---
title: 'EmissionNet (ENV): ML Models for Traffic & Agriculture'
url: https://www.emergentmind.com/topics/emissionnet-env
type: topic
---

# EmissionNet (ENV): ML Models for Traffic & Agriculture

EmissionNet (ENV) denotes two distinct machine-learning constructs in the 2025 literature. In urban transportation, ENV is an Emission–MFD-based, location-aware model for tract-level network emissions, centered on the macroscopic emission fundamental diagram (eMFD) and learned from probe traffic data plus MOVES-Matrix emissions labeling [2511.08722]. In agricultural air-quality forecasting, ENV is a pure convolutional spatio-temporal regression architecture for next-step prediction of gridded agricultural $N_2O$ emissions from multi-channel monthly emissions histories [2507.05416]. A third 2025 paper, on Group Reasoning Emission Estimation Networks (GREEN), does not define or reference a system called “EmissionNet (ENV),” but it describes components that could underpin a practical emissions estimation network in enterprise carbon accounting [2502.06874].

## 1. Scope and nomenclature

The label “EmissionNet (ENV)” is therefore context dependent rather than standardized across a single research lineage. In one usage, it refers to a macroscopic urban traffic emissions model built around $E_s = f(k_s, X_s, Veh_s)$ and intended for real-time monitoring, tract-level inference, and emissions-aware control [2511.08722]. In the other, it denotes a deep convolutional architecture that consumes a $24$-month context of gridded emissions maps and outputs the next monthly $N_2O$ field [2507.05416]. This suggests that the common label reflects a shared concern with emissions estimation or forecasting, but not a shared architecture, data model, or application domain.

| ENV context | Target | Core representation |
|---|---|---|
| Urban traffic | Tract-level emission intensity $E_s(t)$ and derived $\Phi_s(t)$ | $E_s = f(k_s, X_s, Veh_s)$ |
| Agricultural forecasting | Next-month $N_2O$ map $\hat Y \in \mathbb{R}^{H \times W}$ | $f_\theta: X \rightarrow \hat Y$, with $X \in \mathbb{R}^{T \times H \times W \times C}$ |

The distinction is material. The traffic ENV is explicitly location-aware, uses tract descriptors and fleet characteristics, and is empirically grounded in the eMFD literature. The agricultural ENV is a dense convolutional predictor whose inductive bias is multi-scale spatial extraction, dense connectivity, and channel attention over stacked temporal and molecular channels.

## 2. Emission–MFD-based ENV for urban traffic networks

In the urban traffic formulation, ENV is built on the macroscopic emission fundamental diagram for a network such as a census tract $s$ [2511.08722]. For links $a \in \mathcal{A}_s$ with length $L_a$ and $n_a$ lanes, the per-lane link variables are density $k_a(t)$, flow $q_a(t)$, and speed $v_a(t)$, with
$$
v_a(t) = \frac{q_a(t)}{k_a(t)}.
$$
Space-mean network aggregation over lane-miles gives
$$
L_s = \sum_{a \in \mathcal{A}_s} n_a L_a,
$$
$$
k_s(t) = \frac{1}{L_s}\sum_{a \in \mathcal{A}_s} n_a L_a k_a(t),
$$
$$
q_s(t) = \frac{1}{L_s}\sum_{a \in \mathcal{A}_s} n_a L_a q_a(t),
$$
$$
v_s(t) = \frac{q_s(t)}{k_s(t)}.
$$

Link-level running exhaust emission intensity $e_a(t)$ is defined in grams $CO_2$ per vehicle-mile and is obtained by coupling link activity and speed $v_a(t)$ with MOVES-Matrix, indexed by vehicle type, vintage, road type, and speed. Network-level emission intensity is the VMT-weighted average
$$
E_s(t) = \sum_{a \in \mathcal{A}_s} w_a(t)e_a(t),
\qquad
w_a(t) = \frac{q_a^{tot}(t)L_a}{\sum_{b \in \mathcal{A}_s} q_b^{tot}(t)L_b},
$$
where $q_a^{tot}(t) = n_a q_a(t)$ and
$$
VMT_s(t) = \sum_{b \in \mathcal{A}_s} q_b^{tot}(t)L_b.
$$
The total emission rate is then
$$
\Phi_s(t) = \sum_{a \in \mathcal{A}_s} e_a(t) q_a^{tot}(t)L_a = E_s(t)\cdot VMT_s(t).
$$

The eMFD posits a consistent relationship among aggregated traffic states and emissions. In its simplest tract-specific form,
$$
E_s(t) = f_s(k_s(t)),
$$
and more generally,
$$
E_s(t) = f_s(k_s(t), q_s(t), v_s(t)).
$$
The paper further makes the location and fleet dependence explicit:
$$
E_s(t) = f(k_s(t), X_s, Veh_s),
$$
where $X_s$ encodes network, infrastructure, and land-use factors such as development level, street and intersection density, road class composition, job centers, bike/walk potential, and topography, while $Veh_s$ summarizes fleet characteristics such as vintage mix and LDV versus other classes.

