Simulation of the Growth and Yield of Maize (Zea mays L.) on a Loosened Plinthosol Amended with Termite Mound Material in the Lubumbashi Region

1. Introduction

In the Democratic Republic of Congo (DRC), the agriculture is a key component of the economy, employing more than 60% of the Congolese population and contributing 17.4% of GDP [1]. However, agricultural production struggles to ensure food security and generate adequate incomes [2,3,4,5,6]. The main food crops include cassava, plantains, maize, groundnuts, and rice. Commercial agricultural production remains marginal, with the majority of farmers practicing subsistence agriculture [7,8,9,10,11]. In the South-eastern DRC (Haut-Katanga province), in order to alleviate food shortages, the population depends on imports from southern Africa, particularly to meet its needs for maize, a staple food for almost the entire population [12,13,14]. Among other factors, soil degradation remains a significant driver of low crop yield, mainly due to the dominance of highly weathered soils [15,16,17]. The degradation of soil fertility constitutes the major environmental challenge in this region, as demonstrated by scientific studies and confirmed by the perceptions of local communities [18].

These soils are characterized by low chemical fertility and the frequent presence of plinthite on the surface or at shallow depths [15], as well as the occurrence of large termite mounds [19,20]. The presence of plinthite on or near the surface often results in soils that are unsuitable for agriculture or shallow soils with limited trophic volume for crops [21,22]. This significantly reduces the capacity of these soils to support agricultural activities, rendering them unsuitable or marginal [23,24,25,26].

In the Lubumbashi region, a company has implemented an innovative strategy for managing plinthitosols to improve their productivity. This strategy is based on the use of materials from termite mounds built by Macrotermes falciger, which are abundant in the region, combined with subsoiling to break up the plinthite and enable intensive maize cultivation. The termite mounds are excavated, and their materials are then spread over subsoiled plots (Figure 1), thereby expanding the cultivated area, facilitating the use of agricultural machinery, and serving as a soil amendment [27,28,29,30,31,32,33,34,35,36]. This strategy raises a key research question: How does the improvement of plinthosol properties through subsoiling and amendment with termite mound material influence grain maize production?

Crop models are valuable tools for understanding, predicting, optimizing, and managing agricultural systems, thereby contributing to making agriculture more sustainable, productive, and resilient. Among them, APSIM (Agricultural Production System sIMulator) is widely recognized for its ability to model the complex interactions between components of agricultural systems after rigorous calibration and validation [37,38,39]. Operating on a modular and dynamic approach, APSIM allows for integrated simulation of soil-climate-plant-agricultural practice relationships [40,41,42,43,44,45]. This model has been used to evaluate the effect of different fertility management practices, including the application of organic and mineral amendments [46,47,48,49,50,51], as well as soil management practices such as subsoiling [52,53]. These applications illustrate APSIM’s ability to represent in an integrated manner the combined effects of agricultural practices on soil dynamics, crop growth, and crop productivity. The main objective of this study is to use the APSIM model to simulate the growth and yield of grain maize grown on a plinthosol in the Lubumbashi region, under the combined effect of subsoiling and amendment with termite mound materials.

2. Materials and Methods

2.1. Study Sites

This study was conducted in Lubumbashi, Haut-Katanga province, Democratic Republic of Congo, specifically at the FarmCo farm. The study site is located in the suburbs of Lubumbashi, approximately 60 km east of the city, along the Kasenga road in the Kifumanshi river valley (Figure 2). It covers an area of 660 hectares characterized by plinthosols, according to the FAO classification [54]. Systematic excavation of lateritic benches has been carried out, coupled with subsoiling of the entire perimeter. At the same time, materials from giant termite mounds inactive (Macrotermes falciger), which are abundant on the site, have been spread to level the ground and used as fertilizer in the crop plots.

This site is one of the largest farms in the Lubumbashi region, known for its large production of maize grain for human consumption in the form of flour.

The outskirts of Lubumbashi belongs to the Cwa climate (humid subtropical climate with hot summers and dry winters) according to the Köppen classification. The area is characterized by a dry season (May to September), a rainy season (November to March), and two transition months (April and October) [55]. Average annual rainfall is 1,270 mm. The average annual temperature is 20.1°C, with daily minimum and maximum temperatures reaching 8°C during the coldest month and 32°C during the hottest month, respectively [56].

