1 of 19

Towards automation of river water surface detection

Stefano Conversi

July 4^th 2024

Politecnico di Milano – Regione Lombardia

Co-authors:

Daniela Carrion, Francesco Gioia, Alessandra Norcini, Monica Riva

2 of 19

Introduction

Remote Sensing for river monitoring

Integrating optical and radar imagery to enhance river drought monitoring

How to optimize the process?

Towards process automation

Sensitivity analysis and best ML algorithm selection

Conclusions

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

3 of 19

In 2022 Southern Europe suffered one of the most severe drought period of the last century, as shown by Combined Drought Indicator developed by

European Drought Observatory (EDO).

Introduction

A major drought event: Europe 2022

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

4 of 19

Northern Italy was harshly affected, with Piedmont and Lombardy regions particularly impacted.

EDO’s March 2022 Combined Drought Indicator

map for Northern Italy

Po River basin, from Autorità di Bacino Distrettuale del Fiume Po

Introduction

A major drought event: Europe 2022

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

5 of 19

Impacts of drought on territories and inland water systems:

1/3 decrement of national agricultural production (June 2022)
Navigation restrictions in several Po River canals – losses in connected economic sectors (e.g. tourism)
Potential drinking-water rationing
Competent bodies forced to reduce water discharges from artificial reservoirs – hydropower plants affected
Possible impacts on river ecosystem and local biodiversity depletion.

Regional decrees for promoting campaigns of prevention and mitigation of the effects of water scarcity

(Regione Lombardia, DGR April 13th 2022, n. XI/6283)

Necessity of innovative tools for supporting public administration for visualizing and monitoring drought conditions for medium-width river

Assumption: Water surface extent variation as a parameter for estimating drought conditions

Proposed methodology to exploit satellite imagery for mapping water coverage,

by the mean of optical and radar sensors integration.

Introduction

A major drought event: Europe 2022

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

6 of 19

Explore response on several different visible and infrared wavelength intervals
Obtain spectral indices by combining specific bands responses, thus highlighting features of interest (in this case water)
Photointerpretation and indices thresholding
Being an optical source, they are affected by the presence of clouds.

Optical sensors

Radar sensors (SAR)

Record values of backscatter signal of emitted microwaves
Smooth water behaves as a specular surface: clearly distinguishable response, both in values and in appearance (dark pixels)
Thresholding methods
Prone to possible misclassifications:

Roughness of surfaces (e.g. presence of waves on water)
Double bounce and volume scattering (presence of objects in water)
Presence of waterlike surfaces
Background noise (e.g. soil moisture).

Remote Sensing techniques are widely employed in water detection projects, especially in the field of flood-risk monitoring and flooded areas delineation.

Pixels are classified as belonging to a certain class (water/non water) depending on the response recorded by the exploited data source.

Remote Sensing for river monitoring Optical and radar sensors

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

7 of 19

Being an optical source, they are affected by the presence of clouds – missing infomation.

Optical sensors

Radar sensors (SAR)

Prone to possible misclassifications:

Background noise (e.g. soil moisture) – overestimation.

Remote Sensing for river monitoring Optical and radar sensors

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

8 of 19

Sentinel-1 GRD

Sentinel-2 L1C

Speckle filtered

VV band

NDWI

map

SWM

map

VV thresholding

water map

SWM thresholding

water map

Supervised

Random Forest classification

Polygons

training set

Hydrometric

monitoring

system

Classified

water map

Validation

Integrating optical and radar imagery to enhance river drought monitoring

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

9 of 19

Spectral indices evaluation (from S-2):

Normalized Water Index (NDWI)
Sentinel Water Mask (SWM, Milczarek et al., 2017)

Reference water mask maps:

S-1 filtered VV thresholding at -18 dB
SWM thresholding at 1.4.

Supervised classification (integrating optical and radar imagery)

Random Forest algorithm with 10 decision trees
Stacking filtered SAR VV band, NDWI and SWM for obtaining input image
Drawing of a set of polygons deemed as representative of the two classes

Water/Non water
By RGB photointerpretation
Split: 80% train set, 20% test set.

Validation through Total Positive Ratio on a set of 5000 randomly extracted points.

