E-CONDOR: Efficient Contour-Based Detection Of Random Spatial Signals From UAV Observations Using Dual Stochastic Gradient

Maryam Zahra,Homa Tajiani,Hadi Alasti

2024-11-26

Abstract:This paper presents a novel efficient method for spatial monitoring of the distribution of correlated field signals, such as temperature, humidity, etc. using unmanned aerial vehicles (UAVs). The spatial signal is compressed to its iso-contour lines at a number of known levels that are introduced by data fusion center (DFC). The UAV traces a contour line of the field signal at a time, and reports the coordinates of its own traces to the DFC for spatial modeling. The DFC iteratively improves the spatial model of the field signal and assigns a new contour level to each UAV to trace and report its coordinates for spatial model improvement. The selected batch of levels and the start point of the search are introduced by the DFC. In order to reduce the required data for spatial modeling, and accordingly improve the algorithm data efficiency, dual stochastic gradient routines are used at the DFC to find a next proper number of contour levels in the batch, and to eliminate the redundant contour levels, in each iteration. The performance evaluation of the proposed algorithm based on computer simulations demonstrates significantly faster convergence, better signal estimation, and a higher data efficiency against when the stochastic gradient is not used.

Signal Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **How to efficiently use unmanned aerial vehicle (UAV) observation data to conduct spatial monitoring on relevant spatial signals (such as temperature, humidity, etc.)**. Specifically, the research objective is to compress the spatial signal to its contour lines and use the double - stochastic gradient algorithm to select appropriate batches of contour lines, thereby reducing the required amount of data and improving the data efficiency of the algorithm.

### Specific description of the problem:

1. **Unknown distribution of relevant spatial signals**: On a given known area A, there is an unknown relevant spatial signal \( g(x, y) \), with a spatial distribution similar to that shown in Figure 1.
2. **UAV observations**: One or more UAVs fly within the area to measure the intensity of the spatial signal.
3. **Contour line modeling**: Model the spatial signal as its contour lines at multiple known levels, assuming that the spatial signal remains unchanged during the multi - contour line observation process.
4. **Lack of initial information**: The UAV has no prior information about the number of contour lines M and their levels {ℓk} at the beginning.

### Core ideas of the solution:

1. **Role of the data fusion center (DFC)**: The DFC estimates the spatial signal based on the coordinates and signal intensity reported by the UAV, calculates a batch of contour line levels, and then broadcasts these levels to the UAVs one by one.
2. **Double - stochastic gradient algorithm**:
- **First process**: Calculate the increased number of contour lines and find new batches of contour lines.
- **Second process**: Delete redundant contour lines, that is, those contour lines that are too close to the previously drawn contour lines.
3. **Iterative optimization**: Gradually increase the number of contour lines through an iterative process until the learning error drops significantly and the range of spatial signal intensity converges.

### Formula summary:

- **Calculation of contour line levels**:
\[
\ell_i=\frac{\int_{y_i}^{y_{i + 1}}xf_g(x)dx}{\int_{y_i}^{y_{i + 1}}f_g(x)dx}, \quad i = 1,2,\cdots,M
\]
\[
y_i=\frac{\ell_i+\ell_{i - 1}}{2}, \quad i = 1,2,\cdots,M - 1
\]

- **Calculation of learning error**:
\[
\text{Error}_n=\frac{1}{P\times Q}\sum_{i = 1}^{P}\sum_{j = 1}^{Q}|\tilde{g}_n(x_i,y_j)-\tilde{g}_{n - 1}(x_i,y_j)|
\]

- **Update of increment factor**:
\[
\kappa_n=\kappa_{n - 1}+\left\lceil1+\frac{2|\text{Error}_{n - 2}-\text{Error}_{n - 1}|}{|\text{Error}_{n - 2}+\text{Error}_{n - 1}|}\right\rceil
\]

- **Deletion threshold of redundant contour lines**:
\[
\delta_n=\delta_{n - 1}\left|1-\frac{2(\text{Error}_{n - 2}-\text{Error}_{n - 1})}{\text{Error}_{n - 2}+\text{Error}_{n - 1}}\right|

Through the above methods, this paper proposes an efficient spatial signal monitoring algorithm, which can significantly reduce the required amount of observation data and improve the convergence speed and accuracy of the model.