## wpiea2025009-print-pdf

## Source details

**Canonical URL:** [wpiea2025009-print-pdf](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025009-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025009-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025009-print-pdf.pdf.json)

---

### 2.1 Data preparation — Remote sensing inputs and preprocessing
- Data sources and composites:
  - Sentinel-2 Level-2A Surface Reflectance imagery (DEA publicly available geomedian composites) was used.
  - Five geomedian composites were included: one annual geomedian composite, and four 3-month rolling geomedian composites (Jan-Mar, Apr-Jun, Jul-Sept, Oct-Dec) for the year 2022.
  - The merged dataset comprised five geomedian composites (four 3-month composites and one annual composite) totaling 91 bands.
- Geomedian and variability layers:
  - The geometric median (geomedian) synthesizes the most common spectral values at each pixel, reducing spatial noise and highlighting dominant spectral characteristics.
  - Three Median Absolute Deviation (MAD) layers were calculated for each geomedian composite: Euclidean MAD (EMAD), spectral MAD (SMAD), and Bray-Curtis MAD (BCMAD).
    - EMAD measures distance of each pixel from the median pixel in multi-dimensional spectral space.
    - SMAD captures spectral variability between neighboring pixels.
    - BCMAD quantifies compositional dissimilarity between pixels and their neighbors.
- Spectral indices and topography:
  - Six spectral indices were calculated for each geomedian composite and included in the classification: Normalized Difference Vegetation Index (NDVI); Leaf Area Index (LAI); Modified Normalized Difference Water Index (MNDWI); Tasseled Cap Wetness (TCW); Tasseled Cap Brightness (TCB); and TCG.
  - A slope file derived from the Shuttle Radar Topography Mission (SRTM) was included.

### 2.1 Platforms, workflows, and classifier design
- Platforms and workflow:
  - Image classification was conducted on the DEA Sandbox.
  - Training data collection, accuracy assessment, and area estimation were conducted on Google Earth Engine (GEE).
  - The machine learning workflow implemented a modified version of the Digital Earth Africa (DEA) continental cropland mask workflow.
- Land cover mapping definitions and rules:
  - Agriculture definition: a piece of land of minimum 0.16 ha that is sowed/planted and harvest-able at least once within the 12 months after the sowing/planting date. This definition excludes grasslands, unplanted pastures, and perennial/evergreen crops (e.g., mangoes, cashews, grapes, citrus trees).
  - The 0.16 hectares threshold corresponds to about four 20-meter x 20-meter pixels.
  - Random forest classifier was applied to the merged image features and the training data.
  - Masking rules applied to random forest output:
    - Steep slopes (>40°) masked using an SRTM derivative.
    - Water areas masked using the Water Observations from Space (WOfS) dataset.
    - Urban areas masked using the 2019 World Settlement Footprint (WSF-2019) dataset.
- Training data collection:
  - Two-phase approach: an initial province-by-province collection using Sentinel-2 and PlanetScope imagery, followed by iterative augmentation using preliminary classifier outputs to address misclassifications.

### 2.1 Accuracy assessment and area estimation methodology
- Sampling and reference labeling:
  - A simple random sample of 1,200 pixels at 20m resolution was selected for map accuracy assessment.
  - Total sample size was determined using the variance estimator solved for n as described in Cochran (1977), with a target margin of error of 10% of the Agriculture class.
  - Trained interpreters assigned reference land use labels (Agriculture or Other) to each sample unit by examining time-series data of Sentinel-2 and PlanetScope spanning 2019-2022, using the AREA2 toolbox.
  - Each sample labeled with a corresponding confidence level in one of three levels; samples with the lowest confidence values were re-examined and either adjusted or removed (93 reference points removed later; see Section 4.2).
- Estimation:
  - A stratified estimator was applied to the interpreted reference samples to estimate area at 95% confidence intervals (Olofsson et al., 2014).
  - Rationale: simple random sampling was used because a probability sampling design was needed and reference data collection began before the final map was completed, making stratified sampling infeasible.

### 2.1 Training data summary
- Nationwide training dataset size: 34,604 samples (each the size of a 20m Sentinel-2 pixel).
- Training samples visually interpreted and categorized for the year 2022 into two classes: Agriculture and Other.
- Class distribution in training data: approximately 70% labeled as Other and 30% labeled as Agriculture.
- Training site distribution: relatively uniform across the country, with slightly lower densities in larger provinces or areas dominated by uniform non-agricultural land cover.

