Satellite studies of Chinese cities agree that built-up land has grown much faster than population, and that the growth is driven by economics rather than housing need. Manufacturing needs large flat parcels, and local governments that depend on land-sale revenue have reasons to keep converting farmland. These claims, however, rest on maps that record only "impervious surface", meaning any paved or built ground. Nobody has measured, city by city, how much newly built land is factories versus housing versus commercial districts. The industrial explanation is inferred by correlating total built-up growth with aggregate industry statistics. Telling an industrial park from a dense residential compound in imagery is a judgment that needs human experts, and expert labels are scarce.
We propose to spend a fixed expert budget where the classifier is torn between two functional types. This is margin sampling: an active-learning rule, where active learning means the model chooses which examples a human labels next, that queries examples whose top-two predicted class probabilities are nearly tied. We draw those queries randomly, with known probabilities that grow as the margin shrinks. The same labels then do two jobs. They train the classifier, and they correct its city-level area totals so the totals stay unbiased. Our claim is that this design gives tighter and unbiased estimates of industrial and residential land consumption than random labeling at the same budget, and that the usual way of running margin sampling gives biased economic conclusions.
Core Claim
Margin sampling is aimed at pairwise confusion. In the active-learning literature it is preferred over entropy or least-confidence because it targets the boundary between the two leading classes (Schein et al.; Younesian et al.). That is the confusion that matters here: industrial versus residential versus commercial built land (Gap 5 of the file). The obvious use of margin sampling is to train a better classifier and then read areas off its map, and that is a mistake for economic inference. The labeled set is then a non-random sample, and the map's errors are not random either. They sit on the industrial/residential boundary that the economics cares about. Regressions run on such a map inherit bias that no one can measure.
Our fix uses one probability per candidate site. Let $M(x_i)$ be the top-two probability gap for site $i$, $a_i$ its area, $\hat y_i$ the classifier's predicted indicator for "industrial", and $y_i$ the expert label. Site $i$ is labeled with probability $\pi_i = n\,w_i/\sum_j w_j$, where $w_i = \epsilon + 1 - M(x_i)$ and $\epsilon > 0$ guarantees every site some chance of being labeled. Industrial area in city $c$ is then estimated by a difference estimator:
$$\hat A_c = \sum_{i \in c} a_i \hat y_i + \sum_{i \in c \cap S} \frac{a_i (y_i - \hat y_i)}{\pi_i},$$
where $S$ is the labeled set. The estimator is unbiased for any classifier, because the correction term is a standard inverse-probability total. Its variance is driven by the squared errors $(y_i-\hat y_i)^2$ divided by $\pi_i$. Errors concentrate at small margins, so sampling there with high probability is close to the variance-minimizing design. In this setting margin sampling is therefore a variance-reduction design for measurement, and not only a heuristic for training. The labeling is done in rounds, with the classifier retrained between rounds. Each round's correction uses that round's known probabilities, so the per-round estimates remain unbiased and can be averaged.
Research Program
We start from the conversion events that existing work already counts. These are the cells that became impervious between 2000 and 2010 in GlobeLand30 across the prefecture-level cities analyzed by Dai et al.. That study reports a multilevel regression in which GDP growth (coefficient 0.289), the ratio of land-transfer fees to fiscal revenue (0.163), and baseline cultivated land (0.18) predict total impervious expansion.
Converted cells are grouped into parcels, and each parcel gets a high-resolution image chip. A classifier on top of a pre-trained geospatial backbone predicts the functional type: industrial, residential, commercial or public, transport infrastructure, and cleared-but-unbuilt. We use CrossEarth features via VisionXLab/CrossEarth, since it was built to generalize across regions. The cleared-but-unbuilt class matters because land hoarding is one of the distortions Dai et al. blame for gaps between official statistics and satellite data.
Two known failure modes of uncertainty sampling get specific treatment. The first is cold start, where a weak first model queries along the wrong boundary. Following Doucet et al., who found that on pre-trained features one to three rounds of diversity sampling (TypiClust, which picks representative points in dense regions of feature space) suffice before switching to margin sampling, we run a short diversity phase. That phase is also drawn with known probabilities. The second is spatial redundancy: neighboring parcels share a margin, so queries pile up in one corridor. Liu et al. handled this in binary agricultural change detection by checking neighborhood agreement. Rather than reuse that check, we sample within city strata and use systematic spatial ordering. This spreads queries without breaking the known probabilities.
The output is a city-by-type panel of land consumption with standard errors. On that panel we re-estimate Dai et al.'s model separately for industrial and residential conversion. The goal is to test directly a claim that so far has only been inferred: that secondary industry is the dominant consumer of land (Zuo et al.; Ren et al.; Ji et al.). We also ask which functional type carries the land-finance and cultivated-land effects. Jia et al. found that small and large metropolitans grow differently, so city size enters as an interaction term.
Baselines
The naive reference is the classifier's own map with no correction, with its labels chosen at random. This is what most driver studies effectively do with their land-cover inputs. Any design should beat it on bias. It fixes the floor for how wrong area estimates get when classifier errors are ignored.
The strongest competitor for measurement is stratified random labeling by city, combined with the same difference estimator. This is the textbook design for unbiased area estimates. It differs from our proposal only in using uniform rather than margin-weighted probabilities, and it is the real threat. It uses the same backbone, the same per-round retraining, the same total budget, and the same strata. Beating it means a smaller variance for the city-level industrial-area estimate at equal budget. Equivalently, it means a smaller budget for the same confidence-interval width, with coverage of nominal 95% intervals still near 95%.
