at https://rsvqa.sylvainlobry.com/. Examples of RSVQA model predictions are reported in Figure 4 for both resolution images. Note that, for a single image, several questions are possible and the same model is used to answer them. PERSPECTIVES This first work opens a wide range of possibilities for a new line of research toward the next level of human/ image interactions. Nonetheless, all the blocks of the 1) Image Representation 3) Feature Fusion 4) Question Answering Yes No CNN 2) Textual Representation (a) Peri Urban Cropland City Center Increasing " How many cars are there? " RNN 6 7 8 (b) (c) (d) FIGURE 3. An example of an RSVQA system [12]. (a) A remote sensing image and (b) a question in natural language enter two source-specific neural networks, each one outputting a vector representing their information, and (c) both vectors are combined and become (d) the input of a classifier that outputs all the possible answers as separate classes. RNN: recurrent neural network. Q: " Is there a road? " A: " Yes (GT: yes) " Q: " Is there a farm? " A: " Yes (GT: yes) " Q: " How many small residential buildings are there? " A: " Between 10 and 100 (GT: 147) " Q: " Is there a residential building? " A: " Yes (GT: yes) " A: " Yes (GT: yes) " Q: " Are there fewer commercial buildings than water areas? " Q: " How many small residential buildings are there? " A: " Between 10 and 100 (GT: 18) " Sentinel-2 Images (a) Aerial VHR (b) FIGURE 4. The RSVQA system predictions for (a) Sentinel-2 and (b) aerial images, respectively, and different questions for each. The same model is used to answer all of the questions related to one-resolution imagery. GT: ground truth; VHR: very high resolution. (Sources: (a) Copernicus and (b) USGS.) 94 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2021 Q: " Is there a baseball field? " A: " Yes (GT: yes) " Q: " How many commercial buildings are close to a baseball field? " A: " 0 (GT: 0) " Q: " Is there a building south of a road? " A: " No (GT: Yes) " Q: " Is there a commercial building? " A: " No (GT: no) " Q: " How many small buildings are there? " Q: " Are there fewer commercial buildings than roads? " A: " Yes (GT: yes) " A: " 0 (GT: 0) "https://rsvqa.sylvainlobry.com/