---

### 3.2 Mozambique 2022 Active Agriculture Mapping Results — Mapping results and spatial patterns
- Spatial resolution and definition:
  - Active agriculture mapped at 20m spatial resolution for 2022.
  - Active agriculture defined as a piece of land of minimum 0.16 ha that is sowed/planted and harvest-able at least once within the 12 months after the sowing/planting date; this definition excludes grasslands, unplanted pastures, and perennial/evergreen crops.
- Spatial variability highlights:
  - Subset 3A and Subset 3B: densely populated smallholder farms in the provinces of Nampula and Zambezia.
  - Subset 3C: area with lower field density and slightly larger average field sizes (province of Manica).
  - Most agricultural areas found close to urban areas, often forming clusters around settlements.

### 3.2 Key statistics on agricultural extent (2022)
- National and provincial extents:
  - Agriculture accounted for 12% of Mozambique’s total area, covering 90,680.57 km² ± 8,127.49 km².
  - Provincial concentration of agricultural land (roughly 60% concentrated in three provinces):
    - Nampula: 25%
    - Zambezia: 20%
    - Tete: 15%
  - Corresponding provincial population concentrations (2022 projections from the 2017 national census (IV RGPH, 2017)):
    - Nampula: 20%
    - Zambezia: 18%
    - Tete: 10%

### 3.2 Comparison with global products (summary)
- ESRI/Impact Observatory Land Cover product for 2022 significantly underestimates agricultural areas compared to the 2022 map from this study.
- ESA WorldCover map for 2021 aligns more closely but consistently underestimates smallholder agriculture, resulting in a lower overall agricultural land estimate.

### 3.3 Accuracy Assessment and Area Estimation — Results
- Reference sampling and final sample:
  - Simple random sample of 1,200 reference observations; final comparison used 1,107 reference observations after quality assurance.
- Error matrix summary (reference totals and map vs. reference counts):
  - Reference totals: Agriculture = 114, Other = 993, Total = 1,107.
  - Map Agriculture / Reference Agriculture: 92
  - Map Agriculture / Reference Other: 11
  - Map Other / Reference Agriculture: 22
  - Map Other / Reference Other: 982
- Accuracy metrics:
  - User Accuracy:
    - Agriculture: 89.32%
    - Other: 97.81%
  - Producer Accuracy:
    - Agriculture: 83.02%
    - Other: 98.71%
  - Area Estimate [km²]:
    - Agriculture: 90,680.57
    - Other: 696,309.07
  - Area 95% Confidence Interval [km²]:
    - Agriculture: 8,127.49
    - Other: 8,127.49
  - Margin of Error [%]:
    - Agriculture: 8.96%
    - Other: 1.17%
  - Overall Accuracy (Table 1): 96.90%
  - Additional reported overall map accuracy in prose: 96.09%
- Interpretation:
  - All map classes exceed the goal of accuracies above 70% and meet the objective of staying below a 10% margin of error.
  - Slightly lower values for the Agriculture class indicate challenges in accurately mapping cropland in this region (see Section 4.2).

---

### 4. Discussion — Agricultural extent validation, challenges, and applications

H3: 4.1 Agricultural extent in Mozambique: validation and comparison
- National-scale estimate and comparisons:
  - Mozambique's agricultural land estimated at approximately 9.1 million hectares (first national-level estimate based on rigorous accuracy assessment and area estimation methods).
  - Mozambique Ministry of Agriculture’s 2023 census: small and medium farmers utilize approximately 6.9 million hectares, representing about 8.7% of the country’s total area. Differences attributed to:
    - Ministry’s definition of agricultural land excludes large farming enterprises.
    - Ministry methodology relies on sampling through field surveys (limited geographic coverage).
  - FAO estimate: 5.6 million hectares currently cultivated, representing about 7.1% of the country’s total area.
- Limitations of Ministry and FAO estimates:
  - Neither provides accuracy assessments or error terms.
  - Lack spatially explicit information at sufficient granularity for detailed land management.
- Cross-country context (excerpted FAO figures):
  - United States — Country area (in 1,000 ha): 983,151; Agriculture area (in 1,000 ha): 427,475; Share of agri. land use: 43%
  - China — 960,001; 521,395; 54%
  - Brazil — 851,577; 228,489; 27%
  - Australia — 774,122; 377,002; 49%
  - India — 328,726; 178,528; 54%
  - Argentina — 278,040; 117,679; 42%
  - South Africa — 121,909; 96,341; 79%
  - France — 54,909; 28,524; 52%
- Conclusion:
  - Despite methodological differences, all estimates indicate a relatively low agricultural land usage rate in Mozambique compared to other agriculturally significant countries, suggesting potential for growth and increased productivity within the existing land base.