A central methodological point is that the learned $f$ is nonparametric. The study represents the eMFD with tree ensembles rather than a closed-form polynomial or spline. A convenient parametric representation with XGBoost is
$$
f_\theta(x) = \sum_{m=1}^M \lambda h_m(x),
\qquad
x = [k_s, X_s, Veh_s],
\qquad
\lambda > 0,
$$
where $h_m$ are shallow regression trees. The resulting tract-specific eMFDs support region-wide emissions monitoring and provide a basis for assignment and perimeter-control formulations.

## 3. Data, learning protocol, and deployment in the traffic setting

The traffic ENV draws on HERE probe vehicle data for the full U.S. from September to November 2019 at $15$-minute resolution, with variables including network geometry, traffic counts, speeds, and number of probes [2511.08722]. Segment-level speed and volume are aggregated in space-mean fashion to derive per-link flow $q_a(t)$ and density $k_a(t)$, then aggregated to tract-level $k_s(t)$, $q_s(t)$, and $v_s(t)$. Emissions labels are created by querying MOVES-Matrix using link average speed $v_a(t)$ and categorical inputs for vehicle type, vintage, and road type. In the study, New York, Colorado, Texas, and Georgia have emissions labels via MOVES-Matrix; across these states, $13{,}022$ urban census tracts and $15$-minute intervals produce $6{,}629{,}818$ rows for modeling.

Location features are transformed into tract-level factors following the probe-data macroscopic modeling framework cited as [7] in the source paper. These factors include development level, network complexity, local roads share, principal/non-freeway arterials, freeway share, long streets, job centers, bike/walk potential, topography such as hilly/circular roads, and median travel. Fleet is represented with bounding light-duty vehicle vintages, “older” ($2000$ and earlier) versus “newer” ($2018$), yielding indicators such as `vehtype_L1`. Preprocessing includes normalization and stratification by road types and development factors to ensure consistent aggregation across heterogeneous networks.

The target variable is tract-level emission intensity $E_s(t)$ in $g\ CO_2/\text{veh-mile}$. The evaluated models are Random Forest, XGBoost, LightGBM, and Linear SVM, trained with an $80\%$ training and $20\%$ testing split. Performance is reported in terms of $R^2$, MAE, RMSE, and MAPE. XGBoost is the best-performing model on the test set, with $R^2 = 0.92$, MAE $= 16.25\ g/\text{veh-mile}$, RMSE $= 49.83\ g/\text{veh-mile}$, and MAPE $= 2.87\%$. Random Forest and LightGBM both achieve $R^2 = 0.91$ but with substantially larger errors, while Linear SVM reaches $R^2 = 0.82$.

Interpretability is handled with TreeExplainer and SHAP interaction values with density. The most influential features on the eMFD are “development level” and `vehtype_L1`. Both exhibit pronounced divergence beyond $k_s \approx 20\ \text{veh/lane/mile}$. High development level tracts show lower $E_s$ at the same $k_s$, and newer LDVs yield lower $E_s$ relative to older LDVs, especially under high density. The paper also reports location heterogeneity in MFD and eMFD shapes, including a contrast between tracts with lower and higher network capacity and a New York City example in which tracts with high density at $5$ PM also exhibit higher per-mile $CO_2$ emission intensities.

Although trees perform best, the paper provides a neural alternative aligned with the same framework. This ENV variant takes
$$
x = [k_s(t), q_s(t), v_s(t), X_s, Veh_s]
$$
as input, uses hidden layers such as `Dense(64, ReLU) → Dense(64, ReLU) → Dense(32, ReLU)`, outputs $\hat y = E_s(t)$, and minimizes
$$
L(\theta) = \frac{1}{N}\sum_{i=1}^N \left(E_s^{(i)} - f_\theta(x^{(i)})\right)^2 + \lambda \|\theta\|_2^2.
$$
Training uses Adam, early stopping on validation MSE, an $80/20$ split, input standardization for continuous features, and one-hot or embeddings for categorical road types. For deployment, the study emphasizes that inference can use time-resolved aggregated traffic measurements at $15$-minute granularity or faster, together with static $X_s$ and $Veh_s$ profiles. Given $k_s(t)$ and tract descriptors, ENV outputs $E_s(t)$, after which $\Phi_s(t)$ follows from $VMT_s(t)$. Space-mean aggregation is presented as a robustness mechanism, with temporal smoothing, imputation, and regional fallbacks suggested for missing data, and periodic retraining such as quarterly updates suggested for adaptation.