Thematic map of water coverage

Estimate of water surface in a buffer of 600 m from the river vector

Algorithm run

Integrating optical and radar imagery to enhance river drought monitoring

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

10 of 19

Integrating optical and radar imagery to enhance river drought monitoring

Period	Water surface [km²]	TPR [%]	Level Spessa Po [m]	Level Piacenza [m]
2023/5/26-31	19.25	96.8	57.982	56.596

Period	Water surface [km²]	TPR [%]	Level Spessa Po [m]	Level Piacenza [m]
2022/8/29-31	12.53	86.4	55.697	54.246

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

11 of 19

How to optimize the process?

River surface mapping tool to be integrated in public administration procedures

Personnel with no mastery in Remote Sensing or GIS environments:

Avoid training polygons drawn by users
Automate the process a much as possible

Provide the service through an user friendly WebApp

The system must be capable of achieving autonomously reliable water/non water masks to be used in classification algorithms training

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

12 of 19

Otsu algorithm (bimodal histogram images, distinction background and foreground)

Automatic thresholding:

Towards process automation

Otsu algorithm

Otsu algorithm may not work for segmentation of images requiring more than two classes or for complex backgrounds (as in this case)

Bmax Otsu algorithm (Markert et al., 2020).

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

13 of 19

Bmax Otsu algorithms applies the classic method to non-bimodal images (thus working for all the considered satellite input)

Towards process automation

Bmax Otsu algorithm

Automatic thresholding based on Otsu algorithm:

Each image is subdivided in cells of user-defined dimensions (chessboard segmentation)

Cells go trough a bimodality test (estimates maximum normalized Between-Class Variance)
Otsu algorithm for image thresholding is applied only to the bimodal cells

An overall threshold for the image is evaluated.

Grid size =

River width [m] ⋅ 0.008/250

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

14 of 19

Goal: Automatic Training/test set extraction

Representative pixels should be extracted from areas that are certainly containing water/non water

Each input band (VV, NDWI, SWM) is thresholded by the means of Bmax Otsu algorithm

Three rough classified maps («masks») of water/non water are obtained

Which pixels are in common for the three masks?

Among these, which pixels remain consistently classified in the reference period?

A new reliable mask of water/non water pixels for the reference period is obtained combining the pixels satisfying the queries

Training and test set are randomly extracted from the combined mask, considering a fixed percentage of pixels (calibrated: 0.15%).

500 m

Common water areas

Common non-water areas

Training points water

Training points non-water

Towards process automation

Training/Test set extraction

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

15 of 19

Sensitivity analysis

Calibration of Training Points amount

Accuracy estimated on the Spessa Po – Piacenza study area, on three reference periods, validated against a set of known photointerpreted pixels

Analysis performed comparing results for different ML classification algorithms:

Random Forest (RF)
Classification And Regression Tree (CART)
Support Vector Machine

Defined result: 0.15% of the total pixel amount in the area.

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

16 of 19

Best performing

ML classification algorithm

New study area (Po River, from Borgoforte, Mantova, to Casalmaggiore, Cremona)
Tested algorithms: RF, CART SVM
Tested for single sensors and for their integration
5 epochs compared
Intercomparison of results with the outcomes of supervised RF «original method».

Optimal performing solution:

S1 + S2 integration, through SVM-based pixel classification.

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

17 of 19

The proposed integration allows to overcome typical limitations of the single sensors

The methodology is in principle repeatable with whichever optical/radar couple of sensors

Validation showed good results in water surface detection

Multitemporal analysis proved the applicability of the methodology in a wide range of conditions (especially extreme ones)

The correlation between water surface and hydrometric level corroborates results

The whole process is strongly dependent on data (imagery) quality and availability

The methodology is efficient for short-medium time monitoring, at the moment not suitable for real-time

Google Earth Engine is a proprietary platform: public administrations should establish agreements.

Conclusions

Key points

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

18 of 19

Additional code refinement

Enhancement of time window selection, and reduction of user-defined parameters

New case studies analysis

Further validation of the methodology, limitations due to river morphology or geographical context

Final WebApp production

Product to be delivered for Regione Lombardia’s users.

Goal: enhance PA capabilities of mitigating CC-related risks.

Conclusions

Expected developments

Conversi, S., Carrion, D., Gioia, F., Norcini, A., and Riva, M.: Towards automation of river water surface detection, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLVIII-4/W12-2024, 19–27, https://doi.org/10.5194/isprs-archives-XLVIII-4-W12-2024-19-2024, 2024.

Towards automation of river water surface detection

Stefano Conversi

July 4^th2024

19 of 19

Thanks for your attention!

stefano.conversi@polimi.it

Integrating optical and radar imagery to enhance river drought monitoring

Towards automation of river water surface detection

Google Earth Engine code