H3: 4.2 Land cover mapping challenges
- Primary technical challenges identified:
  - Sub-pixel mixtures of land cover, especially in regions with sparse or low-stature vegetation (grasslands, mixed open forests).
  - Agriculture omission errors primarily observed along borders of fields (transition zones) and within fields with dispersed non-plantation tree cover.
  - Commission errors predominantly observed in grassland areas; distinguishing grasslands from rain-fed agriculture is challenging because spectral characteristics are relatively similar throughout the year.
  - Mozambique's agricultural sector is ~70% rain-fed (Silva & Matyas, 2014).
  - Removal of 93 reference points from the final reference dataset due to systematically lower confidence in reference labels.
- Mitigation and remaining opportunities:
  - Inclusion of DEA 3-month rolling geomedian composites in final classification helped mitigate commission errors by incorporating multiple seasons.
  - Absence of very high spatial resolution imagery (<1m) such as Google Earth or drone imagery limited the ability to distinguish active agriculture along transition boundaries at pixel scale.
  - Further integration of other high-resolution datasets may improve precision.

H3: 4.3 Leveraging remote sensing for monitoring and infrastructure insights
- Value of remote sensing:
  - Enables large-scale, timely monitoring of agricultural systems and environmental/infrastructure factors (e.g., flooding, prolonged droughts).
  - Automates data processing and reduces reliance on manual interpretation to improve accuracy, reproducibility, and reliability.
- Recommendations for future research and applications:
  - Adopt dynamic time series models (e.g., Continuous Change Detection and Classification (CCDC)) to capture spatiotemporal dynamics and continuous monitoring of agricultural trends.
  - Use the dataset from this study to support time-series change assessments for detecting land abandonment, crop rotation, and climate-related disruptions.
  - Derive proxy indicators from satellite data to identify likely irrigated areas.
  - Leverage high-resolution imagery (e.g., PlanetScope data) to assess rural road networks and other infrastructure critical for market connectivity and agricultural productivity.

---

### 5. Agricultural development — A historic perspective: sources, impacts, and lessons for Mozambique
- Sources of agricultural development:
  - Technological and chemical innovations:
    - Mechanization: improvements in plowing/tilling, planting, irrigation (sprinkler and drip), combined and specialized harvesters, remote monitoring via drones and satellites, precision agriculture using GPS/GIS, sorting/grading machines, mechanized storage, renewable-energy powered machinery.
    - Chemical inputs and biotechnology: synthetic fertilizers (Haber-Bosch), Enhanced Efficiency Fertilizers, Integrated Pest Management, biological/botanical pesticides, nanotechnology formulations, genetically modified crops for insect resistance and herbicide tolerance.
  - Economic and policy factors:
    - Agricultural policies (subsidies, direct payments, tariffs, trade agreements, environmental regulations, land use policies) and R&D underpin productivity and resilience.
  - Social and demographic changes:
    - Population growth and urbanization altered rural labor supply and demand for processed foods.
  - Market access, energy, and input costs:
    - Transportation, electrification, globalization, and technology transfer affected market reach, competition, and exposure to price volatility.
- Green Revolution case studies and impacts:
  - Mexico: semi-dwarf, disease-resistant wheat varieties led to self-sufficiency by early 1960s and export capacity.
  - India: Green Revolution from late 1960s produced food self-sufficiency in many staples within roughly a decade; diffusion expanded beyond wheat.
  - Brazil: multi-decade transformation via EMBRAPA, cerrado development, no-till, GMOs, precision agriculture, subsidies, credit, and tax incentives; environmental and social concerns noted.
- Linkages to economic development and poverty alleviation:
  - Agricultural productivity increases raise rural incomes, improve food security, reduce poverty, and stimulate non-agricultural demand.
  - Structural transformation releases labor from agriculture, supporting broader economic diversification and sustained growth.
  - Empirical references indicate agriculture-driven growth has larger poverty-reducing impacts and higher poverty-growth elasticity for Mozambique (figures referenced).
- Lessons and constraints for Mozambique:
  - Mozambique has opportunity to expand underutilized agricultural area and raise productivity.
  - Mozambique has imported about double the amount of food in terms of value compared to what it has exported, in contrast to South Africa which has exported about 70 percent more than it paid for the import of food products (figure referenced).
  - Agriculture is the main source of income for more than 70 percent of the population, provides employment for about 80 percent of the workforce, and contributes to over a quarter of the country’s GDP.
  - Impediments (consensus view):
    - Key Infrastructure Gaps: rural roads, electricity, irrigation, and ICT deficits.
    - Underdeveloped Input Markets: limited supply, weak demand, high prices, low quality of inputs.
    - Poor Organization and Functioning of Output Markets and Value Addition.
    - Lack of Access to Finance: limited credit, lack of agricultural insurance, exposure to weather risks.
  - Stylized facts:
    - Large gap in fertilizer usage between Mozambique and regional peers.
    - Positive relationship between farmer wealth and propensity to use productive inputs and tools.
  - Constraints for large-scale agribusiness overlap with smallholder constraints (infrastructure, red tape, governance, coordination).