2.2. APSIM Model Description

APSMI NextGen version 2024.10.7607.0 was used for scenario analysis in this study. This model was developed to simulate biophysical processes in agricultural systems in response to environmental variations [38]. A wide range of models are available in APSIM for major crop, pasture, and tree species, as well as key agricultural system processes at a daily time step [57]. The APSIM-Maize crop module was used to simulate the growth and yield of grain maize on plinthite soil over three growing seasons. The processes that affect growth and yield in APSIM-Maize are simulated by interactions between daily weather data, cultivar, soil properties, and different management practices [37,38].

The parameters calibrated to simulate maize growth and yield at the study site are detailed below and were integrated using a step-by-step approach. Phenological phase progression is expressed in thermal time, thus expressing values in degree days [58].

2.3. Parametrization and Calibration

2.3.1. Climate Data

The agro-meteorological data was obtained from NASA’s Prediction of Worldwide Energy Resource (POWER CERES/MERRA2) satellite database (https://power.larc.nasa.gov/data-access-viewer/) for the period from January 1, 2016, to December 31, 2024, in order to cover the study period. This source was used due to the lack of operational local weather stations in the study area. The data covers temperature, precipitation, relative humidity, wind speed, and solar radiation (Figure 3). The reliability of NASA data has already been proven by several studies [59,60,61,62].

2.3.2. Reference Evapotranspiration

The FAO Penman-Monteith approach [63] was used to calculate evapotranspiration from maize plants during the study period. This method, which requires the use of meteorological variables, is widely recognized for its ability to provide accurate and unambiguous estimates of ET0 in various environments around the world [64,65,66,67]. Figure 4 shows the results of reference evapotranspiration variation from 2016 to 2024 at the study site.

2.3.3. Maize Grain Yield

Maize grain yield was measured in subplots measuring one square meter each, clearly marked out in the different blocks (Figure 5). At least five measurements per block were taken during the first (2022-2023) and second (2023-2024) growing seasons, and the average yield per block was then calculated for each season. The yield for the 2016-2017 growing season was considered the control (B0), corresponding to the period prior to the implementation of subsoiling and termite mound material spreading. As a result, only blocks B1 to B10 benefited from these interventions.

In terms of crop management, all blocks, including the control block B0, benefited from the same technical itineraries, including mechanical tillage, harrowing, sowing, mineral fertilization, and the application of a pre-emergence herbicide. Maize was sown at the same time as the application of NPK (10-20-10) base fertilizer at a rate of 200 kg/ha, followed by a urea supplement applied 45 days after sowing at a rate of 200 kg ha^-1. The SC719 maize variety was grown at a spacing of 0.75 m between rows and 0.25 m within rows, giving a density of 53,333 plants per hectare. Harvesting took place when the plants reached harvest maturity (Table 1), i.e., when the seeds had a moisture content of 13%.

2.3.4. Soil Properties

A detailed description of a soil profile of ± 1 m in each bloc was carried out in each block following the FAO soil description guidelines [68]. Soil samples were taken in 2022 from the diagnostic horizons and wereair dried (for 5 days in the open air), sieved on a 2 mm mesh, packaged, and prepared in accordance with ISO 11074 for chemical analysis. Soil analyses carried out at the soil chemistry laboratory at Gembloux Agro-Bio Tech, Belgium include measurements of pH in water and KCl 1N, organic carbon determination by wet method (Walkey-Black method), total nitrogen was determined using the Kjeldahl method total and determination of available elements by extraction with ammonium + acetate EDTA (Ethylene Diamine Tetra Acetic Acid) at pH 4.65. The extracts obtained were measured using a colorimeter for phosphorus and ICP OES (ICP-Optical Emission Spectrometry) for available cations (calcium, magnesium, and potassium). Particle size distribution was determined using the pipette method, in accordance with the XFX 31-107 standard protocol. Undisturbed samples were collected from the horizons using Kopecky cylinder. These samples were then sent to the laboratory for bulk density measurement and study of soil water dynamics at the Soil Physics and Mechanics Technical Platform (PhyMeSol), Agro-Bio Tech, Belgium. Saturated hydraulic conductivity (Ksat) was measured using the constant head permeameter method and soil water retention was measured using Richard’s apparatus. For the control (B0), the results of soil analyses carried out in Belgium in 2016 were used. The same analytical protocols were applied, with the exception of hydraulic properties, which were estimated using the pedotransfer functions of Saxton & Rawls [69]. The results relating to water content, physicochemical properties (Table 2) and particle size distribution (Table 3) were incorporated into the APSIM model for the simulations.