The strongest competitor for classifier quality is semi-supervised training from a small random labeled set, in the style of RS-MTDF. That method distills DINOv2 and CLIP into a student model and reports 51.21% mIoU on LoveDA with 1% of labels. For comparison, fully supervised HRNet reports 49.79% test mIoU (LoveDA). For classifying the functional type of newly built land we found no reported result on Lacuna, so that task has no published reference number.
Two comparisons are ablations rather than competitors. The first is deterministic argmin-margin selection with no correction, the usual way margin sampling is run. It isolates the cost of losing known probabilities. The second keeps the full weighted design but sets $w_i$ by entropy instead of margin. It isolates whether the pairwise top-two gap matters, given that the confusions of interest are between pairs of classes. All arms share the backbone, tuning budget, and total label count.
Why not just…?
Why not just label a stratified random sample? It is unbiased and simple, and if classifier errors were spread evenly across sites it would match our design. The case for weighting rests on errors being concentrated at small margins. If they are not, the gain vanishes. We measure this directly by comparing variance ratios, both on held-out data and on LoveDA, where full labels let us compute the exact sampling variance of each design. If the variance reduction is negligible, we will report stratified random sampling as the recommendation.
Why not just use points of interest or other existing functional maps as labels? Points-of-interest data already enter urban simulation as predictors (Liu et al., 3D FLUS), and they cost nothing. But they record where businesses are listed, not what is built on a parcel. Their coverage is also likely to be thinnest in the small cities where Jia et al. found the most direct economic sprawl. We include a classifier trained on labels derived from points of interest as an arm and score it against the expert evaluation sample. If it is as accurate, expert labeling is unnecessary.
Why not just run semi-supervised learning on a foundation model and skip query design? Semi-supervised training improves the classifier, but it does not make area estimates unbiased. On LoveDA, the self-training adaptation methods CBST (41.32% mIoU) and IAST (40.48%) stay below the supervised oracle (45.67–46.79%). That gap shows how pseudo-labels can systematically mislead. The two approaches are complementary: a better classifier simply shrinks the residuals our estimator corrects. We therefore also run our design on top of an RS-MTDF-style student model.
Why not just apply active learning to the suitability models of land-use simulators such as PLUS? In those models the training labels are historical transitions that are already fully observed (Fu et al.; He et al.). Labels are free, so query selection saves nothing. The scarce quantity is the functional type of the converted land, which is what we target.
Proposed Experiments
The first check is a controlled simulation on LoveDA, where every pixel is labeled. We treat building area per administrative district as the target and replay each design many times. Across replays we measure bias, variance, and interval coverage. The prediction is that argmin margin is visibly biased, and that the margin-weighted design has lower variance than stratified random sampling at budgets of 1–5% of tiles.
The second experiment is the real study. We label the functional type of converted parcels in Dai et al.'s cities under each design. A separate stratified random evaluation sample, double-annotated by experts, serves as ground truth for both map accuracy and city totals.
The decisive economic check asks whether conclusions change across designs. We compare the land-finance and GDP coefficients for industrial versus residential conversion under the corrected margin design and under argmin margin. If the coefficients differ beyond their standard errors, uncorrected active labeling would have changed the substantive answer.
We also test generalization across city tiers by holding out small cities, which tests whether the cold-start diversity phase is needed when the economic regime shifts. The theory of Balcan and Zhang on margin-based learning under heavy-tailed distributions assumes isotropic data, which spatially clustered parcels violate. We therefore treat it as motivation only and rely on the empirical variance ratios as evidence.
References
- Active learning for logistic regression: an evaluation
- End-to-End Learning from Noisy Crowd to Supervised Machine Learning Models
- Examining the Driving Factors Causing Rapid Urban Expansion in China: An Analysis Based on GlobeLand30 Data
- CrossEarth: Geospatial Vision Foundation Model for Domain Generalizable Remote Sensing Semantic Segmentation
- Bridging Diversity and Uncertainty in Active learning with Self-Supervised Pre-Training
- Change Detection for Hyperspectral Imagery Based on Multi-layer Cascade Screening Strategy
- Analysis of Temporal and Spatial Characteristics of Urban Expansion in Xiaonan District from 1990 to 2020 Using Time Series Landsat Imagery
- Evaluating the effects of urbanization on land use and landscape patterns in southern China (1990–2020)
- Spatial-Temporal Changes in Land Use and Their Driving Forces in the Circum-Bohai Coastal Zone of China from 2000 to 2020
- Characterizing the Growth Patterns of 45 Major Metropolitans in Mainland China Using DMSP/OLS Data
- RS-MTDF: Multi-Teacher Distillation and Fusion for Remote Sensing Semi-Supervised Semantic Segmentation
- LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation
- Three-Dimensional Simulation Model for Synergistically Simulating Urban Horizontal Expansion and Vertical Growth
- Dynamic Simulation and Prediction of Carbon Storage Based on Land Use/Land Cover Change from 2000 to 2040: A Case Study of the Nanchang Urban Agglomeration
- Landscape Ecological Risk Assessment of Kriya River Basin in Xinjiang and Its Multi-Scenario Simulation Analysis
- Sample and Computationally Efficient Learning Algorithms under S-Concave Distributions