In our previous analysis of Seongnam City using our cutting-edge change detection application, we delved into the automatic extraction of temporal changes.
In this insightful article, we will check the accuracy of the model's inference results.
With our esteemed partner, K-water's keen interest in Seongnam City, we have conducted thorough analyses leveraging diverse datasets and our advanced change detection model.
In the initial aerial image, the ground appears obscured, but in the newly taken photo, the landscape has been transformed into a forest. It is possible that this area is of a sensitive nature, prompting speculation that the Korean government intentionally masked it by creating a forest cover.
Perhaps, this area is considered a confidential facility, and the Korean government should have hidden the area by drawing a forest on the ground from the beginning.
This is believed to be a problem caused by a change in aerial photography processing policy, and in the future, when aerial photography is first taken, important facilities must be covered with forests to avoid being detected by machine learning models.
We also confirmed that our change detection model was good at detecting newly constructed buildings on the ground.
Deep Block's Change Detection model incorporates a convolution layer for feature extraction. However, as highlighted in our deep learning course, convolution of an image may lead to compression and loss of some information within the image. As you can see, the road is thin and long, but these lanes are lost because of convolution.
Check how the convolution layer works at our free deep learning course.
While this is a current limitation of the model available online, we are enhancing our capabilities and exploring new architectures for our change detection model. If you are seeking a more advanced model, please reach out to us via our website, deepblock.net/contact.
Despite these limitations, our application has successfully identified sections of newly constructed roads. Deep Block's model boasts a higher level of accuracy than perceived. Even if the extracted area does not cover the entire region, we encourage remote sensing image analysts utilizing Deep Block to closely examine the surroundings of the identified changes.
If the internet connection is unstable or Deep Block traffic increases, it may take time for loading the image.
While not immediately apparent in the image, the area where specific structures once stood has now been transformed into a forest.
Our model successfully identified the construction of a building that previously did not exist. The standout feature of Deep Block is its exceptional ability to detect subtle changes, especially in high-resolution and spatially detailed images. We fearlessly tackle large-scale photos, showcasing our strength in analyzing vast amounts of data with precision.
The disappearance of the greenhouses signifies the emergence of new roads in the landscape.
While the roads may not be easily identifiable, the presence of small polygons marking the lanes is a key feature. Despite being a limitation of our model, these small polygons should not be disregarded, as they serve as indicators of newly constructed roads.
The change in the color of the greenhouse may seem insignificant at first glance. However, our focus lies on scrutinizing alterations on the roof. Factors such as heavy objects on the roof or unauthorized installation of solar panels can pose risks of structural instability and durability issues.
But some people just want to judge it as negative. In this case, the pre-trained model we provide must be trained in a different way. Please contact our team if you want to re-train our change detection model.
A new access road has been created on the right side of the highway, and it can also be seen that buildings have disappeared.
A newly constructed baseball stadium has been unveiled, but only a part of the ground is visible. This discrepancy arises because the model tends to ignore seasonal color variations in the field.
False positive cases can also be found. Although people were able to detect that the roof had changed, the model determined that the shape of the roof looked slightly different. However, since there appears to be no significant change to the naked eye, this is an incorrect detection result.
It is essential to acknowledge that machine learning models, much like humans, are susceptible to errors. However, our change detection application remains really useful due to its minimal false negative cases and high-speed inference capabilities.
To manually observe ground changes, one would need to sift through extensive 20GB satellite images repeatedly each minute. This labor-intensive task is currently undertaken by governments and militaries worldwide, but the good news is that our automated system can streamline this process efficiently.
The accuracy of orthoimages can be compromised by the shooting angle towards the surface, especially in the presence of tall buildings. This can lead to inaccuracies in orthoimage generation, making it challenging to detect ground changes effectively. Although the image is an orthophoto, it has not undergone precise preprocessing, resulting in incorrect detection outcomes by the model.
We know how to ignore slight errors in orthophotos. Please contact us if you are interested in our technology.
You can also check that new buildings have been constructed through our application’s inference.
You can also see that the color of the roof changes and the building changes.
We keep checking out new buildings.
Although the construction of the building was initially incomplete, it has now been finished, resulting in a noticeable change in the roof color and the addition of multiple structures on top.
The roof has changed, and if you look closely, a building right next to it is gone.
The issue seems to have arisen due to the improper creation of the ortho image.
Nevertheless, it is reassuring to note that Deep Block has the capability to detect the construction of tunnels.
Furthermore, our analysis has verified the construction of a new roadway in the area.
The Seongnam area seems to have changed quite a bit.
Buildings are appearing and disappearing.
The location where the greenhouse once stood has now been transformed and no longer features the structure.
This structure covers the highway, serving as a noise-reducing barrier. The Korean government installs walls along highways to mitigate sound pollution.
A newly constructed parking lot can be observed in the area, with a closer look revealing the presence of a newly built building indicated by a small polygon on the right side.
This location was also blurred by the Korean government at first and later covered with a forest.
If you look closely, you can see that a new building has been constructed, and solar panels have been installed on the right building.
Even when a model's IoU is not good, Deep Block's machine learning model still produces fairly accurate comparison results.
A newly developed area can be observed on the left side, showcasing recent changes in the landscape.
On the contrary, a false positive case was identified on the right side, attributed to inadequate orthophoto processing of a high-rise apartment complex.
Although the model could not detect the entire building, it was able to identify at least small areas of change.
Conclusion
The combination of Deep Block's change detection model and its powerful algorithm showcases impressive accuracy in performance. Despite this success, we have identified areas for improvement, particularly in road detection and the challenges faced in detecting changes in areas with tall buildings where orthoimages may not be accurately generated.
At present, our support is limited to optical images. If you have an interest in observing changes in SAR satellite images, please reach out to us for further information.
Fortunately, our inference result analysis reaffirms that Deep Block offers a straightforward and precise tool for remote sensing image analysts. We invite you to stay tuned for updates on our team's ongoing efforts to automate the analysis of large-scale remote sensing images.
Thank you!
