Potential for mapping the world’s land resources using satellites and artificial intelligence: an Australian case study in land use.
The Queensland Government is using machine learning and computer vision to automatically map and classify land use features in satellite imagery. Successfully applied to the mapping of banana plantations, the method is extremely efficient compared to current methods of mapping compilation. Using this technology the Queensland Government can accurately map and classify the land use in a timely manner, aiding response to biosecurity and natural disaster events.
The Queensland Land Use Mapping Program (QLUMP) has been mapping and assessing land use patterns and changes throughout Queensland, Australia for twenty years. Although many advances have been made automating processes, the main methodology of QLUMP requires manual interpretation of imagery and ancillary data at the desktop, hand digitising features to map on-screen and classification of land use features by a team of skilled staff. Queensland has an area of 1.7 million square kilometres (seven times the size of Great Britain and two and a half the size of Texas) and as a result, this process is time and resource intensive. To map land use for the whole state would take one person approximately 30 years to complete.
Machine learning as a sub-discipline of artificial intelligence has in recent years progressed with enhanced computing power so that methods of computer vision and deep learning in image analysis and classification approaches are now viable. Machine learning algorithms allow for computers to train on data inputs and use statistical analysis in order to output values that fall within a specific range. For these reasons, machine learning facilitates the building of models from existing land use data in order to predict the land uses of a new image.
Following these principles, we used existing land use mapping of banana plantations to test the capability of computer vision to map land use features as a proof of concept. We trained the computer to map banana plantations in the Johnstone catchment in north Queensland and used the model to predict banana plantation in another area, the Tully catchment. The results were very encouraging with an overall mapping accuracy of 97%, well above the 80% requirement.
With a trained model we can now readily update banana plantation mapping and rapidly respond to critical events such as natural disasters and biosecurity incidents - Queensland is threatened episodically by tropical cyclones and recently the viability of the local banana industry has been threatened by the detection of a very damaging fungus disease, Panama Tropical Race 4.
Future work will expand this method into other land use classes in an attempt to fully or semi-automate land use mapping in Queensland. Using the manual method, it took one person approximately six weeks to map all land uses in the Tully catchment, including banana plantations. Running the trained model on the same area to map banana plantations took a matter of minutes. With a fully trained model consisting of all land uses, we estimate we could reduce the state wide mapping of land use from 30 years to less than a year.
The proof of concept using machine learning and computer vision on existing land use data was highly successful for banana plantations and it is anticipated to be just as successful for other land use classes. This is a paradigm shift for the way land use is mapped from the traditional manual hand digitising to full or semi-automation.
A rapid approach to mapping land use, land use change, land management practices and the impact of disasters such as floods, fires, tsunamis, typhoons, hurricanes and tropical cyclones would be a major benefit to the developing world and disaster zones. With machine learning in fusion with high performance supercomputing, cloud computing, big data technologies, and readily available global satellite imagery from international sources like the European Copernicus and the US Geological Survey programs combined with growing commercial providers, a new era in land resource mapping can be envisaged.
What Makes Your Project Innovative?
Traditional remote sensing classification approaches use the colours (spectral bands) of individual image pixels or dots to determine what the feature is. The problem is that different features can have similar colours such as banana plantations and rain forests or even grasslands as they are all green. In interpreting land use, humans also take into account the context of the whole image so we can see that the banana plants are planted in rows, near roads and packing sheds and have a distinct leaf texture. This is why using computer vision has been so successful; the computer learns about all the components which make up a banana plantation, not just the colour.
Traditional approaches are also resource intensive, the benefits of computer vision are their capacity to process big data, which is increasingly being acquired at ever greater spatial and temporal resolutions, to output extremely timely land use information.
What is the current status of your innovation?
As of October 2018, we are writing a paper on the proof of concept with the hope it will be peer reviewed and published to aid the scientific community. We are also seeking to operationalise the method into the land use mapping program and extend its capability to inform other land use classes. Initial work has commenced to automate sugar cane crop mapping with promising results.
We are looking for opportunities to work with other states and territories of Australia to assist with their mapping programs and international collaborations to use these technologies to assist other countries to map their features of interest.
Collaborations & Partnerships
The Queensland Government initially collaborated with a private company, Envista, in the initial proof of concept stage. Envista assisted with their computer vision experience to accelerate our learning in this area and introduced us to a new state-of-the art algorithm.
We have also started initial collaborations with the New South Wales Government and the Australian Federal Government to apply our methods to other parts of Australia.
Users, Stakeholders & Beneficiaries
Other Queensland Government agencies, industry groups and catchment managers can now can access an updated banana plantation map. This is of great benefit to the current biosecurity response and containment incident affecting the banana industry in Queensland.
The Queensland Government is part of the Australian Collaborative Land Use and Management Program so other states and territories will be able to access this technology and apply it to their own land use mapping programs.
Results, Outcomes & Impacts
Using a stratified random sample of points throughout the catchment, the accuracy assessment of banana and non-banana plantations was 97% accurate. Assessing the banana plantations separately from other land use features we found the model was 80% accurate. The main confusion the model had was separating banana plantation from other tree fruit crops such as paw paws. Further training of the model using additional data will likely resolve some of these issues and increase the accuracy. We anticipate expanding the methodology to other land use classes to create a fully or semi automated land use map of Queensland. This could potentially reduce the time it takes one person to map the state of Queensland from 30 years to less than a year.
Challenges and Failures
The biggest challenge was the lack of literature in this application. This is a new areas science and limited research has been conducted.
Managing expectations of colleagues and other interested parties was also a challenge. These methods do not work without training and training data. I had to manage the expectation held by some colleagues that this technology can map anything and everything automatically.
Securing on-going funding to continue this research and to operationalise and integrate this method into the existing land use mapping program has been a challenge. Although this is cutting edge science producing encouraging results, the current economic environment has restricted the available resources to continue this work.
Conditions for Success
This work is computationally intensive and high end Graphics Processing Units (GPUs) are required to process the data. We found using a GPU (Nvidia Tesla P100) was 70 times faster than using the computers Central Processing Unit (CPU) (Intel Xeon).
Support from management and securing resourcing to conduct research into new methods is imperative.
The results have been encouraging to further pursue the work, the potential efficiencies to expand the computer vision into other land use classes is real.
These methods have started to be replicated to other applications such as mapping sugar cane plantations and woody vegetation. It is anticipated these methods could be used to map other features in satellite imagery and aerial photography such as fire scar mapping, woody vegetation change detection and monitoring of mining and coal seam gas well infrastructure.
Computer vision has the capability to greatly assist the efficiency of compiling mapping to support large-area landscape management and monitoring programs, in support of natural resource management and monitoring by government and non-government and aid organisations.
As Queensland is part of the Australian Collaborative Land Use and Management Program (ACLUMP), this work was presented to all other states and territories at the 2018 ACLUMP workshop. This research has generated interest across the country, particularly in states who do not have enough funding to conduct a state wide land use program.
Data preparation and quality training data are required for an accurate model. Although we already had quality land use mapping data, some errors were identified and fixed to produce quality outputs.
Communication to promote the science, including tailored content suited to both technical and non-technical audiences was vital to promote the work and for support to continue.
Comprehensive field survey data has strengthened the validation of our results. Our research team is fortunate to have access to high performance computing resources and substantial archives of satellite imagery. The choice of machine learning algorithm has been critical for our project.