## 4. Convolutional ENV for agricultural $N_2O$ forecasting

In the agricultural forecasting paper, EmissionNet (ENV) is a pure convolutional architecture for spatio-temporal regression on global gridded emissions data [2507.05416]. The task is next-step forecasting of spatially resolved agricultural nitrous oxide emissions from a multi-year context of monthly emissions maps. The input is
$$
X \in \mathbb{R}^{T \times H \times W \times C},
$$
with $T = 24$ months and $C = 5$ channels corresponding to $CH_4$, $CO_2$, $N_2O$, $CO2bio$, and GWA. The supervised objective is
$$
f_\theta: X \rightarrow \hat Y,
$$
where $\hat Y \in \mathbb{R}^{H \times W}$ is the next-month $N_2O$ field. The primary setup uses a single-step horizon $\tau = 1$, while evaluation also includes auto-regressive multi-step roll-outs in which previous predictions can enter the context window.

The data source is EDGAR GHG emissions from $2000$ to $2023$ at $0.1^\circ$ resolution over latitudes $[-90^\circ,+90^\circ]$ and longitudes $[-180^\circ,+180^\circ]$. Preprocessing pools spatially to $0.3^\circ \times 0.3^\circ$ and discards flux dimensions, yielding a tensor of shape $(5, 288, 600, 1200)$, where $288 = 24$ years $\times 12$ months. A rolling-window context uses $24$ months to predict the next month’s $N_2O$. The split is Train Jan 2000–Mar 2019, Val Apr 2019–Jul 2021, and Test Aug 2021–Jan 2024. The data are described as exhibiting strong seasonality and spatial heterogeneity aligned with agricultural cycles and continental versus oceanic contrasts.

ENV represents time by stacking the $24$ context frames and $5$ molecular channels along the channel axis, so that standard $2$D convolutions jointly mix spatial and temporal/molecular dimensions. The input head consists of two $3 \times 3$ convolution layers with stride $2$, each followed by batch normalization and ReLU:
$$
F_1 = \text{ReLU}(\text{BN}(\text{Conv}_{3 \times 3, s=2}(X))),
$$
$$
F_2 = \text{ReLU}(\text{BN}(\text{Conv}_{3 \times 3, s=2}(F_1))).
$$

The backbone then applies three multi-scale feature extraction modules. Each module uses parallel branches with kernel sizes $1 \times 1$, $3 \times 3$, $5 \times 5$, and a pooling branch, concatenated channel-wise:
$$
B = [H_{1 \times 1}(A), H_{3 \times 3}(A), H_{5 \times 5}(A), H_{pooling}(A)].
$$
This is followed by four implicit deep supervision modules with dense skip-concatenation in a DenseNet-style form,
$$
x_l = H_l([x_{l-1}, x_{l-2}, \ldots, x_0]),
$$
where each $H_l$ is BN $\rightarrow$ ReLU $\rightarrow$ Conv $3 \times 3$. To control channel growth, each IDS module ends with a $1 \times 1$ convolution for channel compression and a $2 \times 2$ max pool with stride $2$.

A further architectural component is squeeze-and-excitation-style channel attention between basic layers in each IDS module:
$$
z_c = \frac{1}{H \cdot W}\sum_{i=0}^{H-1}\sum_{j=0}^{W-1} u_c(i,j),
$$
$$
s_c = \sigma(w_2 \otimes \delta(w_1 \otimes z_c)),
$$
$$
x_{out} = s_c \cdot u_c.
$$
Here, $u_c$ is the $c$-th channel feature map, $z_c$ is a global average-pooled descriptor, $\delta$ is ReLU, and $\sigma$ is sigmoid. The final output head is a $2$D convolution projecting to a single $N_2O$ emission map. No explicit positional encoding or transformer attention is used in ENV.

Training minimizes mean squared error,
$$
L_{MSE} = \frac{1}{N}\sum_{i=1}^N \|y_i - \hat y_i\|_2^2,
$$
with AdamW, dynamic learning rate, warmup, weight decay $1 \times 10^{-4}$, learning rate $3 \times 10^{-3}$, warmup ratio $0.2$–$0.3$, batch size $8$ for deeper models, and $200$ epochs. The paper reports no additional regularization beyond weight decay.

## 5. Empirical performance and comparative behavior

The two ENV systems are evaluated in very different regimes, so their metrics are not directly comparable. The traffic ENV predicts tract-level $CO_2$ emission intensity in physical units, while the agricultural ENV predicts gridded next-step $N_2O$ fields and is primarily scored by MSE [2511.08722; 2507.05416].