---

### 6. Conclusion — Key findings and policy priorities
- Key findings:
  - Over the past few decades, Mozambique has not seen fast agricultural development.
  - Household surveys indicate land usage by smallholder farmers has not increased substantially; agricultural land coverage is contained to a minimal fraction of the overall available land.
  - Productivity of Mozambican farmers is, on average, very low compared to the wider region and has not improved markedly over the recent past.
  - Over 70 percent of the workforce depends on agriculture to sustain lives and livelihoods.
  - Outside options in the labor market are broadly non-accessible to the rural population due to lack of education and limited urban demand for rural, low-skilled workers.
  - The impact on poverty of growth in the agricultural sector is higher than in any other sector of the economy.
- Implications and priorities:
  - Improving agricultural output should be a front-and-center policy objective of any Mozambican government.
  - There are many impediments to the sector and the government should try and address all of them, but an early prioritization and impact assessment is recommendable.

*Source: wpiea2025009-print-pdf — selected content from IMF Working Paper.*

### 2.1 Data preparation ...................................................................................................

### 2.1 Data preparation

### Remote sensing data and preprocessing
- Sentinel-2 Level-2A Surface Reflectance imagery (DEA publicly available geomedian composites) was used.
- Five geomedian composites were included: one annual geomedian composite, and four 3-month rolling geomedian composites (Jan-Mar, Apr-Jun, Jul-Sept, Oct-Dec) for the year 2022.
- The geometric median (geomedian) synthesizes the most common spectral values at each pixel, reducing spatial noise and highlighting dominant spectral characteristics.
- Three Median Absolute Deviation (MAD) layers were calculated for each geomedian composite: Euclidean MAD (EMAD), spectral MAD (SMAD), and Bray-Curtis MAD (BCMAD).
  - EMAD measures distance of each pixel from the median pixel in multi-dimensional spectral space.
  - SMAD captures spectral variability between neighboring pixels.
  - BCMAD quantifies compositional dissimilarity between pixels and their neighbors.
- Six spectral indices were calculated for each geomedian composite and included in the classification: Normalized Difference Vegetation Index (NDVI); Leaf Area Index (LAI); Modified Normalized Difference Water Index (MNDWI); Tasseled Cap Wetness (TCW); Tasseled Cap Brightness (TCB); and TCG.
- A slope file derived from the Shuttle Radar Topography Mission (SRTM) was included.
- The merged dataset comprised five geomedian composites (four 3-month composites and one annual composite) totaling 91 bands.

### Platforms and workflows
- Image classification was conducted on the DEA Sandbox.
- Training data collection, accuracy assessment, and area estimation were conducted on Google Earth Engine (GEE).
- The machine learning workflow implemented a modified version of the Digital Earth Africa (DEA) continental cropland mask workflow.

