Page 109 - EER-2-2

Explora: Environment
and Resource Climate suitability of AWD practice

using the Penman-Monteith equation to characterize A simple averaging technique across the MaxEnt outputs
water and energy exchanges between the land surface and was used as a consensus approach to estimate the central
the atmosphere. These predictors measure the quantitative tendency. 15
surplus or deficit of paddy field water balance, biomass, Further, the geospatial predictive assessment criteria
and rice crop yield. 24
include the area under the receiver operating characteristic
2.2.4. Predictive modeling of paddy rice using MaxEnt curve (AUC) (Table S3). Specificity, the prediction of
background points, sensitivity, and precise presence were
MaxEnt machine learning tool is widely applicable considered for the model’s presentation, which was vital for
for predictive modeling owing to its robust predictive improving the accuracy level of the model. In addition, the
15
performance, small data requirements, and straightforward precipitation threshold helped to define the limits of rainfall
model setup. The model uses the available data and several compatible with AWD practice since areas with consistently
26
defined environmental predictor constraints. Herein, the
27
MaxEnt machine learning model was developed using high rainfall (extreme rainfall) may not be suitable for this
28
the guidelines by Phillips and Dudik. We maintained practice. This is vital to determine when and where natural
rainfall can supplement or replace irrigation and optimize
the default setting of 5000 iterations and the evaluation water use. At the same time, percolation rates assess the
metric stabilization of 1000 background points. The choice 22
for the MaxEnt model was due to its high predictive suitability of different soil types for AWD practice.
accuracy under low sample size and the model’s user- 2.3. Estimation of percolation rates for paddy soils
friendly settings.
The percolation rates for paddy rice fields were estimated
We generated a 30 m × 30 m grid and resampled all by defining the higher limit of the percolation rate from
the predictor data to the exact resolution using the nearest soil texture classes, using data from the AfSoilGrids250m
22
neighbor and majority filter techniques (resampling, resolution depicting soil properties of Africa. Sensitivity
29
mosaicking, clipping, reprojection) for continuous and analysis was conducted to assess the uncertainties of
discrete datasets, respectively. The maps of the selected percolation rates, considering four major sensitivity
environmental predictors considered for mapping are classes, i.e., the lower bound of the potential percolation
provided as supplementary materials (Figure S1). Modeling rate, an upper bound of the possible percolation rate, a
was performed based on the occurrence data of the paddy basic setting corresponding to likely percolation rates
rice. The study applied the Kernel density estimate for the between the lower and upper bounds, and fixed values of
random generation of 1000 as background setting data potential percolation rates at a national scale for Uganda
with 60% and 40% for training and testing, respectively. (Table S3). The potential percolation rates ranged from 1
The training and testing ratio was set to provide a good to 15 mm/day (Table 2).
balance between training the model on sufficient data and
evaluating model performance – to reduce overfitting. By weighing the first three standard depths, we
Likewise, K-fold cross-validation was applied in which the aggregated data from the potential percolation (Pot ) up
pc
data (Table S2) was divided into k-folds, and the model to a depth of 30 cm. Table S4 describes the area covered
was trained on k-1 folds and validated on the remaining by each texture class at each depth over the study region,
fold and final performance metrics as average. Auto feature highlighting the domination of the clay-loam, clay and clay
was used to avoid overfitting by automatically limiting the sandy loam groups. These soil classes minimize percolation
complexity of the model depending on the amount of rates affecting the paddy rice irrigation schedule and
present data available. generally improve biomass production and nutrient uptake
(nitrogen use) by paddy rice. 30
2.2.5. Selection of threshold areas, model evaluation,
and validation 2.3.1. Climatic suitability assessment of AWD
The difference between the suitable and unsuitable areas irrigation practice
was identified based on the continuous probability surface We defined the climatic niche of AWD practice from
generated by the model predictions, where a threshold a simplified hydrological (water balance) model. The
22
value was used to maximize the division of the sum of conceptual illustration of water balance in the paddy field
the sensitivity and specificity by two. The sum maximizer is demonstrated in Figure S2.
threshold performed best across the varied modeling I + P + C =Potential ET + D + S + Pot (I)
settings. A threshold-based approach was used to cp pc
24
minimize the mean error rate for irrigated rice locations Where P , I, Potential ET, C, D, S, and Pot represent
pc
cp
and the error rate for background samples generated. precipitation, irrigated water supply, PET of a specific
Volume 2 Issue 2 (2025) 5 doi: 10.36922/EER025040005