In the traffic setting, the benchmark comparison among tabular regressors is as follows:

| Model | Test performance |
|---|---|
| XGBoost | $R^2 = 0.92$; MAE $= 16.25$ g/veh-mile; RMSE $= 49.83$ g/veh-mile; MAPE $= 2.87\%$ |
| Random Forest | $R^2 = 0.91$; MAE $= 119.32$; RMSE $= 240.50$; MAPE $= 7.58\%$ |
| LightGBM | $R^2 = 0.91$; MAE $= 104.48$; RMSE $= 245.00$; MAPE $= 6.83\%$ |
| Linear SVM | $R^2 = 0.82$; MAE $= 171.46$; RMSE $= 345.03$; MAPE $= 15.42\%$ |

The result that $R^2 \approx 0.92$ indicates that density plus location and fleet features explain most variance in $E_s$ across tracts and times. The paper further states that interactions become pronounced at medium-to-high densities, especially for development level and fleet vintage.

In the agricultural setting, ENV is compared with MLP, ConvLSTM, and EmissionNet-Transformer (ENT):

| Model | Test MSE |
|---|---|
| EmissionNet (ENV) | $0.00010$ |
| EmissionNet-Transformer (ENT) | $0.00068$ |
| ConvLSTM | $0.00156$ |
| MLP | $0.06288$ |

The paper reports relative improvements of ENV versus ConvLSTM at approximately $93.6\%$, versus ENT at approximately $85.3\%$, and versus MLP at approximately $99.8\%$. It also reports a parameter count of approximately $11.7$M for ENV, versus approximately $0.71$M for ENT, approximately $1.3$M for ConvLSTM, and approximately $32$M for the MLP. ENV has the best single-step accuracy, whereas ENT is described as more stable in long auto-regressive roll-outs, with RMSE approaching an asymptote over extended horizons. By contrast, ENV exhibits faster error growth during long roll-outs due to weaker modeling of long-range temporal dependencies.

The traffic and agricultural studies also differ in interpretability strategy. The traffic paper uses TreeExplainer and SHAP interaction values to identify development level and fleet vintage as dominant modifiers of the density–emissions relation. The agricultural paper primarily uses qualitative map comparison, noting that ENV avoids ocean offsets observed in ConvLSTM and maintains high fidelity in both high-gradient and smooth regions. Formal saliency, SHAP, or occlusion analyses are not reported იქ in the agricultural study; the paper explicitly notes that no explicit calibration or uncertainty quantification is reported.

## 6. Related frameworks, limitations, and prospective extensions

A separate line of work, GREEN, is adjacent to but distinct from ENV [2502.06874]. That paper states explicitly that it “does not define or reference a system called ‘EmissionNet (ENV).’” GREEN is an end-to-end enterprise emissions estimation framework based on text-driven sector classification, Group Reasoning over the NAICS ontology, and EE-MRIO-linked carbon intensity assignment. The same source further states that a practical EmissionNet could adopt GREEN’s components: an enterprise-to-sector classifier framed as information retrieval, a Group Reasoning hierarchical ensemble, an aligned economic model linking sector labels to carbon intensity factors, and an emissions inference module computing emissions from intensity and revenue. This suggests that “EmissionNet” is also emerging as a broader naming pattern for emissions-oriented ML systems beyond the two ENV definitions above.

The limitations of the traffic ENV are domain specific. The paper lists probe bias due to varying HERE coverage, emissions labeling coverage limited to four states, a simplified fleet representation using only two LDV vintages, transferability challenges for unseen cities, model drift as infrastructure and travel demand evolve, the need for dynamic models and equity/access constraints in policy integration, and limited mechanistic interpretability of tree ensembles. It proposes richer fleet mix modeling, extension beyond MOVES-Matrix coverage, fine-tuning for new cities, and hybrid approaches such as GAM plus trees or physics-informed ML.

The limitations of the agricultural ENV are likewise explicit. The study relies on single-step training with auto-regressive evaluation, which makes long-horizon forecasts susceptible to compounding error. It reports no explicit uncertainty quantification, no explicit meteorological drivers beyond the emissions channels, and no graph-based modeling or adjacency matrices. Suggested directions include scheduled sampling, teacher forcing, multi-horizon loss, multi-resolution heads, hybrid conv-attention designs such as ENT, positional encodings, uncertainty-aware objectives, and integration of additional environmental drivers.

Across these usages, ENV functions less as a single canonical model than as a recurring label for emissions-focused ML systems. In transportation, it operationalizes a learned, location-aware eMFD that links network density, infrastructure, land use, and fleet composition to tract-level $CO_2$ intensity and total emissions. In agricultural forecasting, it denotes a deep convolutional architecture that exploits stacked temporal context, multi-scale spatial processing, dense connectivity, and channel attention to predict monthly $N_2O$ maps. The shared theme is the replacement of sparse empirical or physics-driven formulations with learned nonlinear mappings whose structure is tailored to the aggregation level and control objective of the application domain.

Source: https://www.emergentmind.com/topics/emissionnet-env