Automatic measurement of coverage area of water-based pesticides-surfactant formulation on plant leaves using deep learning tools

TL;DR

The paper addresses quantifying foliar wetting by water-based pesticide formulations on cucumber leaves, arguing that wet-area measurements offer a better proxy for spray efficacy than contact-angle metrics. It develops an automatic measurement pipeline by applying a deep learning model (HED-UNet) to overhead video frames to segment wet areas and convert pixel counts to physical areas, enabling analysis of wetting as a function of surfactant concentration. Direct CMC measurement yields $80 \\pm 5 \\mu\text{L}/\\text{L}$, and experiments cover concentrations up to $11.25$ times the CMC, showing that maximum wet area grows with surfactant concentration. The DL approach accelerates processing (seconds per image) and handles image artifacts better than threshold-based methods, though some underestimation occurs due to residual false negatives; overall, the method provides a transferable framework for precision agriculture wetting studies, with limitations tied to training-set size.

Abstract

A method to efficiently and quantitatively study the delivery of a pesticide-surfactant formulation in water solution over plants leaves is presented. Instead of measuring the contact angle, the surface of the leaves wet area is used as key parameter. To this goal, a deep learning model has been trained and tested, to automatically measure the surface of area wet with water solution over cucumber leaves, processing the frames of video footage. We have individuated an existing deep learning model, reported in literature for other applications, and we have applied it to this different task. We present the measurement technique, some details of the deep learning model, its training procedure and its image segmentation performance. Finally, we report the results of the wet areas surface measurement as a function of the concentration of a surfactant in the pesticide solution.

Figures (10)

• Figure 1: Plot of the data for CMC measurement. On the vertical axis are reported the values for the surface tension (SFT), which has been measured with the Wilhelmy plate method, and on the horizontal axis are reported the values for the volume concentration. According to the theory of CMC, the surface tension is supposed to follow an exponential decay as the concentration of surfactant grows, until it reaches the critical micelle concentration (CMC), ant then it is supposed to continue linearly. So, the data have been plotted in a semi-logarithmic plane, in order for the exponential part to appear linear. In the figure the datapoints are represented in two different colors to higlight those belonging to the two different regimes (exponential in red, and linear in green). Then, two linear best fits are performed, for the initial and the final regime (see blue lines), and the CMC is estimated as the concentration at the interception of the two lines (see black circle in the plot). The estimate for the CMC is $80 \pm 5 \mu l / l$. It is expressed in $\mu l$ of surfactant, as it is delivered in the vendor's container, per liters of water.
• Figure 2: Example of a single frame from a video recording, on the left, and the correspondent output of the Deep Learning model segmentation, on the right. In this example, in the center, due to the geometry of leaf, a series of several bright reflection spots are present, aligned in a vertical row, along big part of the image. On the right we can see that the Deep Learning model has correctly identified the reflection bright spots as belonging to the wet area.
• Figure 3: Another example of a single frame from one of the video footage, and the correspondent mask recognized by the Deep Learning model. In this example, we can see how the Deep Learning model has been able to recognize as "dry" a spot of dry leave surface, all encircled by wet surface, at its bottom (pointed by the red arrow). In this image we can also see how the DL model can correctly neglect the black background present in the image.
• Figure 4: Schematic of the structure of the HED-UNet model (based on the schemes present in Heidler2022. In the lower left corner a depiction of a possible input image is shown. Only this image, and the final prediction mask, are shown in this schematic in their extension: all the other images, feature masks, and intermediate predictions are represented just as lines, as if the images were seen from an edge. The input image is fed into an encoder-decoder structure, which consists of in subsequent hidden layers that first reduce the original size of the image, and then re expand it, back to the original size (n.b. the encoder-decoder used here consist of 6 hidden layers of lower resolution, the number of layers depicted is smaller for graphic clarity). The encoder-decoder structure makes it possible to exploit both the localized information gathered in the high resolution layers, and the global, nonlocal information obtained from the low resolution layers. The skip connections are parts of the network that connect the encoder layers and the decoder layers. One of the most relevant features of the HED-UNet model is the "deep supervision"; it consists of part of the network that extract information from the lower resolution layers, independently, and generate separate predictions. This mechanism is used for both the edge detection and the segmentation parts of the model. Another relevant feature of the model is the "attention mechanism". This consists in special 'attention maps', i.e. weight coefficients, that are multiplied, pixel per pixel, against the intermediate predictions. In this way different areas contribute in different ways depending on different degrees of 'importance'. After combining all the intermediate prediction outputs, the final prediction is obtained.
• Figure 5: Plots of the training loss and the validation loss as functions of the number of epochs. We can observe a stable loss of around 0.001, computed with the torch.nn.BCEWithLogitsLoss() loss function explicitly reported in equation \ref{['eq:balancedBCEloss']}.