The colour image of the area of interest, which covers approximately 12k square miles. The area is South Kalimantan Indonesia.
The importance of each variable in the model. The created variables designed to detect Vegetation (TNDVI) is the most important
The map with land use classified by the model. Light Green is Palm Plantation, Dark green is Forest, White is Cloud, Blue is water.
Plantations from the SPOTT database overlaid on the classified map
Calculated Values based on SPOTT data. area in Km^2. According to data 47 percent of Seruyan's output is Conflict
To prevent further deforestation in tropical rainforest for the production of palm oil. Focusing on the early detection of land use in conflict areas. The Spectral analysis provides a clear way of detect changes in land which can be investigated and checked against certified land. This allows stake holders easy access to data to check new farms are not in a conflict area. This can aid registration to RSPO or prevent. information can be shared with manufactures and the wider public to reassure them in the sustainability in the Palm Oil product.
What it does:
The system has two components, which work together for enhanced sensitivity to changes in land use.
- Change detection for early intervention in vulnerable sites.
- Land classification to monitor land use and identify conflict palm areas.
Change Detection Uses spectral analysis to detect changes to satellite imaging in conflict areas. Using open source API's to detect changes in land use, and checking against data of certified areas. Using machine learning to highlight possible conflict zones to be highlighted. Web app receives updates from server displaying polygon to show location of land change. Images update every 5 days and one pixel represents 10m.
Land Classification A secondary system classifies Landsat 8 images by area class, these classes are Forest, Plantation, Water, Cloud. This classification allows the detection of conflict palm by comparing the areas classified as Plantation with the SPOTT database of Oil Palm plantations. This allows the model to estimate the legal and illegal output of each administrative region in indonesia as well as it's value.
How it was built:
Due to there different functions the models are built using different systems the results of which are then integrated.
Change Detection using google icloud platform, debian8 and squl server. Python backend with a google map API interface. Data from Sentinel2 MSI (multi spectral interface ) using 2,3,and 4 correspondence to re, green and blue band width bulian cirrus data. Python back end with opencv API implementing a function which subtracts last Sentinel-2 as a baseline. update data reference to change of potential conflict areas are identified by opencv subtract method. subtract baseline from updated data to give geospatial sat of polygons (defence) corresponding to area of potential conflict. exported as a vector to google map API. stored as a squl.
Land Classification The area class classifier was built using TIFF files from landsat 8, the area in the test is south Kalimantan, specifically the administrative regions of Seruyan and Kotawaringin Timur. The Tiff files are loaded into R along with shape files defining areas of Water, Cloud, Plantation and Forest, a multilevel XGboost model is then trained and tested, before being deployed to classify the entire map. The resulting classifier is then replotted as a map. Once done the map is compared with known plantations and conflict palm identified
The biggest challenge is time. the ability to create a working model in such a short time frame. Learning to work with TIFF files and Geospatial data was challenging especially, given the size and moving them around in the cloud.
Accomplishments that we are proud of:
The system provides an early detection , especially useful in hard to reach areas that could go unidentified for long periods of time. Thus helping to maintain the biodiversity of the local area.
The predictive model is very accurate having prediction accuracies in the high 90's. The ability to convert the observed oil palms into a dollar value is also helpful to help non-specialists appreciate the size of the issue. Incorporating SPOTT data to quantify conflict is very helpful and allowed us to make statement about the state of conflict palm based on real data as shown in the table below.
|2||Total Plantation (Km^2)||1426|
|3||Legal Plantation (Km^2)||749|
|4||Conflict Plantation (Km^2)||677|
|5||Total Forest (Km^2)||1157|
|6||Estimated Production (Tonnes)||526194|
|8||Production Conflict (Tonnes)||247311|
|9||Value Conflict (MUSD)||171|
What we learned :
How satellite imagery can be used to identify land. That changes can be identified on a regular time intervals. The extent of current deforestation levels and the need for action. Time - creating an idea and concept takes longer than a couple of days.
What's next for Reading Palms?
The combination of change detection and land classification systems reduces computational load so improving scalability. As a result the system can be used to identify any changes in land mass use, across the globe. This can include other conflict areas, or to track progress of reforestation projects.