Detect tree wilt with machine learning

1. Scientific objective

The objective is to reproduce a binary remote-sensing classifier that assigns a wilt score to a segmented QuickBird image object. The source study targeted foliage discolouration associated with Japanese oak wilt and Japanese pine wilt. In a research workflow, the model can help prioritize image segments for map review and field inspection; it does not identify a pathogen or confirm disease in an individual tree.

2. Data and provenance

The downloadable tree_wilt.csv contains 574 rows, five numeric inputs and one binary target, with no missing values.

Derived tutorial subset

The UCI release provides separate training and testing files. This tutorial table is a derived, rebalanced subset: it contains all 500 rows from the source testing file plus the 74 diseased-tree rows from the source training file. It excludes the 4,265 other-land-cover rows from that training file. The resulting table contains 261 positive and 313 negative rows and must not be compared directly with results from the original UCI split.

Outlier cleaning before modelling

Neural Designer applies Tukey’s rule to the numeric variables: values outside [Q1 − 3×IQR, Q3 + 3×IQR] are flagged. The per-variable counts are shown below; they overlap, so they do not sum to the number of excluded rows.

Seven overlapping rows (1.22% of the table) are marked unused. All seven have wilt = 0. The model therefore uses 567 rows: 261 positive and 306 negative.

3. Model

The initial network standardizes the five inputs, passes them through one dense layer with three tanh neurons and returns one sigmoid score. The 5–3–1 architecture contains 22 trainable parameters and is used as the starting point before hidden-layer selection.

4. Training strategy

The classifier minimizes class-weighted squared error with the quasi-Newton method. The stored configuration uses positive and negative weights of 1.0862 and 0.9265, respectively, and no explicit regularization.

For the initial three-neuron network, training error falls from 0.982 to 0.169 over 98 epochs. Selection error reaches its lowest region much earlier—approximately 0.158 near epoch 24—and then rises to 0.548. This divergence is evidence of overfitting in the displayed run; it is one reason not to judge the model from training loss alone.

5. Model selection and baseline

Growing-neurons selection evaluates hidden-layer sizes from one to ten, with three trials per size. The lowest reported selection error is 0.1657 at four hidden neurons; the corresponding training error is 0.1380.

Selected architecture

The selected 5–4–1 network has 29 trainable parameters and is the model used for the final testing analysis, browser demonstration and downloadable Python export.

6. Scientific validation

Final performance is calculated once on 113 rows assigned to the testing subset. At the fixed 0.50 threshold, the classifier detects 38 of 49 wilt-labelled segments and correctly rejects 47 of 64 other-land-cover segments.

7. Inference and reproducibility

A defensible inference pipeline starts upstream of the neural network. Image segmentation and the five feature definitions must remain identical to the study pipeline, and every prediction must retain acquisition and segment identifiers for spatial review.

Held-out calculation example

The rounded values below correspond to a row assigned to the testing subset. Its supplied target is other land cover. The exported model returns a wilt score of 0.020051 and therefore assigns the negative class at the 0.50 threshold.

Reproduce the calculation

The Python package contains the exact exported model, ordered input schema and executable example. The Neural Designer package contains the cleaned sample roles, initial and selected architectures, training settings and regenerated analyses.

8. Validity, uncertainty and limitations

• Derived sampling design. The tutorial table keeps all 500 source testing rows but adds only the 74 positive rows from the source training file. It excludes 4,265 negative training rows, then creates a new random split. The reported metrics are therefore not an evaluation on the official UCI test protocol.
• Outlier-selection boundary. Tukey thresholds are calculated before the model split and exclude seven negative rows. A confirmatory workflow should estimate cleaning thresholds from training data only and report sensitivity analyses with and without exclusions.
• Spatial dependence is untested. The table has no coordinates, acquisition blocks or segment-group identifiers. Random row splitting can place nearby or related image objects in different subsets.
• Image-processing contract is incomplete. The CSV does not package the pansharpening, segmentation, georeferencing or feature-extraction implementation needed to reproduce descriptors from raw imagery.
• Transfer is untested. Sensor calibration, illumination, season, atmospheric correction, geography, species mix and disease stage may shift the feature distributions.
• Training-curve warning. The initial run shows increasing selection error after its early minimum. Checkpointing at the best selection epoch and repeated grouped validation should precede any scientific conclusion.
• Uncalibrated decision score. The sigmoid output ranks candidates but is not demonstrated to be a calibrated probability. The 0.50 threshold has not been selected from surveillance costs.
• Endpoint boundary. A positive segment is a remote-sensing disease candidate, not laboratory confirmation of Japanese oak wilt, Japanese pine wilt or a causal organism.

References

• Johnson B. Wilt [Dataset]. UCI Machine Learning Repository; 2013. CC BY 4.0.
• Johnson BA, Tateishi R, Hoan NT. A hybrid pansharpening approach and multiscale object-based image analysis for mapping diseased pine and oak trees. International Journal of Remote Sensing. 2013;34(20):6969–6982.