MODIS algorithm, the Global Flood Database algorithm (GFD) for the same area and period.

The developed methods shows similar trends with GFD, but a more realistic yearly maximum according to the time series analysis above.

This inferred data permits to construct more realistic return period estimates.

↓ 

One of the main outputs of this work is to publish a publicly available high resolution flood dataset. The selected flood events are based on existing flood event datasets (xBD, Sen1FLoods11, NASA IMPACT, FloodPlanet (in house)), in order to be able to compare with other lower resolution datasets, but also label a few new events.







In the context of this work, I have mainly been involved in writing a pipeline to download and process PlanetScope scenes into chips, from which we sample a representative subset which is then labelled (c.f. figure above).

The events have been selected to coincide in time with available Harmonized Landsat Sentinel-2 (HLS) or Sentinel-1 data, such that deep learning models can be developed to compare the commercial vs public data source.

For each event, the selected chips have been hand labelled with three classes: no water, low confidence water, high confidence water (c.f. figure).

The data processing and labelling of all events has been completed, and we are currently working on releasing the dataset along with baseline model results (deep learning both on public and commercial data).

↓ 

This project is done in collaboration with the Pima County Government, where we are currently assessing the utility of satellite imagery to detect floods in the desert, in particular the advantage of high frequency PlanetScope commercial data.


Flash floods in the desert during the monsoon season recess rapidly, and capturing them with satellite images appears difficult.

PlanetScope images, with their near global and near daily images, provide an interesting product to detect the differences between consecutive images, as often only wet soil or sand remains as witness of the flood itself.

In this context, I have developed a Google Earth Engine Applet to visualize the daily precipitation over Pima county, in an effort to aid the selection of flood events to be modelled, and hint towards their location in space (c.f. figure).

Additionally, for another project, in collaboration with a PhD student, Alex Saunders, we developed a tool to map the available satellite images around a specific date. This is particularly useful here as it allows us to select flood events where satellite imagery has a potential (c.f. figure).






I have contributed to hand labeling flood events and developed machine learning algorithms to classify the wet sand we observed.

One of the promising methods we developed is based on selecting two consecutive PlanetScope images, one on the morning of the event, and one on the day after.


This permits to identify the changes in the landscape, and seems to be most promising with the algorithms we developed.

↓ 

In order to accurately model these species, it is of crucial importance to understand their drivers for presence and dynamics.



The goal of this study is to develop an applied metapopulation framework to reproduce observed dynamics of mountain species in space and time, driven by Earth Observation and calibrated on Insitu Observations.













A metapopulation is commonly described as a population of populations of a single species , i.e. a population comprised of multiple subpopulations, called local populations inhabiting the same landscape, but in distinct geographical locations. The metapopulation concept is an abstraction of the concept of population, where an ensemble of interacting individuals is considered to a higher level, and where interacting subpopulations are considered.

The specific metapopulation model used here is called Spatially-Explicit Stochastic Patch Occupancy Model (SPOM).

Building on a grid with N cells, SPOM computes a possible distribution of occupied cells at every simulation time t by considering extinction and colonisation processes, whose rates depend on the species properties and on the landscape features. A binary state variable w_i (t) is set to 1 when the cell i is occupied and 0 when empty (i = 1, ... , N).

Starting from a given initial distribution of occupied cells, at each time step, the model allows unoccupied cells to be colonised by surrounding occupied cells with a certain probablity.

Then, the cell becomes occupied at time t + ∆t depending on a random sample from a Bernoulli distribution. Similarly, species in occupied cells can go extinct with a certain probability. SPOM works as a Markov chain, where, for each cell, the probabilities of colonisation and extinction events are modelled depending on presence and conditions.



In the context of this work, the colonisation and extinction probabilities are based on using Earth Observation data to characterise the niche of the species.

The factors are combined in a non-linear logistic function to best describe the niche.

The different EO factors are selected to best describe the environmental conditions relevant for the modeled species.








The developed framework proposes a three step algorithm:
1. Estimation of the initial presence of the species in the landscape through a Species Distribution Model (SDM);
2. Calibration of the partially observed markov process (POMP) through iterated filtering (IF);
3. Simulation of the species in the landscape with the calibrated model (inference).



The iterated filtering algorithm allows data only partially observed in time to be accommodated by solely updating the calibration at time steps where the data is available.

This developed framework has been tested on two species of Carabids in the Gran Paradiso National Park (GPNP), where the presence has been modelled in space and time.


Given the process based nature of this model, the framework has the potential to propagate the presence of a species in space and time, and potentially inform on the effects of climate change.