2.3.5. Leaf Area Index

The Leaf Area Index (LAI) was determined for maize crops using Sentinel-2 satellite images, available via Google Earth Engine (GEE, https://code.earthengine.google.com). LAI is a key parameter for quantifying leaf area per unit of soil surface area and is crucial for assessing crop health and productivity. In the APSIM model, LAI directly controls the solar radiation intercepted by the canopy and, consequently, the ability to capture light for photosynthesis. Remote sensing is an effective tool for studying the seasonal dynamics of LAI, particularly in the absence of field measurements [70,71]. Sentinel-2 images are particularly well suited to this study due to their high spatial resolution (10 m) and extensive temporal coverage (5-day), allowing for accurate and regular monitoring of crops throughout their growth cycle [72]. Satellite data were filtered to cover the period from sowing to harvest for all study blocks. However, cloud-free satellite images are often too limited for the period between crop emergence and the V5/V6 stages to provide sufficient degrees of freedom for observations. As a result, a cloud cover criterion was applied, retaining only images with less than 20% cloud cover (Table 4).

The study area was delimited by a Shapefile representing the geographical boundaries of the area, thus enabling accurate geospatial analysis.

The satellite images were then preprocessed in GEE, with the main step being the calculation of the Simple Ratio (SR), a spectral index widely used to estimate LAI due to its high sensitivity to vegetation density [72,73]. SR was therefore used as an intermediate variable in estimating LAI, its use being justified by several studies that demonstrated good performance, with coefficients of determination (R²) greater than 0.70 and low errors, confirming a strong correlation between SR values and actual LAI measurements [74,75,76]. This method was chosen after comparison with other approaches, particularly those based on the NDVI.

The SR is derived from the NIR (B8) and red (B4) spectral bands of Sentinel-2 images, using the equation (SR; Equation (1)). In order to improve the accuracy of LAI estimates, supervised classification was performed using the SVM (Support Vector Machine) algorithm. This algorithm is known for its robustness in separating complex and heterogeneous spectral classes. The application of SVM made it possible to better discriminate between vegetated and non-vegetated areas, thereby reducing noise in the data and improving the reliability of the derived LAI values [77].

$$\mathit{S}\mathit{R}={\displaystyle \frac{\mathit{B}8}{\mathit{B}4}}$$

(1)

This calculation yields values ranging from 0 (bare soil) to higher values for more densely vegetated areas. The SR is particularly relevant in this study, as this ratio is sensitive to the leaf area present in the canopy, a key factor in LAI estimation.

LAI was then estimated from SR using an empirical relationship specific to maize (LAI; Equation (2)), where the factor 0.323 and the bias: -0.058 are empirical coefficients specific to maize cultivation and can sometimes be adjusted according to the region [78]. These coefficients have been specifically adjusted for maize cultivation based on agronomic research and are widely used to estimate LAI from SR in studies on this crop.

$$\mathit{L}\mathit{A}\mathit{I}=\mathrm{0,323}\mathbf{X}\mathit{R}\mathit{S}-0.058$$

(2)

LAI values were extracted for each image based on the previously defined crop blocks. For each image in the collection, the average LAI was calculated for each block using the 'reduceRegion' function in GEE. This operation made it possible to determine the average LAI for each crop area on each image capture date. These results enabled a visual and temporal assessment of crop health, tracking the progression of leaf development from sowing to harvest, which was important in calibrating the APSIM model. In addition, all LAI values for each block and each image were exported as a CSV file, allowing for detailed analysis and comparison with APSIM-simulated values and performance statistics.

