PhytNet -- Tailored Convolutional Neural Networks for Custom Botanical Data

TL;DR

It is shown that PhytNet is a promising candidate for rapid disease or plant classification and for precise localisation of disease symptoms for autonomous systems and that the most informative light spectra for detecting cocoa disease are outside the visible spectrum.

Abstract

Automated disease, weed and crop classification with computer vision will be invaluable in the future of agriculture. However, existing model architectures like ResNet, EfficientNet and ConvNeXt often underperform on smaller, specialised datasets typical of such projects. We address this gap with informed data collection and the development of a new CNN architecture, PhytNet. Utilising a novel dataset of infrared cocoa tree images, we demonstrate PhytNet's development and compare its performance with existing architectures. Data collection was informed by analysis of spectroscopy data, which provided useful insights into the spectral characteristics of cocoa trees. Such information could inform future data collection and model development. Cocoa was chosen as a focal species due to the diverse pathology of its diseases, which pose significant challenges for detection. ResNet18 showed some signs of overfitting, while EfficientNet variants showed distinct signs of overfitting. By contrast, PhytNet displayed excellent attention to relevant features, no overfitting, and an exceptionally low computation cost (1.19 GFLOPS). As such PhytNet is a promising candidate for rapid disease or plant classification, or precise localisation of disease symptoms for autonomous systems.

Figures (4)

• Figure 1: Distributions of non-photochemical quenching (NPQt)(a) and photosynthetic yield (Phi2)(b) of cocoa trees with different disease states. Box plots show the interquartile range with whiskers at 1.5 times the IQR from the first and third quartiles. Raw data points plotted as white circles. Measurements taken from cocoa trees in five disease states with a MultispeQ v2.0. BPR n=10, FPR n=5, Healthy n=9, Witches' broom (adjacent tissue) n=5, Witches' broom (affected tissue) n=5.
• Figure 2: Mean reflectance spectrum measurements, ±1 standard error (a) and feature importance scores (b) derived from a random forest classifier trained on the same data. Spectroscopy data were gathered from cocoa pods that were either healthy or showing early to mid-stage frosty pod rot or black pod rot symptoms. The standard error is shown by grey shading and the dotted lines at 400 and 720 nm show the bounds of the human visible spectrum. Random forest classifier train accuracy: 95.45% and test accuracy: 78.26%.
• Figure 3: Infrared images with class activation heatmaps produced using Grad-CAM and four CNNs. Models used are PhytNet, ResNet18, EfficeintNet-b0 and EfficientNetV2 (left to right). The leftmost column shows raw input images with ground truth labels in white, other white labels are predicted by each model.
• Figure 4: Violin and box plots of 10-fold cross-validation results for PhytNet and ResNet18 trained(a) and validated(b) on infrared or RGB images of cocoa disease. Shown here is the Gaussian density function, medium and interquartile ranges for mean F1, per class F1, precision and recall. PhytNet was trained only on IR data, while ResNet18 was trained on IR or RGB data. The datasets had four classes: Black pod rot, Frosty pod rot, Healthy and Witches' broom disease. n=70 images per class of early to mid-stage diseased or healthy cocoa with a 90%:10% train, validation split.