### Land cover mapping: definitions and classifier
- Agriculture was defined as: a piece of land of minimum 0.16 ha that is sowed/planted and harvest-able at least once within the 12 months after the sowing/planting date. This definition excludes grasslands, unplanted pastures, and perennial/evergreen crops (e.g., mangoes, cashews, grapes, citrus trees).
- The 0.16 hectares threshold corresponds to about four 20-meter x 20-meter pixels.
- Random forest classifier was applied to the merged image features and the training data.
- Masking rules applied to random forest output:
  - Steep slopes (>40°) masked using an SRTM derivative.
  - Water areas masked using the Water Observations from Space (WOfS) dataset.
  - Urban areas masked using the 2019 World Settlement Footprint (WSF-2019) dataset.
- Training data collection proceeded in two phases: an initial province-by-province collection using Sentinel-2 and PlanetScope imagery, followed by iterative augmentation using preliminary classifier outputs to address misclassifications.

### Accuracy assessment and area estimation methodology
- A simple random sample of 1,200 pixels at 20m resolution was selected for map accuracy assessment.
- Total sample size was determined using the variance estimator solved for n as described in Cochran (1977), with a target margin of error of 10% of the Agriculture class.
- Reference labeling:
  - Trained interpreters assigned reference land use labels (Agriculture or Other) to each sample unit by examining time-series data of Sentinel-2 and PlanetScope spanning 2019-2022, using the AREA2 toolbox.
  - Each sample labeled with a corresponding confidence level in one of three levels.
  - Samples with the lowest confidence values were re-examined and either adjusted or removed.
- A stratified estimator was applied to the interpreted reference samples to estimate area at 95% confidence intervals (Olofsson et al., 2014).
- Rationale for sampling design:
  - Simple random sampling selected because a probability sampling design was needed and reference data collection began before the final map was completed, making stratified sampling infeasible.

### Results summary (training data)
- Nationwide training dataset size: 34,604 samples (each the size of a 20m Sentinel-2 pixel).
- Training samples visually interpreted and categorized for the year 2022 into two classes: Agriculture and Other.
- Class distribution in training data: approximately 70% labeled as Other and 30% labeled as Agriculture.
- Training site distribution: relatively uniform across the country, with slightly lower densities in larger provinces or areas dominated by uniform non-agricultural land cover.

*Source: wpiea2025009-print-pdf — 2.1 Data preparation (IMF Working Paper).*

### 3.2 Mozambique 2022 Active Agriculture Mapping Results

### 3.2 Mozambique 2022 Active Agriculture Mapping Results

### Mapping results and spatial patterns
- Active agriculture mapped at 20m spatial resolution for 2022 (Figure 3.2).
- Active agriculture defined as a piece of land of minimum 0.16 ha that is sowed/planted and harvest-able at least once within the 12 months after the sowing/planting date. This definition excludes grasslands, unplanted pastures, and perennial/evergreen crops.
- Spatial variability:
  - Subset 3A and Subset 3B: densely populated smallholder farms in the provinces of Nampula and Zambezia.
  - Subset 3C: area with lower field density and slightly larger average field sizes (province of Manica).
  - Most agricultural areas found close to urban areas, often forming clusters around settlements.

### Key statistics on agricultural extent (2022)
- Agriculture accounted for 12% of Mozambique’s total area, covering 90,680.57 km² ± 8,127.49 km² (Table 1).
- Provincial concentration of agricultural land (roughly 60% concentrated in three provinces):
  - Nampula: 25%
  - Zambezia: 20%
  - Tete: 15%
- Corresponding provincial population concentrations (2022 projections from the 2017 national census (IV RGPH, 2017)):
  - Nampula: 20%
  - Zambezia: 18%
  - Tete: 10%

### Comparison with global products (summary)
- ESRI/Impact Observatory Land Cover product for 2022 significantly underestimates agricultural areas compared to the 2022 map from this study.
- ESA WorldCover map for 2021 aligns more closely but consistently underestimates smallholder agriculture, resulting in a lower overall agricultural land estimate.