2.3.6. Crop Parameters and LAI Calibration

Although the APSIM model includes default genotypes, mainly calibrated for certain regions of the world, most of these genotypes correspond to grain maize varieties that do not reflect the genotype used in this study. As a result, a new genotype was created in the model based on data from the maize variety technical data sheet provided by SEED-CO Zambia and field observations, in order to better reflect the specific phenology of the crop considered in the study. For LAI calibration, the values estimated on each Sentinel-2 image acquisition date were used to represent leaf growth dynamics throughout the crop cycle, using a cross-validation approach (2/3 of the data for calibration and 1/3 for validation), similar to that applied for grain maize yield (Table 5).

2.4. Cross-Validation

Cross-validation was performed by randomly partitioning all growing seasons, with two-thirds of the blocks used for model calibration and one-third used for validation. This approach ensures a strict separation between model training and evaluation, thus providing a reliable estimate of its generalization capacity. Commonly used in agro-environmental modeling, this method allows the robustness of the APSIM model to be evaluated by testing its ability to predict yields on blocks not used during calibration. The model was evaluated using the following statistical metrics:

-

Coefficient of determination (R²), which measures the proportion of variance explained by the model (R2; Equation (3)). A value close to 1 indicates good performance.

$${\mathit{R}}^2={\displaystyle \frac{\sum ({\mathit{Y}}_{\mathit{o}\mathit{b}\mathit{s}}-{\overline{\mathbf{Y}}}_{\mathit{o}\mathit{b}\mathit{s}})({\mathit{Y}}_{\mathit{s}\mathit{i}\mathit{m}}-{\overline{\mathbf{Y}}}_{\mathit{s}\mathit{i}\mathit{m}})}{\sqrt{\sum {({\mathit{Y}}_{\mathit{o}\mathit{b}\mathit{s}}-{\overline{\mathbf{Y}}}_{\mathit{o}\mathit{b}\mathit{s}})}^2}\sum {({\mathit{Y}}_{\mathit{s}\mathit{i}\mathit{m}}-{\overline{\mathbf{Y}}}_{\mathit{s}\mathit{i}\mathit{m}})}^2}}$$

(3)

-

RMSE (Root Mean Square Error), quantifying the average standard deviation between simulated and observed yields (RMSE; Equation (4)). The lower the RMSE, the more accurate the model.

$$\mathit{R}\mathit{M}\mathit{S}\mathit{E}=\sqrt{{\displaystyle \frac{1}{\mathit{n}}}\sum {({\mathit{Y}}_{\mathit{o}\mathit{b}\mathit{s}}-{\mathit{Y}}_{\mathit{s}\mathit{i}\mathit{m}})}^2}$$

(4)

-

NSE (Nash-Sutcliffe Efficiency), which evaluates the accuracy of the model by comparing it to the average of the observations (NSE; Equation (5)). When NSE is close to 1, the model performs well, and when it is less than 0, the model performs less well than the average of the measured yields.

$$\mathit{N}\mathit{S}\mathit{E}=1-{\displaystyle \frac{\sum {({\mathit{Y}}_{\mathit{o}\mathit{b}\mathit{s}}-{\mathit{Y}}_{\mathit{s}\mathit{i}\mathit{m}})}^2}{{({\mathit{Y}}_{\mathit{o}\mathit{b}\mathit{s}}-{\mathit{Y}}_{\mathit{o}\mathit{b}\mathit{s}})}^2}}$$

(5)

-

MAE (Mean Absolute Error) estimates the average of the absolute differences between simulated and observed values, providing an assessment that is less sensitive to extreme errors (MAE; Equation (6)).

$$\mathit{M}\mathit{A}\mathit{E}={\displaystyle \frac{1}{\mathit{n}}}\sum \left|{\mathit{Y}}_{\mathit{o}\mathit{b}\mathit{s}}-{\mathit{Y}}_{\mathit{s}\mathit{i}\mathit{m}}\right|$$

(6)

where Yobs represents the observed value, Y(sim) represents the simulated value, $\overline{\mathbf{Y}}$represents the mean of the values, and n is the number of observations.

