A Novel Wireless Leaf Area Index Sensor Based on a Combined U-Net Deep Learning Model

Abstract

Leaf area index (LAI) is an important parameter for forestry vegetation canopy structure investigation and ecological environment model study. Traditional ground direct measuring method is too time and labor consuming, while the remote sensing technique lacks of adequate validation and comparative analysis. Here, a novel wireless LAI sensor based on a lightweight deep learning model (LAINET) has been designed with a Raspberry Pi microcomputer and a LoRa transceiver. The mainly metering pattern of sensor system is the digital hemispherical photography (DHP) methodology based on Beer-Lambert law: firstly, the crown canopy's image is captured and segmented by LAINET, then the vegetation gap fraction can be extracted to calculate the LAI value. Our proposed LAINET consists of a lightweight convolutional neural network (CNN) and a generative adversarial network (GAN). The average accuracy of semantic segmentation (i.e. CNN part) could reach 0.978, and the combination of GAN for image super-resolution reconstruction can improve the accuracy of gap fraction measurement more by 5.5%. In addition, LAINET effectively solves the problem of low segmentation accuracy brought by environmental effects, the separation accuracy in direct sunlight or clear weather has been improved significantly. So the ultimate LAI value can be calculated precisely and stably. Experiment results show that the proposed sensor obtains a fine measuring error of less than 4% when comparing with the commercial plant canopy analyzer HM-G20. Combined with Uninterruptible Power Supply module of 5200 mAh, the sensor can work effectively for about 8 months, principally meeting the deployment and measurement criteria of forestry LAI. Therefore, the wireless sensor presented in this paper has a great application prospect.

GS3: A Lightweight Method of Generating Data Blocks With Shuffling, Scrambling, and Substituting Data for Constrained IoT Devices

“…A comparison between Raspberry Pi platform and (TelosB, MicaZ and Iris) platforms is presented in [49], showing that Raspberry Pi devices have some strengths: processing power, memory, connectivity, and flexibility. Two recent use cases to design a WSN using Raspberry Pi as a sensor platform are: In [50], it is created a low-cost WSN for monitoring and measuring the Leaf Area Index (LAI), which is an important parameter for forestry vegetation canopy structure investigation and ecological environment model study. In [51], it is analyzed the performance of the deep learning algorithms in compressing medical images and the efficiency of the Raspberry Pi WSN in transmitting the compressed images across the WSN nodes.…”

“…In this section, we describe the experiments (or benchmark tests) to assess and validate our proposed method. Firstly, we suppose that our IoT devices transmit multimedia data, such as the use cases in [50] or [51]. Therefore, some of the tests that we carried out are histogram analysis, Chisquare, and Entropy tests.…”