Landscape-aware routing trades distance for lower urban conflict

A landscape-aware ant colony optimization framework produced lower modeled environmental conflict in urban waste-collection routes while maintaining a favorable trade-off between travel distance and spatial intrusion. On the primary synthetic benchmark, the approach reduced its Landscape Conflict Index by 31.1% at a 13.9% increase in route distance. A separate 80-customer Zhengzhou-inspired scenario showed a 27.3% reduction in conflict for a 7.8% distance increase. The study is a computational evaluation of simulated routing problems, not a trial on a live municipal fleet, and its results indicate how collection planners could represent the locations of visual, acoustic and ecological sensitivity rather than treating every street as equivalent.

The researchers formulated the collection task as a capacitated vehicle-routing problem with a homogeneous fleet, a central depot and customer locations with assigned demand. To represent environmental differences among places, they constructed a continuous landscape-sensitivity field using superimposed Gaussian radial basis potentials. Each potential has an intensity and a spatial decay scale, allowing designated features to be represented as peaks that influence surrounding road segments. Separate layers can describe visual, acoustic and ecological sensitivity, including anchors associated with scenic areas, schools, hospitals, wetlands and culturally sensitive locations. For each candidate edge between routing nodes, the field was sampled at ten evenly spaced points and averaged along the edge to produce a landscape cost.

From those edge costs, the team derived a dimensionless Landscape Conflict Index, or LCI, defined as the total landscape cost of selected route segments divided by their combined length. A lower LCI indicates that a route is concentrated in less-sensitive areas, whereas a higher value indicates greater overlap with sensitive landscape. Unlike a fleet-wide measure of emissions, LCI is explicitly spatial and can be compared across route configurations. It can also be separated by road segment. In the primary instance, about 12% of edges accounted for more than half of the total modeled landscape cost, while the distribution of edge-level values was right-skewed, with a skewness of 1.42. This decomposition makes it possible to identify geographic segments that dominate the model’s conflict estimate.

The resulting Landscape-aware Ant Colony Optimization algorithm, LA-ACO, embeds landscape information in two stages of its search. During route construction, a generalized heuristic makes candidate edges with high landscape cost less likely to be chosen, suppressing sensitive edges before a complete route is assembled. At the end of each iteration, the pheromone update also accounts for LCI, giving stronger reinforcement to solutions with lower conflict. A generalized objective combines travel distance and weighted landscape cost, and an iterative 2-opt refinement evaluates route rearrangements using that same objective rather than distance alone. Environmental sensitivity therefore shapes both short-term exploration and the algorithm’s accumulated memory instead of being applied only as a post hoc penalty.

Across ten random seeds on the primary benchmark of 50 customers, LA-ACO reached a mean generalized cost of 2,602.7, compared with 2,882.4 for standard ACO, a 9.4% reduction. The improvement was present in nine of ten paired runs; one outlier favored standard ACO by 7.6%. A one-sided paired Wilcoxon signed-rank test gave a p-value of 0.002. The reported final costs of genetic algorithm, particle swarm optimization and simulated annealing were 46–58% above that of LA-ACO. LA-ACO also had the narrowest 95% confidence band, with a coefficient of variation below 3.8% at convergence. Its normalized Shannon entropy for edge-use diversity changed only from 0.886 to 0.883 over 80 iterations, a 0.3% decline, so the narrow spread was not accompanied by an obvious collapse in colony diversity.

Comparison with NSGA-II exposed a different computational trade-off. On the primary instance, LA-ACO achieved a Pareto hypervolume of 198.4, compared with 104.1 for NSGA-II, and the lowest inverted generational distance, 0.048. Its reported front fully dominated those of NSGA-II, genetic algorithm and particle swarm optimization in the set-coverage analysis. Simulated annealing, however, remained mutually non-dominated with LA-ACO and could reach lower absolute LCI at individual operating points by accepting substantially longer routes. Across the complete distance–LCI curves for all five benchmark types, LA-ACO had a total hypervolume of 459.8 against 401.0 for simulated annealing. The latter dominated only a small portion of the radial instance’s LA-ACO front. Reproducing the full LA-ACO sweep required 28.5 seconds on the same hardware, compared with 6.0 seconds for one NSGA-II run at 50 customers, because LA-ACO recovers the front by sweeping the landscape weight.