2.5. Simulation Runs

All simulations were carried out using a common set of calibrated parameters. They covered the period from sowing to harvest (Table 1), for each block and season.

Figure 5 below shows the methodological diagram used for the construction and calibration of the APSIM model.

Figure 5. Flowchart of the APSIM model construction, calibration, and validation process.

Figure 5. Flowchart of the APSIM model construction, calibration, and validation process.

2.6. Data Analysis

All statistical analyses in this study were performed using R software, version 4.4.1.

3. Results

3.1. Performance of the APSIM Model for LAI Simulation

The in-depth evaluation of the APSIM model in cross-simulation of the leaf area index (Table 6) shows a good correspondence between the simulated and observed values, with R² between 0.85 and 0.87 in calibration and validation, reaching 0.86 across all blocks. This trend is confirmed by the proximity of the points along the 1:1 line (Figure 6), illustrating good overall consistency between observations and simulations. The Nash-Sutcliffe efficiency (NSE) varies from 0.61 to 0.71 in validation and calibration, respectively, indicating an overall acceptable performance of the model, with a slight expected decrease during validation (overall NSE = 0.67). The root mean square errors (RMSE) and absolute errors (MAE) remain moderate, with values between 0.25 and 0.35 m² m^-2, confirming the satisfactory accuracy of the model predictions.

The simulated leaf dynamics during the crop cycle show that the APSIM model is capable of accurately representing maize growth through the simulation of the leaf area index (Figure 7). The simulated values generally follow the trend of the observations across all blocks, with a peak corresponding to the phase of full leaf cover, followed by a gradual decline at the end of the maize crop cycle (senescence).

For the 2022-2023 growing season (red box), the linear adjustments between observed and simulated values show a strong correlation, confirming that the model is well suited to the field data. In 2023-2024 (blue box), the trends remain similar, although some blocks show more marked deviations, particularly block B8, which performs less well (NSE = 0.50; RMSE≥ 0.30 m² m^-2), suggesting more pronounced heterogeneity in plant structure.

For the 2016-2017 season, the model shows a satisfactory ability to reproduce LAI variations (NSE = 0.67), including under control conditions (B0) without amendment or subsoiling.

3.2. Evaluation of APSIM Model Performance in Cross-Validation of Grain Yield

Cross-validation of the APSIM model, performed using a 2/3 split for calibration and 1/3 for validation, shows robust performance and good generalization ability of the model for simulating maize yields (Table 7). Analysis of the statistical metrics reveals an overall coefficient of determination (R² = 0.91), indicating a strong correlation between observed and simulated yields. In addition, the Nash-Sutcliffe efficiency (NSE = 0.90) confirms this adequacy, highlighting the satisfactory accuracy of the simulations.

The calibration set (n = 12) and validation set (n = 9) show comparable performance, with R² values of 0.92 and 0.89, respectively, and NSE values of 0.99 and 0.88, indicating an excellent match between simulated and observed yields, with a slight decrease in validation, which is expected in a robust modeling process. Error analysis, using RMSE (0.47 t ha^-1) and MAE (0.45 t ha^-1), shows that the differences between predictions and measured values remain small, confirming the accuracy of the model. Furthermore, the small difference between RMSE and MAE suggests a uniform distribution of errors without significant outliers, ensuring reliable prediction.

These results demonstrate that the APSIM model is capable of accurately reproducing maize yields over different years and experimental blocks, while maintaining good stability between calibration and validation. Slightly more efficient on the training set, the model shows no signs of overfitting, which reinforces its applicability for larger-scale simulations.

The integration of all seasons, including the 2016-2017 control (Figure 8), reinforces the overall consistency of the model, with a very strong relationship between simulated and observed values (Y = 1.00X + 0.01; R² = 0.91; NSE = 0.90), reflecting low residual variability and good generalization capacity of the model across seasons and different agronomic conditions.

