Internal Texture Model Evaluation Report

Summary of Results

• The texture model is performing acceptably and does not have any obvious issues, and should be accepted for deployment
• There are no extreme outliers or anomalies found
• The CST Model is worse overall than the model we showed A&L, but not significantly worse, see notes below
• New CST Model ALP test pass rate is high for clay (78%), but not as good as initially expected for Sand (65%) and Silt (70%)
• CST Model CV outliers are worse for silt and sand, but better for clay than the A&L-accepted model.

Baseline: A&L-Approved Cross Validation

A&L accepted the side-by-side data for a grid model we developed in June 2025. This model architecture was incorrect - it was overfitted. As a result, we will likely never see this level of accuracy in a texture model. The results are here anyways as a reminder that this is what A&L thinks is possible / this is what they roughly expect. We don’t have inference for the ALP trays or other test sets with this model, so we only can compare new models to the cross validation results from this A&L-approved grid model, and evaluate test sets independently. The CV results are below.

Figure 1

A&L-Approved CV Results r2 Original Grid Model

Stat	Clay	Sand	Silt
R2	0.651	0.763	0.651
MAE	7.049	9.387	8.435

Table 1 Shows the R2 and MAE values to beat in the original approved grid texture model.

Baseline: A&L-Approved CV Outliers

Figure 2

New CST Model

The current CST model being evaluated predicts texture and is up for deployment at A&L on EVT27.

CST Model Cross Validation Result

The CST Model was trained on 300 A&L samples with hydrometer data only.

Figure 3

CV Results r2 New CST Texture Model

Stat	Clay	Sand	Silt
R2	0.745	0.690	0.491
MAE	5.998	11.105	9.557

Table 2

Figure 4

Result:

The new Model r2 and MAE are overall worse than the original overfitted model. The new r2 and MAE are better for Clay, but worse in Sand and Silt compared to the original model A&L accepted. Based on the CV data, the new CST model is better than the original model on outliers for clay, but has more outliers in sand and silt than the original A&L accepted model.

ALP Proficiency Tray Pass/Fail Result (Averaged Data)

This tray is used to evaluate performance on the proficiency program

Figure 5

ALP Pass Rate of Sand, Silt and Clay (Non-Averaged Data)

ALP Test Tray EVT27 Pass Summary when Replicates are not averaged within sample

Clay Pass Rate (%)	Sand Pass Rate (%)	Silt Pass Rate (%)
78.33	65	70

Table 3

Precision Tray

This tray has been scanned extensively across digitizers over time. We use the precision tray to assess cross-digitizer performance in precision and accuracy over time.

Figure 6

Texture Accuracy Tray

This tray evaluates predictions against NAPT-proficient particle size boundaries for an indication of the model’s geographic generalizability.

Figure 7

SOM Eval Tray

The evaluation of this tray is meant to asses the precision of the model on repeated samples from different geographic origins to the training dataset, and to assess the accuracy and precision of the model on a widely-used internal QC sample, SRS2001.

Figure 8

Canadian Precision Tray

The evaluation of this tray is meant to assess the precision of the model on sample replicates, and generalizability of the model to samples of diverse geographic origin.

Figure 9

NAPT test samples

Figure 10