nitin-45/modem

Identify the monuments from Satellite Images using Deep Learning and Integration of Interpretability for the predicted outcomes (Explainable AI).

MOnument DetectiOn from satEllite iMage (MODEM)

PROBLEM STATEMENT

---

To identify the monuments from Satellite Images using Deep Learning and Integration of Interpretability for the predicted outcomes (Explainable AI).

TEAM MEMBER DETAILS

---

Aniket Thakur (Team Lead)
Aditi
Ashita Priya
Nitin Kumar
Piyush Kumar Singh
Soham Dutta

SOLUTION

---

1. For a long time, finding undiscovered monuments have been a monumental task for archaeologists. We intend to simplify that by making a computer learn to detect monuments.

2. Imagine how easy it would have been to detect the Khajuraho Temple if satellite images were available back in early 1800s. A British army captain, T.S. Burt, discovered this temple accidentally hidden behind the palm leaves (and hence the name Khajur-aho!).

3. Telling if a satellite image is a monument might sound like a simple ML problem which can be attacked using CNNs, but we need sufficient proof for our prediction. However, with CAM (Class Activation Map) techniques at our disposal, we can predict the pixels which led the CNN model to make decisions. Thus, we can explain the decision made by the model.Besides the model, the data collection itself is a challenging task, which we try to address in our solution by using Transfer learning to learn from minimal amount of data.

4. The inductive bias here is that most of the man-made structures and monuments have symmetry and patterns – a quality which is best grasped and learned through CNNs.

5. We can use standard training and testing dataset split to figure out the accuracy of our model. Human verification of the activated pixels can be used to judge the veracity of the explanation generated by the model.

6. Frameworks/Technology used: Python 3, NumPy, Pandas, TensorFlow, PIL, selenium, gekodriver, Firefox

ARCHITECTURE

---

For image recognition we are using pre trained CNN model (VGG-19 / GoogLeNet / ResNet)

Classification part with fully-connected and softmax layers

METHODOLOGY

---

Satellite Image Collection

1. For collecting the images of Indian monuments, we have used the satellite images from ISRO .

2. We can identify a monument with the help of its latitude and longitude. So we only need to create a CSV file which contains name, latitude, and longitude of the monuments, e.g., Taj Mahal , India Gate

3. Now simply automate the process of image collection, and this can be done in python.

4. The images attached here are basically the cropped screenshots from the above links. Video demonstration is also there.

Taj Mahal
India Gate

Societal Impact & Business Relevance

---

1. Discovering unknown monuments will benefit the tourism industry, bringing revenue to the state. Discovering monuments is a very difficult task, specially in locations which are geographically isolated.
2. There are numerous examples, in which monuments were accidentally discovered in the middle of jungles like Khajuraho, Angkorwat etc.
3. Now that we have satellite images to our rescue, we can expedite such searches and reveal new tourist spots in a given region. This will also help the society in preserving the cultural heritage of the area, helping people to discover their glorious unknown history!

Future Scope

---

Since, monuments are man-made structures, we can modify our solution to detect unknown build-ups on terrains like military structures etc.

---

THE WORK IS STILL IN PROGRESS!!