3.3. Performance of the APSIM Model in Predicting Yields

Analysis of simulated and observed yields in the different blocks indicates that the APSIM model is capable of accurately reproducing grain maize production trends (Table 8). The results show that for the 2022-2023 and 2023-2024 seasons, the simulated values generally remain close to the observations, with moderate differences between the two. The blocks that benefited from subsoiling and amendment with termite mound materials (B1-B10) show higher yields, reaching 10.4 t ha^-1 in 2023-2024 for B4, reflecting the positive effect of these interventions on productivity. Conversely, the control block (B0) in 2016-2017, without subsoiling or the addition of termite mound materials, showed a lower yield (4.1 t ha^-1 observed versus 4.4 t ha^-1 simulated), highlighting the limiting effect of the properties of unamended plinthosols on maize production.

Although the accuracy of the model is generally satisfactory, some differences are observed, particularly in B5 (2022-2023) and B9 (2023-2024), where the differences between observed and simulated yields exceed 0.5 t ha^-1, suggesting possible adjustments to the model based on the specific conditions of these blocks.

This section may be divided by subheadings. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn.

4. Discussion

4.1. Evaluation of the Accuracy of LAI Derived from Sentinel-2 and Its Simulation with APSIM

The use of Sentinel-2 images coupled with the Google Earth Engine platform to estimate leaf area index (LAI) proved effective in monitoring maize growth dynamics in different blocks. This remote sensing approach has provided LAI estimates consistent with field observations, confirming its relevance for agronomic [70,76,79,80,81]. The Simple Ratio (SR) algorithm, selected after comparison with other indices such as NDVI, has shown good performance in previous studies, with coefficients of determination greater than 0.70 and low errors [74,75,76]. White et al. [73], indicated that SR is particularly well suited for estimating LAI, as it is more sensitive to variations in leaf density and less prone to spectral signal saturation in dense stands. These results are consistent with the work of Cohrs et al. [72], who demonstrated the robustness of SR for monitoring plant biomass.

The LAI values calculated by SR for each date and each block, then integrated into APSIM, enabled the model to perform well. The coefficients of determination (R²) between 0.85 and 0.87 in calibration and validation (overall R² = 0.86) reflect a strong correlation between simulated and observed values. The Nash-Sutcliffe efficiency (NSE), ranging from 0.61 to 0.71 (global NSE = 0.67), confirms the model’s ability to reproduce observed trends, exceeding the threshold of 0.50 generally accepted as an indicator of acceptable performance [82]. The low RMSE (0.25-0.35 m² m^-2) and MAE (0.27 m² m^-2) values also confirm the accuracy of the model’s predictions. These results support the validity of integrating remote sensing-derived LAI into the calibration process and demonstrate the potential of APSIM to represent the physiological processes governing maize growth in challenging soil conditions.

The overall assessment shows that APSIM correctly reproduces the temporal dynamics of LAI, with a simulated evolution consistent with observations, namely a rapid growth phase leading to peak leaf area index, followed by a gradual decline towards senescence [37]. These results confirm the model’s ability to represent the physiological processes of maize development, as reported by Song et al. [83], Jin et al. [84], Chauhdary et al. [85], Santos et al. [86], Beah et al. [87], Dilla et al. [88]. The high coefficients of determination (R²) ranging from 0.85 to 0.93 and NSE values varying mainly between 0.50 and 0.77 indicate that APSIM effectively captures the spatio-temporal variations in LAI, in agreement with the conclusions of Keating et al. [57] on the robustness of the model. An NSE greater than 0.50 indicates acceptable simulation performance [82].

Analysis of the differences between simulated and observed LAI values reveals that the APSIM model performs satisfactorily overall, while highlighting certain variations depending on the blocks and growing seasons. In 2022-2023, linear adjustments between simulated and measured values show a strong correlation (NSE = 0.57-0.77), reflecting an accurate representation of leaf dynamics and confirming the model’s ability to reproduce maize growth under improved growing conditions, in line with the observations of Keating et al. [57] and Holzworth et al. [38]. In 2023-2024, the general trends remain similar, but more pronounced differences appear in certain blocks, particularly block B8 (NSE = 0.50; RMSE ≥ 0.30), suggesting more pronounced heterogeneity in plant structure, possibly linked to local variations in soil conditions in this block, which may affect water and nutrient uptake and plant cover density, as highlighted by Mountrakis et al. [77] and Prudnikova et al. [89]. Finally, the control block (B0) in 2016-2017 shows a satisfactory ability of the model to reproduce LAI variations (NSE = 0.67), despite the absence of amendment and subsoiling, confirming that APSIM can simulate maize growth even under more restrictive soil conditions, with limited water and nutrient availability [39,86,88], which reinforces its relevance for modeling crops on low-fertility soils.