A full factorial analysis varied the landscape weight, alpha, and the heuristic exponent across a 10-by-8 grid, with three replications at each setting. A stable region emerged when alpha was between 2 and 6 and the heuristic exponent was between 1.0 and 3.0. At alpha equal to zero, landscape information was ignored, and the primary benchmark produced a mean route distance of about 742 kilometers and an LCI of 0.72. Raising alpha generally increased distance and reduced conflict, but the gains diminished at high values. At the primary case’s empirically identified knee point of alpha two, LCI fell by 31.1% while distance rose by 13.9%, equal to 2.2 percentage points of LCI reduction for each percentage point of additional distance. The researchers emphasized that alpha two is specific to the tested configuration and should be recalibrated for a different city, fleet, sensitivity field or policy objective.

Robustness testing simultaneously perturbed the intensities, spatial scales and positions of sensitivity anchors to represent uncertainty in environmental mapping. At relative perturbations of up to 30%, the median and interquartile range of LCI shifted by less than 0.04, while the coefficient of variation remained below 4.2%. At the extreme 50% perturbation, the coefficient rose to 5.8%, although the median LCI remained within 0.06 of baseline. The edge-cost calculation integrates ten sampled sensitivity values, which the authors argued can attenuate independent point-level errors and help retain the relative ordering of high- and low-conflict edges. In the Zhengzhou-inspired case, moving from alpha zero to two reorganized routes around overlapping visual, acoustic and ecological peaks, producing the reported 27.3% LCI reduction with a 7.8% distance increase.

Additional tests examined whether LCI corresponded to a measurable physical surrogate and whether the approach could operate beyond a Euclidean layout. Across 54 edges from an optimized route ensemble, edge-level LCI was strongly associated with predicted road-traffic noise from the ISO 9613 model: Pearson’s r was 0.933 and Spearman’s rho was 0.918, both with p-values below 0.001. These are associations between a modeled sensitivity metric and a modeled acoustic quantity, not measurements from collection vehicles or residents. On a synthetic 100-node Manhattan-grid road network using city-block distances, the framework again identified alpha two as a knee point. That operating point reduced landscape conflict by 20.4% for a 3.2% increase in road distance. The road test was nevertheless based on a synthetic grid, because the researchers could not retrieve a full OpenStreetMap city network in their computing environment.

Several limitations constrain how the findings can be interpreted. The sensitivity field is static and therefore does not represent time-dependent disturbance around schools, parks, residences or special events. The Zhengzhou-inspired instance is also synthetic, and the study did not use historical GPS trajectories, actual collection schedules or comprehensive municipal sensitivity data. Local search primarily rearranges customers within individual routes rather than moving them between vehicles, correlated errors in sensitivity mapping were not fully examined, and a systematic scalability study across larger problem sizes was not conducted. A single global alpha also simplifies policy preferences that may vary among districts and environmental criteria. The transferable contribution is therefore not a universal setting of alpha, but a procedure for constructing and calibrating a spatial sensitivity field, tracing the distance–conflict frontier and identifying an operating point that can be recalibrated for a particular collection network.

Subject of Research: Multi-objective optimization of urban waste-collection routes using spatial landscape-sensitivity modeling

Article Title: A landscape-aware multi-objective ant colony optimization framework for urban waste collection vehicle routing

Article References: Ma, S., & Liu, C. (2026). A landscape-aware multi-objective ant colony optimization framework for urban waste collection vehicle routing. Discover Artificial Intelligence, 6(1), Article 1088. https://doi.org/10.1007/s44163-026-02159-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02159-2

Keywords: urban waste collection, vehicle routing, ant colony optimization, landscape sensitivity, landscape conflict index, multi-objective optimization, urban logistics, route planning, environmental disturbance, landscape-aware, multi-objective, colony