---

### Accuracy Assessment and Area Estimation (Section 3.3)
- Reference sampling:
  - Simple random sample of 1,200 reference observations; final comparison used 1,107 reference observations after quality assurance (Figure 3.4).
- Error matrix summary (Table 1):
  - Reference totals: Agriculture = 114, Other = 993, Total = 1,107.
  - Map vs. reference counts:
    - Map Agriculture / Reference Agriculture: 92
    - Map Agriculture / Reference Other: 11
    - Map Other / Reference Agriculture: 22
    - Map Other / Reference Other: 982
  - User Accuracy:
    - Agriculture: 89.32%
    - Other: 97.81%
  - Producer Accuracy:
    - Agriculture: 83.02%
    - Other: 98.71%
  - Area Estimate [km²]:
    - Agriculture: 90,680.57
    - Other: 696,309.07
  - Area 95% Confidence Interval [km²]:
    - Agriculture: 8,127.49
    - Other: 8,127.49
  - Margin of Error [%]:
    - Agriculture: 8.96%
    - Other: 1.17%
  - Overall Accuracy (Table 1): 96.90%
- Additional reported overall map accuracy in prose: 96.09%.
- Interpretation:
  - All map classes exceed the goal of accuracies above 70% and meet the objective of staying below a 10% margin of error.
  - Slightly lower values for the Agriculture class indicate challenges in accurately mapping cropland in this region (see Section 4.2).

---

### 4. Discussion

### 4.1 Agricultural extent in Mozambique: validation and comparison
- National-scale estimate from remote sensing:
  - Mozambique's agricultural land estimated at approximately 9.1 million hectares (first national-level estimate based on rigorous accuracy assessment and area estimation methods).
- Comparison with national and international estimates:
  - Mozambique Ministry of Agriculture’s 2023 census: small and medium farmers utilize approximately 6.9 million hectares, representing about 8.7% of the country’s total area. Differences attributed to:
    - Ministry’s definition of agricultural land excludes large farming enterprises.
    - Ministry methodology relies on sampling through field surveys (limited geographic coverage).
  - FAO estimate: 5.6 million hectares currently cultivated, representing about 7.1% of the country’s total area. FAO relies on annual questionnaires and self-reported data.
- Limitations of Ministry and FAO estimates:
  - Neither provides accuracy assessments or error terms.
  - Lack spatially explicit information at sufficient granularity for detailed land management.
- Cross-country context (Table 2 excerpted figures; Source: FAO):
  - United States — Country area (in 1,000 ha): 983,151; Agriculture area (in 1,000 ha): 427,475; Share of agri. land use: 43%
  - China — 960,001; 521,395; 54%
  - Brazil — 851,577; 228,489; 27%
  - Australia — 774,122; 377,002; 49%
  - India — 328,726; 178,528; 54%
  - Argentina — 278,040; 117,679; 42%
  - South Africa — 121,909; 96,341; 79%
  - France — 54,909; 28,524; 52%
- Conclusion: despite methodological differences, all estimates indicate a relatively low agricultural land usage rate in Mozambique compared to other agriculturally significant countries, suggesting potential for growth and increased productivity within the existing land base.

### 4.2 Land cover mapping challenges
- Primary technical challenges:
  - Sub-pixel mixtures of land cover, especially in regions with sparse or low-stature vegetation (grasslands, mixed open forests).
  - Agriculture omission errors primarily observed:
    - Along borders of fields (transition zones).
    - Within fields with dispersed non-plantation tree cover.
  - Removal of 93 reference points from the final reference dataset due to systematically lower confidence in reference labels.
- Commission errors:
  - Predominantly observed in grassland areas; distinguishing grasslands from rain-fed agriculture is challenging because spectral characteristics are relatively similar throughout the year.
  - Mozambique's agricultural sector is ~70% rain-fed (Silva & Matyas, 2014).
- Mitigation steps and remaining opportunities:
  - Inclusion of DEA 3-month rolling geomedian composites in final classification helped mitigate commission errors by incorporating multiple seasons.
  - Absence of very high spatial resolution imagery (<1m) such as Google Earth or drone imagery limited the ability to distinguish active agriculture along transition boundaries at pixel scale.
  - Further integration of other high-resolution datasets may improve precision.

### 4.3 Leveraging remote sensing for agricultural monitoring and infrastructure insights
- Value of remote sensing:
  - Enables large-scale, timely monitoring of agricultural systems and environmental/infrastructure factors (e.g., flooding, prolonged droughts).
  - Automates data processing and reduces reliance on manual interpretation to improve accuracy, reproducibility, and reliability.
- Recommendations for future research and applications:
  - Adopt dynamic time series models (e.g., Continuous Change Detection and Classification (CCDC)) to capture spatiotemporal dynamics and continuous monitoring of agricultural trends (see Zhu & Woodcock, 2014; Arevalo et al., 2020).
  - Use the dataset from this study to support time-series change assessments for detecting land abandonment, crop rotation, and climate-related disruptions.
  - Derive proxy indicators from satellite data to identify likely irrigated areas (Xie & Lark, 2021).
  - Leverage high-resolution imagery (e.g., PlanetScope data) to assess rural road networks and other infrastructure critical for market connectivity and agricultural productivity.