4.2. Evaluation of the Performance of the APSIM Model for Simulating Maize Grain Yields

The cross-validation conducted confirms the robustness of the APSIM model for simulating grain maize yields. The high overall coefficient of determination (R² = 0.91) demonstrates its strong explanatory power, consistent with other studies that have validated APSIM for maize cultivation [47,83,88,90,91,92,93,94,95,96,97]. Furthermore, the overall NSE of 0.90 indicates the model’s ability to accurately reproduce observed yield trends, which is a key criterion for evaluating agro-climatic models [82].

Separate analysis of the calibration (n = 12) and validation (n = 9) phases shows similar performance, with an expected slight decrease in R² (0.92 to 0.89) and NSE (0.99 to 0.88) values in validation. Chauhdary et al. [98] observed R² and NSE values of 0.79 and 0.75 in calibration and 0.76 and 0.69 in validation, respectively. This phenomenon is expected in robust modeling processes, as a well-calibrated model should retain satisfactory predictive power when applied to new data [99,100]. The absence of a marked decrease in validation performance indicates that the model is not overfitted and retains good potential for generalization to other agroecological contexts [38]. The low RMSE (0.47 t ha^-1) and MAE (0.45 t ha^-1) values confirm the accuracy of the predictions, and their proximity reflects a homogeneous distribution of errors without systematic biase [37], which corroborates the results of [87,101,102,103] on the use of APSIM in low-fertility soil contexts in Africa.

The simulated yields are generally consistent with the observed values but reveal a marked contrast between the control block (B0) and the improved blocks (B1-B10). The blocks that benefited from subsoiling and amendment with termite mound material had the highest yields, reaching 10.4 t ha^-1 (B4 in 2023-2024), while block B0 only reached 4.1 t ha-1 observed (4.4 t ha^-1 simulated). This clear difference illustrates the limiting effect of the physicochemical properties of unimproved plinthosols on maize productivity, in particular their low porosity, high compactness, and limitation of root growth [104,105,106,107,108,109,110]. Conversely, improving soil structure and fertility through subsoiling and the addition of termite mound material has significantly increased water and nutrient availability, promoting deeper root development and better plant nutritional status [28,30,31,34,35,111,112].

These results confirm that subsoiling and soil amendment interventions played a decisive role in increasing soil productivity and improving the model’s fit with observations on blocks B1-B10.

However, some disparities remain, namely, the differences between simulated and observed values exceed 0.5 t ha^-1 for B5 (2022-2023) and B9 (2023-2024). This could reflect the omission of certain factors when parameterizing the model. More detailed optimization of the coefficients relating to soil properties and biotic factors could improve the accuracy of the simulations [113,114,115,116]. Nevertheless, Zhang et al. [117] point out that certain uncertainties may be linked to the structural limitations of the model itself.

5. Conclusions

This study demonstrated the ability of the APSIM model to simulate maize growth and yield on a subsoiled plinthosol enriched with termite mound (Macrotermes falciger) material in Lubumbashi. The approach combining remote sensing and modeling made it possible to obtain accurate estimates of the leaf area index, with high coefficients of determination and moderate to low simulation errors. A similar trend was observed for yield, where there was a strong correlation between simulated and observed values (R2 > 0.90; NSE > 0.80 and RMSE≤ 0.5 t ha^-1), confirming the robustness of the model in predicting grain maize yield.

Yield measurements further showed that soil improvement practices, particularly subsoiling and the application of termite mound materials, led to a significant increase in maize grain production. The APSIM model successfully reproduced these yield differences between blocks, thus demonstrating its ability to represent the effects of management practices when properly parameterized.

However, discrepancies between simulations and observations were noted in some blocks, highlighting the importance of finer model calibration. The integration of additional data and a more precise adjustment of maize physiological parameters would refine the model’s predictions, thereby enhancing its usefulness as a decision-making tool in sustainable agriculture.