Influence of sample distribution and prior probability adjustment on land cover classification
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Machine learning algorithms are widely used for remote sensing land surface characterization. Successful implementation requires a representative training sample for the domain it will applied in (i.e. area of interest or validation domain). However, accessibility and cost strongly limit the acquisition of suitable training samples for large regional applications. Further, it is often desirable to use previously developed datasets where significant resources have been invested, such as data developed from extensive field survey or high resolution remotely sensed imagery. These data often only partially represent the domain of interest and can lead to various forms of sample bias (land cover distribution or class properties). Classifier spatial extension is an extreme case, where a sample is trained from one region (i.e. sample domain) and applied in another (i.e. application domain). This approach is desirable from a cost perspective, but achieving acceptable accuracy is often difficult. In this research we investigate two approaches to account for possible differences between the sample and application domain land cover distributions. The first is an iterative resampling approach to predict the application distribution and adjust the sample distribution to match. The second is the use of prior probabilities to adjust class memberships. Results reinforce the importance of the land cover distribution on accuracy for algorithms that are designed to minimize the classification error with training data. Of the adjustments methods tested resampling was superior if the application domain distribution was well known. However, if it is not then the use of prior probabilities performed similarly overall. A generic model was developed to predict if resampling or prior adjustment should be applied to enhance accuracy.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it