Share
Embed
Identification of Wastewater CH4 Emission Sources with Computer Vision and Sentinel-2 Observations
Poster at the ESA Phi-week 2020
Carlos Alberto Gomez Gonzalez
October 02, 2020
Transcript
19 slides
-
Identification of Wastewater CH4 Emission Sources with Computer Vision and
Sentinel-2 Observations Carlos Alberto Gómez Gonzalez Gerard Aceves Soley Marc Guevara Vilardell Kim Serradell Maronda ESA Phi-week 2020 -
Background • CH4 is a greenhouse gas not as prevalent
in the atmosphere as CO2 but far more potent (IPCC report, 2019) • According to the Emissions Database for Global Atmospheric Research (EDGAR), a global emission inventory, the wastewater disposal sector accounted for 5 to 20% of the CH4 emitted in 2012 (Janssens-Maenhout et al., 2019) • Current emission inventories lack spatial representativeness due to the usage of the population density to allocate emissions from wastewater treatment plants (WWTP) EDGAR emissions dataset -
Aims of this study • Develop a Deep Learning-based methodology
for the identification and localization of Spanish CH4 emission sources, with a focus on WWTPs • Using European Space Agency images from the Sentinel-2 mission • Design a methodology that could be applied to other geographical areas where the identification of emission sources is more scarce • Enhance the spatial representativeness of current emission inventories/databases -
Aim of the project The Urban Waste Water Treatment database
(UWWT-DB) is used for extracting the location of the WWTPs in Spain -
Object detection Object Detection is one of the fundamental, and
therefore well studied, computer vision problems -
Object detection Li et al. 2019 Detecting objects on satellite
images differs from the task of object detection on natural images. Examples: -
TensorFlow object detection API • In this study we use
Tensorflow, an open source machine learning platform • It facilitates the construction, training and deployment of object detection models • Contains state-of-art models that can be fine-tuned to learn new classes of objects Two Object Detection models considered (both on top of the ResNet50 backend): 1. Faster R-CNN 640x640 resnet50. Ren et al. (2015) 2. EfficientDet 640x640 resnet50. Tan et al. (2019) -
Workflow Satellite images Training and validation sets (90%) Test set
(10%) 3x Model training Parameters: batch size step size learning rate l.r. decay epochs WWTP detector (fine-tuned object detection model with a new class) Model evaluation Object detection models pre-trained (COCO 2017 dataset) -
Sentinel-2 images • Red, Green, Blue bands at 10m resolution
• 100 km x 100 km images • Earth imagery every 5 days • The characteristic size of the WWTP (meters to tens of meters) presents a challenge to the usage of Sentinel-2 images -
Creating the dataset Sentinel-2 bands (R,G,B) Image of the WWTP
WWTP coordinates from the UWWTP-DB Extract an window of 1x1 km, join 3 bands Bounding Boxes Label manually each image (using LabelImg) lon lat -4.913.763.548.201.410 4.647.908.205.880.020 -1.171.405.003.801.900 4.415.578.654.792.910 -3.191.401.053.333.200 4.408.318.555.398.290 -1.867.400.114.640.220 4.135.515.700.517.950 . . . Image file Annotation file -
Sentinel 2 vs. Google Earth A Google Earth dataset was
created containing the same WWTPs for a comparison with the Sentinel-2 counterpart. • Left: Sentinel-2 image (10 m resolution) • Right: Google Earth image (1 m resolution) -
Predictions format The highest rated score prediction is chosen amongst
the multiple predictions 0: [x1 y1 x2 y2 ] score 0 1: [x1 y1 x2 y2 ] score 1 2: [x1 y1 x2 y2 ] score 2 3: ... -
Evaluation metrics Intersection Over Union (IOU) Usually, a valid prediction
corresponds to a IOU > 0.5 -
Evaluation metrics Precision - Recall curve mean Average Precision (mAP)
for different IOU thresholds -
Results • 500 WWTP samples in total (90% for training/validation
and 10% for testing) • mAP on the test set for IOU > 0.5, taking the mean over 3 runs with different train/test shuffles With the Sentinel-2 images lower detection performance is obtained. Still the results are decent considering the WWTP test the resolution limits of the instrument. mAP (Google Earth data) mAP (Sentinel-2 data) Faster R-CNN 0,79 0,62 EfficientDet 0,89 0,70 -
Prediction examples on Sentinel-2 images It would be extremely hard
to find the wastewater treatment plants by eye -
Conclusions and next steps • State-of-the-art object detection models give
promising results in localizing wastewater plants even at the 10m resolution of Sentinel-2 images • Expand the training dataset with more examples (from the UWWT-DB) • A simple analysis shows a relationship between the plant size and its water treatment capacity (Pearson correlation coefficient: 0,64) • Couple a regression model to predict the plant capacity • Use WWTP capacity as proxy for CH4 emissions to improve the existing methodologies for spatial distribution of emissions -
Thank you [email protected] This study has been possible thanks to
the BSC International Summer HPC Internship Programme and the work of Gerard Aceves Soley during his internship.