CNN Model Comparison: A Study on Pequi Tree Branch Blight Detection

Convolutional neural networks (CNNs) are moving from the lab to the field in agricultural disease identification. A study on Pequi tree (Caryocar brasiliense) branch blight (seca-dos-ponteiros) systematically compared the performance of five CNN architectures—AlexNet, Inception, LeNet, VGG16, and MobileNet—in disease detection. This research isn't just about the health of a single tree; it touches on a core question at the intersection of computer vision and ecological conservation: when algorithms confront the complex pathological features of the real world, which architecture can truly deliver?

Why Focus on Pequi Tree Branch Blight

The Pequi tree is an iconic species of Brazil's Cerrado biome, and its fruit holds both economic value and ecological significance. Branch blight poses a serious threat to tree development and fruit yield, causing economic and ecological losses that cannot be ignored. Traditional diagnosis relies on manual observation of texture and color changes in leaves, branches, and flower panicles—a process that is inefficient and constrained by subjective experience.

The entry point of this study is: can CNNs automatically identify these pathological patterns and precisely distinguish healthy from diseased plants? If feasible, the window for early intervention would open much earlier.

Five CNN Architectures Go Head-to-Head

The five architectures selected for the study span different stages of CNN development and design philosophies:

LeNet is one of the earliest CNNs, lightweight in structure and suited to small-scale image classification tasks. AlexNet rose to fame in the 2012 ImageNet competition, proving the potential of deep CNNs. VGG16 is known for its orderly stacks of 3×3 convolutions, reaching a depth of 16 layers—strong feature extraction capability but high computational cost. Inception introduces multi-scale parallel convolutions, capturing features at different granularities within the same layer. MobileNet, meanwhile, is purpose-built for mobile and edge devices, dramatically compressing parameter counts through depthwise separable convolutions.

Comparing MobileNet and VGG16 on the same task inherently embeds an engineering trade-off question: in agricultural scenarios, which matters more—accuracy or deployment cost?

Depthwise Separable Convolution is the core design of MobileNet. It decomposes standard convolution into two steps: first, spatial convolution is performed independently on each input channel (depthwise convolution), then a 1×1 convolution combines features across channels (pointwise convolution). This decomposition can reduce computation and parameter counts by roughly 8 to 9 times, at the cost of delaying cross-channel information interaction and slightly diminishing representational capacity. For tasks like agricultural disease detection—which typically lack large-scale annotated data—a smaller parameter count actually helps reduce overfitting risk, which is one reason MobileNet is often prioritized in small-sample scenarios.

Multi-scale parallel convolution is the signature design of the Inception module. Within the same layer, the network simultaneously uses convolution kernels of different sizes (such as 1×1, 3×3, 5×5) and pooling operations, then concatenates the outputs along the channel dimension. This structure enables the model to capture both fine-grained textures (such as spots on leaves) and coarse-grained shapes (such as the overall outline of blighted branches) within a single layer—particularly well-suited to agricultural images where pathological feature scales vary widely. The cost is a complex internal module structure with many computational graph branches, requiring strong operator fusion capabilities from inference frameworks during actual deployment.

Technical Approach and Data Dimensions

The study employed a Python-ecosystem computer vision toolchain: Scikit-learn for model evaluation and data splitting, NumPy for numerical computation, TensorFlow for building and training networks, and Matplotlib for visualization analysis.

Model inputs consisted of three categories of image data: leaves, branches, and flower panicles. These organs exhibit texture changes and color abnormalities after infection with branch blight, and CNNs need to learn these pathological patterns from pixel-level features. The design of multiple input dimensions increases the model's generalization challenge—pathological manifestations differ significantly across organs, and whether a single model can handle them uniformly is an important consideration in architecture selection.

Data splitting and model evaluation are the credibility cornerstones of this type of comparative study. When sample sizes are limited, k-fold cross-validation is typically needed rather than a single train/test split, to reduce variance in evaluation results. The stratified sampling provided by Scikit-learn ensures consistent proportions of healthy and diseased samples across folds, avoiding inflated accuracy caused by class imbalance. Furthermore, evaluation metrics cannot rely on accuracy alone—when diseased samples are a minority, recall and F1 score better reflect the model's actual detection capability, which is especially critical for early disease intervention decisions.

The Methodological Significance of Multi-Architecture Comparison

The value of such comparative studies lies not only in selecting the best model. Differences among CNN architectures in feature extraction methods, receptive field sizes, and parameter efficiency can amplify into performance gaps in specific tasks. For scenarios like Pequi tree branch blight—with limited sample sizes and subtle pathological features—networks that are too deep may overfit, while those too shallow may underfit.

The ultimate goal of the study is to identify the architecture that most accurately distinguishes healthy from diseased plants. This conclusion has direct guiding significance for subsequent field deployment: if MobileNet can maintain acceptable accuracy while substantially reducing computational requirements, then real-time detection on edge devices becomes possible; if VGG16 or Inception significantly leads in accuracy, then it becomes necessary to weigh whether higher accuracy justifies the computational cost.

Overfitting and underfitting are opposing constraints in architecture selection. Overfitting refers to a model performing excellently on the training set but experiencing a sharp drop in accuracy when generalizing to new samples—common in deep networks with far more parameters than samples. Underfitting refers to model capacity being insufficient to capture pathological patterns in the data, manifesting as low accuracy on both training and test sets. In tasks like Pequi tree branch blight, where sample sizes are limited and pathological features are subtle, network depth and parameter count need to match the data scale. Transfer learning—initializing networks with pretrained weights from large-scale datasets like ImageNet—is a common means of mitigating overfitting, and it also keeps deep architectures like VGG16 and Inception competitive in small-sample agricultural tasks.

From a Single Study to the Logic of Agricultural AI Deployment

Applying CNNs to disease detection for a specific tree species is a microcosm of agricultural AI moving from general scenarios toward vertical specialization. The methodology of the Pequi tree branch blight study can be transferred to disease management for other cash crops and ecological tree species. The key lies in: the quality of annotated data, the visual separability of pathological features, and the model's robustness to variations in lighting and shooting angles.

This study doesn't stop at the level of "can CNNs detect disease" but further asks "which CNN is most suitable." This kind of fine-grained architecture-level comparison provides actionable selection criteria for agricultural practitioners and technology developers. When algorithms truly enter orchards and forests, choosing the right tool matters more than simply using a tool.

One-line Summary

A study systematically compares five CNN architectures for Pequi tree branch blight detection, providing a basis for agricultural AI model selection.

Paragraph Summary

A study on Pequi tree (Caryocar brasiliense) branch blight systematically compared the performance of five convolutional neural network architectures—AlexNet, Inception, LeNet, VGG16, and MobileNet—in disease detection. Using three categories of images (leaves, branches, and flower panicles) as input, and building a technical pipeline based on the Python ecosystem with Scikit-learn, NumPy, TensorFlow, and Matplotlib, the study aimed to precisely distinguish healthy from diseased plants. The article notes that the methodological value of such architecture-level comparisons lies in revealing the trade-off between accuracy and deployment cost: if MobileNet can reduce computational demands at acceptable accuracy, real-time edge device detection becomes feasible; if VGG16 or Inception significantly leads in accuracy, then it becomes necessary to weigh whether higher computational cost is justified. The study provides actionable reference criteria for model selection in agricultural scenarios.