These researchers are trying to figure out a way to automatically identify buildings in satellite imagery without having to manually label every single pixel or use fancy algorithms that require a lot of computing power (which can be expensive). Instead, they’re using deep learning techniques like convolutional neural networks (CNNs) and fully connected layers to extract features from the images and classify them as either “building” or “not building”.

Here’s how it works: first, you feed a bunch of satellite imagery into the model. The model then processes this data using various filters and convolutions to identify patterns that are characteristic of buildings (like straight edges and rectangular shapes). This process is called feature extraction, and it helps the model learn what features are most important for identifying buildings in remote sensing images.

Once the model has learned these features, it can then classify new satellite imagery as either “building” or “not building”. The accuracy of this classification depends on a number of factors, including the quality of the data and the complexity of the algorithms used to train the model. However, in general, deep learning techniques like CNNs are very effective at identifying buildings in remote sensing images with high levels of accuracy (around 95% or better).

One example of how this technique can be applied is in urban planning and development. By automatically extracting building footprints from satellite imagery, planners and developers can create detailed maps that show the location and size of buildings within a given area. This information can then be used to identify areas where new buildings are needed or where existing buildings may need to be demolished or renovated.

Another example is in disaster response and recovery efforts. By using satellite imagery to automatically identify damaged buildings, emergency responders can quickly assess the extent of damage and prioritize rescue operations accordingly. This information can also help with post-disaster reconstruction efforts by providing a detailed map of which buildings need to be rebuilt or repaired.

In terms of selecting an appropriate learning rate for the model, we conducted several comparative experiments using different values (0.001, 0.01, 0.02, and 0.05) to investigate their effect on iteration. We found that a learning rate greater than 0.001 resulted in slower convergence of the loss function, while a value of 0.08 led to much larger values for the loss function compared to other learning rates. As a result, we trained our RU-Net model using a learning rate of 0.001 on the building dataset.

Regarding sample size and batch size, we tested three sets of comparative models with different combinations (512×512,16), (256×256,32), and (128×128,64) to determine which provided the best performance. We found that the combination (512×512,16) outperformed the other combinations for all evaluation metrics but consumed almost twice as much time as (256×256,32), and seven times as much time as (128×128,64).

In terms of selecting an appropriate learning rate for the model, we conducted several comparative experiments using different values (0.001, 0.01, 0.02, and 0.05) to investigate their effect on iteration. We found that a learning rate greater than 0.001 resulted in slower convergence of the loss function, while a value of 0.08 led to much larger values for the loss function compared to other learning rates. As a result, we trained our RU-Net model using a learning rate of 0.001 on the building dataset.

I’m interested in understanding the technical details behind it.