TL;DR
This paper introduces a novel method for cleaning up noisy electrical signals from plants by using an optimized Infinite Impulse Response (IIR) filter. This filter effectively removes low-frequency "drifts" caused by environmental factors or measurement issues, allowing for a clearer analysis of the plant's actual responses to stimuli. The optimization process leverages wavelet packet energy to ensure that the baseline plant signals remain consistent across different conditions, paving the way for more reliable plant-based sensing applications.
What Problem Does This Paper Address?
Analyzing electrical signals from plants is crucial for understanding their responses to various environmental stimuli, which has potential applications in areas like robotics and environmental monitoring. However, these plant signals are notoriously "noisy." A significant challenge is the presence of low-frequency "drifts" – slow, baseline shifts in the signal that aren't related to the plant's actual response. These drifts can arise from a multitude of factors, including leaf movement, changes in humidity or temperature, variations in the recording equipment, and even the plant's internal biological state. These drifts can obscure the subtle electrical signals that indicate a plant's reaction to a stimulus, making it difficult to quantify and interpret the plant's behavior accurately. Unlike more established bio-signals like ECG or EEG, there isn't a standardized pre-processing framework for plant electrical signals, largely due to the lack of a clear understanding of these artifacts.
Methods
The core of this paper is the design of an "optimum" Infinite Impulse Response (IIR) filter. The goal is to create a filter that removes these troublesome low-frequency drifts while preserving the important high-frequency components that represent the plant's actual response.
Here's a breakdown of their approach:
- IIR Filter Design: They explore four common types of IIR filters: Butterworth, Chebyshev Type I, Chebyshev Type II, and Elliptic. These filters are designed to act as high-pass filters, specifically targeting the removal of low-frequency components.
- Wavelet Packet Transform (WPT): To optimize the filter parameters, they employ the Wavelet Packet Transform. WPT decomposes the signal into different frequency bands (nodes) at various levels. The energy within each of these nodes is then calculated.
- Optimization Criterion: The key idea is to use the energy distribution of the pre-stimulus (baseline) plant signals. The filter is optimized such that the WPT energy distributions of these baseline signals, recorded under different experimental conditions, are as similar as possible (i.e., they form "clusters" that are close together in a feature space). This ensures that the filter doesn't introduce biases that differentiate between plants or experimental setups before a stimulus is applied.
- Objective Function: They define a cost function (J) that quantifies the "distance" between the centroids of these energy distributions for different pre-stimulus signals. Minimizing this function leads to the optimal filter parameters.
- Nelder-Mead Simplex Algorithm: This standard optimization algorithm is used to find the filter parameters (like cutoff frequency, order, and ripple) that minimize the objective function.
Datasets
The study utilized experimental data collected from 11 tomato plants (Solanum lycopersicum, cv. Shiren) aged between 3-5 weeks. These plants were grown in a controlled greenhouse environment. The electrical signals were recorded at a rate of 10 Hz from three points on the plant stem (top, middle, and base as a reference).
The plants were subjected to four different stimuli:
- Ozone (O3)
- Sulfuric acid (H2SO4, 5 ml, 0.05 M)
- Sodium Chloride (NaCl, 5 ml, 3mM)
- Sodium Chloride (NaCl, 10 ml, 3mM)
The data was collected in a dark room within a Faraday cage to minimize external electromagnetic interference.
Key Insights and Ideas
- Drift is a Major Hurdle: The paper strongly emphasizes that low-frequency drifts are a significant impediment to analyzing plant electrical signals, often masking the true biological responses.
- Wavelet Energy as a "Signature": The core innovation is using the distribution of wavelet packet energy in the pre-stimulus signal as a consistent "signature" of the plant's baseline state. By forcing these signatures to overlap after filtering, they ensure the filter is unbiased.
- Chebyshev Type II Shines: Among the tested IIR filter types, the Chebyshev Type II filter consistently performed best in minimizing the objective function, indicating its superior ability to achieve overlapping pre-stimulus energy distributions.
- Robustness to Segmentation Size: The optimization approach proved robust across different data segmentation sizes (256, 512, and 1024 samples), suggesting that the chosen filter settings are generally applicable.
- Wavelet Basis Independence: The effectiveness of the optimized Chebyshev Type II filter was also shown to be largely independent of the specific wavelet basis used for WPT, highlighting the robustness of the overall methodology.
Key Results
- Optimized Filter Performance: The study successfully designed optimized IIR filters for plant electrical signal processing. The Chebyshev Type II filter, with specific parameters (e.g., a cutoff frequency around 0.77 Hz and order 6 for 256 samples), achieved the minimum cost function (Jmin) values, indicating the best overlap of pre-stimulus energy distributions. For instance, a Jmin of 11.64 was achieved for 256 samples.
- Visual Confirmation: Figure 5 visually confirms that the pre-stimulus wavelet energy distributions for different stimuli (H2SO4, Ozone, NaCl) are indeed highly overlapped after filtering with the optimized Chebyshev Type II filter.
- Stimulus-Induced Changes: In contrast, Figure 6 demonstrates that after applying stimuli, the spread and distribution of wavelet energies do change significantly, indicating that the filtered signal is now effectively capturing the plant's response to the stimuli.
- Robustness Demonstrated: Figures 7 and 9, along with Table I, show that the filter's performance remains consistent across different data segmentation sizes and that the choice of wavelet basis (amongst the common ones tested) does not drastically alter the outcome.
Limitations and Future Work
- Limited Stimuli Types: The study focused on a specific set of chemical and gaseous stimuli. Future work could explore a wider range of stimuli, including physical ones.
- Plant Species Specificity: While the study used a single species (tomato), it's acknowledged that plants from different species or grown in different environments might exhibit variations. Further research could investigate the generalizability of the method across diverse plant types.
- Classification of Stimuli: The current work focuses on drift removal and ensuring baseline consistency. The authors suggest that future research could build upon this by developing methods to classify the specific external stimuli based on the filtered plant electrical signals, moving beyond just detecting a response.
- Real-time Applications: While not explicitly stated as a limitation, the computational cost of WPT and optimization might be a consideration for real-time applications, which could be an area for future exploration.
Summary updated