Project description for website
What it does
LandPro (Land Productivity monitoring app) is a ML based interface that harnesses the power of earth observation satellites, translating the data in relevant information for farmers, land managers and more.
LandPro’ workflow includes 4 steps: 1- The user indicates the Area of Interest through a simple web interface 2- We retrieve relevant data from Earth Observation satellite image repositories, including: Multispectral images (SENTINEL 2 satellites) NDVI time series Soil organic carbon estimation 3- A Deep learning model for semantic segmentation divides the area into subpolygones with homogenous land cover 4- For each subpolygone two co2 estimates are calculated :
- Vegetation co2 absorption is based on the NDVI, and is calculated as the integral of the NDVI for the last year
- Soil organic carbon (SOC) stock is a vital component of soil with important effects on the functioning of terrestrial ecosystems. Its values were obtained from Lugato, Emanuele, et al. "A new baseline of organic carbon stock in European agricultural soils using a modelling approach." Global change biology 20.1 (2014): 313-326.
All this data is then displayed for the end user through a simple & clear map based interface.
LandPro can help self-evaluation by farmers, comparing different managements of cropland and extracting trends and other significant variations.
Where is my crop growing well? Which part of my park should I restore? What are the damages of the last fire/drought/flood? Which part of this forest should be protected? Every day, farmers, construction companies, forest and park managers, environmental operators, environmental conscious tourists ask these kinds of questions, but the answers requires long-term knowledge of the area and deep understanding of the ecosystem. Satellite image can provide us the insights we need by looking at the past and comparing different areas. This landscape approach to satellite-based information system is based on published and peer reviewed scientific research by team members. See for example: Jucker Riva, M., Daliakopoulos, I. N., Eckert, S., Hodel, E., & Liniger, H. (2017). Assessment of land degradation in Mediterranean forests and grazing lands using a landscape unit approach and the normalized difference vegetation index. Applied Geography, 86, 8–21. https://doi.org/10.1016/j.apgeog.2017.06.017
How we built it
CO2 Metrics SOC estimation: SOC data of Europe in the form of tif file was obtained from JOINT RESEARCH CENTRE EUROPEAN SOIL DATA CENTRE (ESDAC), based on Lugato, Emanuele, et al. "A new baseline of organic carbon stock in European agricultural soils using a modelling approach." Global change biology 20.1 (2014): 313-326. The polygon area of selection in the form of a geojson file was loaded. Based on the polygon area of selection, using this as a mask over the SOC data, the SOC area of interest was clipped. The mean of SOC data was then calculated for the polygon. NDVI time series:
Challenges we ran into
There is little data available on land classification, especially we have missing a dataset with masks that can be used efficiently for machine learning. Data about soil is scattered across a thousand different databases with no homogeneity!
Accomplishments that we're proud of
The whole structure of the app was built in 5 days!!! Without even meeting each other in real life!
What we learned
Matteo has learned how to connect Google Earth Engine with python and tensorflow to run cloud based deep learning models. Also, it is the first time the overall structure of my app is not a mess ;)
What's next for LandPRO
Currently LandPro is a simple interface to harness data from different sources in a meaningful way. In time LandPro can become a reference for validating self-assessment, sharing successes and improve adoption of positive practices among farmers, land managers and more.
Better estimates of co2 absorptions /emissions can be calculated combining land cover type with multispectral images. Several scientific papers have been published on the topic, and we expect methods and models to improve greatly in the near future
More interactivity: modifying polygones boundaries, adding spatial elements, describing land management practices each subareas are functionality that can be easily integrated within the user interface
Visualisation of co2 changes in time: What is the trend, have there been major changes in the recent past. SOC estimates could be updated real time based on sentinel images and latest sampling data.
Comparison with neighbouring areas: LandPro can easily provide information not only on the area of interest entered by the user, but also for the neighbouring similar areas. This could provide the user of an estimates of its performance, but also suggest locations where better practices may be found
Technical improvements: Using tiled data for spatial information would reduce memory usage and speed up the application Larger training of deep learning models would reduce dependencies from external services