- Posted by MG UG
- On May 11, 2017
- 0 Comments
Deep Learning Neural Networks and Remote Sensing in Manitoba
The province of Manitoba is tasked with producing a land use/land cover (LU/LC) map derived from satellite image classification, for the province in a timely manner. However, due to a variety of constraints, meeting this deadline has proven challenging. In many cases it may take several years to produce a LU/LC map for the southern agricultural region of the province.
After many discussions with GeoManitoba, the agency tasked with producing this map, Drs. Christopher Storie (Geography) and Christopher Henry (Applied Computer Science) have proposed a new method for producing these maps. By using the previous three LULC maps produced and the original Landsat data used to produce these maps these faculty members are able to train a deep learning neural network to automate the classification of future datasets based upon the knowledge learned from the already classified data.
The process involves training a pre-defined neural network architecture that has been adapted to use satellite imagery. The adaptation requires the algorithm to handle six separate bands instead of the typical three (red, green and blue). However, this adaptation also increased the computational load of the training process as every calculation is being done six times. The result is that the neural network will be trained on an NVIDIA DIGITS system (an interactive deep learning GPU training system). By using this architecture the computational demand of the training is dramatically reduced as the training is an iterative process with the error being controlled until it reaches an acceptable level. Training of this nature can take days to weeks until the error becomes acceptable. Once the system is properly trained it can be “sealed” and used on a normal computer to classify future datasets automatically.
The project is underway and the original data sets have been prepared and are almost ready for the training states. Approximately 52,000 images that are 250 x 250 pixels in size, for the three study years have been generated and are ready for input. Once completed, the goal of the system is to produce LU/LC maps, once the data is ready, in much shorter period of time rather than the normal “weeks to months to years” it currently takes.