A Novel System for Potential Mosquito Breeding Hotspot Intimation and Monitoring Using MLOps and Improved YoloV3

Figure 2 shows the mosquito breeding site notification and monitoring system. The longitude and latitude of the detected part will be sent to the remote storage. Remote storage maintains the history of the hotspots. Further, this data can be verified to assess the impact of the spraying performed by the local authorities. Public cameras can be of two types a)Analog and b) Digital Hd cameras. The minimum resolution would be 1080 pixels. The flow of the activities for the implemented system would be as below.

Workflow

Step 1: Dataset preprocessing was done for the stagnant water dataset.

Step 2: Training and validating the deep learning model were done.

Step 3: The trained model was deployed on the cloud.

Step 4: Public Camera Video streaming was sent to the cloud.

Step 5: A Deep learning algorithm was executed for detecting stagnant water.

Step 6: If water was found, then location coordinates will be sent to the remote server else the next video was considered for frame detection.

Step 7: The history of hotspots was stored on the remote server.

2.1 Cloud infrastructure

We trained our model on Google Colab and then it was deployed on the cloud whilst detection was done for the received video recording from the public cameras.

To provide continuous monitoring the MLOps service of Microsoft Azure Cloud was used [11, 12]. Further, the corresponding location coordinates were stored at the remote server for vector control action to be done by the local authorities.

2.2 Deep learning algorithm

Deep learning is a subset of machine learning methods based on artificial neural networks. Here we used a deep learning algorithm known as YoloV3, to provide accurate real-time detection.

As shown in Figure 2, the algorithm includes two steps: 1) Data Preprocessing and 2) Training and Validation.

2.2.1 Data Preprocessing

To validate the model, the stagnant water image and video datasets were published on the Mendeley repository [13, 14]. This dataset contains 1976 labeled images of stagnant water. These images are street view images taken from top and side view for different water surfaces such as green, muddy, black, and shiny. The images were resized to 256 x 256 pixels and then labeling was done. These images were collected from two cities in India at different times of the day. In the detection step, part of the image will be classified into water or wet surface class. Hence images were labeled in Yolo format with a ‘0’ value for the water class. While ‘1’ is taken for the wet surface class. A few images were rotated by 90 degrees to increase the data samples.

2.2.2 Training and Validation

We trained Yolo family algorithms on Google Colab using transfer learning. For training and validation, the distribution of the dataset was done as per Table 2. For the dataset of 3812 images, the sources used, and the number of images captured are shown in the table. After experimenting with anchor box size [15], we chose Yolov3 out of other existing algorithms.

To increase object detection performance, data fusion was done by adding google images with real-time images [16]. Meng et al. [17] thought that “information fusion is the study of efficient methods for automatically or semi-automatically transforming information from different sources and points in time into a representation that provides effective support for human or automated decision making.” The corresponding experimentation results are given in Table 3.

Table 2. Dataset distribution

Table 3. Model accuracy

The following training parameters were employed: number of batches=64; number of subdivisions=32; maximum number of iterations=45,000; and the learning rate=0.001. After 16 hours, the training was interrupted with a validation loss value of 0.04, at iteration 41,200.

Yolov3 Detection Process

Yolov3 mainly utilizes residual layers and a certain number of convolutional layers to complete the detection process and uses the entire image features to predict each bounding box. This algorithm performs real-time detection at a fast speed. At the same time, it predicts all classes of all bounding boxes for the complete training, which maintains maximum average accuracy and strong real-time performance [16].

It divides the input image into N×N grids and assigns one anchor bounding box for each ground truth object. Equations 1-4 show four coordinates (tx, ty, tw, th) predicted by the network for each bounding box, and then a function was used to predict three corresponding parameters in the form of coordinates like the center point coordinates (cx, cy) of the bounding box, the width- bw, and the height- bh [17]. Where σ (•) is the sigmoid activation function, which was used to limit the center like the center point coordinates (bx, by) of the bounding box, the width bw, and the height bh. The confidence in the detected object was calculated by the formula given below,

bx=σ (tx) + cx                 (1)

by=σ (ty) + cy              (2)

bw=p_w e^tw                    (3)

bh=p_h e^th                   (4)

Pruning

As Yolov3 has 106 layers, model compression was needed to reduce the model parameters and make it easy to deploy. The methods of model compression are categorized into quantization [18-20], pruning [21, 22], low-rank decomposition [23, 24], and knowledge distillation [25, 26]. Out of mentioned techniques, we used layer pruning, wherein layers were removed according to the loss and Bfvalue. Yolov3 has good detection accuracy, but to reduce time complexity and memory constraints, we performed pruning which contributes to the carbon footprint reduction as well [27].

Evaluation Metrics

As the problems of missed detection and false detection happened in the monitoring process of water detection, hence for experimentation mAP, Precision, and Recall were used as evaluation parameters. The precision was the ratio of the number of correctly detected water/wet areas to the total number of detected water/wet areas. The Recall was the ratio of the number of correctly detected water/wet areas to the total number of water/wet in the data set. The calculation method is shown in Eqns. (7) and (8).

Precision $=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$           (5)

Recall $=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$               (6)

$m A P=\frac{1}{\mathrm{nc}} \sum_{\mathrm{i}=0}^{\mathrm{nc}} \mathrm{APi}$                 (7)

$\mathrm{APi}=\frac{1}{n d} \sum_{j=0}^{n d} P i j$               (8)

In these formulas, True Positive (TP) indicates the number of correctly detected water/wet surfaces, True Negative (TN) indicates the number of correctly detected backgrounds, False Positive (FP) indicates the number of incorrect detections, and False Negative (FN) indicates the number of missed detections, respectively. As shown in equations 9 and 10, mAP is the average of AP (average precision), and AP is the average of all categories, i.e., classes used for labeling. The mean average precision was calculated by taking an average of precision for several recall values.

Sustainability

Sustainability contributes towards fulfilling the needs of present generations without compromising the needs of coming generations, and at the same time maintaining a balance between the environment and economy for social well-being. This system indirectly contributes to the United Nations Sustainability Goal i.e UNSDG (3) “Good Health and well-being” which aims to avoid premature deaths by preventing disease spread.