*Source: IMF Working Paper — 3.2 Mozambique 2022 Active Agriculture Mapping Results (wpiea2025009-print-pdf: selected content).*

### 5. Agricultural development - a historic perspective

### 5. Agricultural development - a historic perspective

### The sources of agricultural development
- Technological and chemical innovations have been central drivers of rising agricultural output, including:
  - Mechanization:
    - Improvements in plowing and tilling, planting, and irrigation (sprinkler and drip technology).
    - Combined harvesters and specialized harvesters (e.g. potato diggers, cotton pickers).
    - Remote monitoring via drones and satellites; precision agriculture using GPS and GIS.
    - Sorting and grading machines reduce post-harvest loss; mechanized storage and preservation extend shelf life.
    - Better transportation and logistics; renewable-energy powered machinery improve energy efficiency.
  - Chemical inputs and biotechnology:
    - Synthetic fertilizers (Haber-Bosch process) increased nitrogen availability and crop yields.
    - Recent improvements: Enhanced Efficiency Fertilizers; Integrated Pest Management; biological and botanical pesticides; nanotechnology for nano-formulations of fertilizers and pesticides; soil improvement products; water-soluble fertilizers.
    - Genetically modified crops foster insect resistance and herbicide tolerance (Qaim, 2009).
- Economic and policy factors:
  - Agricultural policies: subsidies, direct payments, preferential tariffs, trade agreements, environmental regulations, and land use policies shape incentives and market outcomes.
  - Research and Development (R&D): public and private investment in high-yield varieties, precision technologies, biofortified foods, and climate-resilient crops underpin productivity and resilience (Pingali, 2012; Lobell et al., 2011; Fuglie et al., 2011).
- Social and demographic changes:
  - Population growth and urbanization: from about one billion in 1800 to more than 8 billion today; urban migration reduced rural labor supply and increased demand for processed foods and longer shelf life products.
- Market access, energy, and input costs:
  - Transportation networks (railroads, highways) and electrification expanded market reach, competition, diffusion of know-how, and use of advanced machinery.
  - Globalization expanded market access, specialization by comparative advantage, and technology transfer, while introducing price volatility and exposure to exchange rate and trade policy shifts.

### Green Revolution – the fast catching up of some countries
- Broad impact:
  - The Green Revolution markedly influenced hunger reduction and agricultural advancement, particularly in Asia, and contributed to several countries' transitions to middle-income status and beyond.
  - Despite achievements, high levels of malnutrition persist in many regions; substantial rural populations in Sub-Saharan Africa and South Asia remain reliant on low productivity agriculture.
- Case studies:
  - Mexico:
    - Norman Borlaug and the Rockefeller Foundation developed semi-dwarf, disease-resistant wheat varieties responsive to fertilizers and irrigation.
    - By the early 1960s, Mexico achieved wheat self-sufficiency and became an exporter; environmental and biodiversity concerns followed.
  - India:
    - Green Revolution began in the second half of the 1960s; within roughly a decade India achieved food self-sufficiency in many staples.
    - Initial gains concentrated in northern regions and wheat; later diffusion covered most crops including rice.
    - Proliferation of private tube-wells (groundwater irrigation) aided rapid technology adoption.
  - Brazil:
    - Over five decades, Brazil moved from net importer to leading producer/exporter through EMBRAPA (established 1973), cerrado development, no-till farming, GMOs, precision agriculture, subsidies, credit, and tax incentives.
    - Exports include soybeans, beef, sugar, coffee, and orange juice; environmental concerns include deforestation in the Amazon and cerrado, plus social issues around land rights and indigenous displacement.

### Linkage between agricultural progress, broader economic development and poverty alleviation
- Agriculture’s roles:
  - In low-income countries, agricultural development is pivotal for food production, self-sufficiency, and as a catalyst for economic growth and poverty reduction.
  - Agricultural productivity increases raise rural incomes, improve food security, and reduce poverty; they also provide affordable food to urban areas and create demand for non-agricultural goods and services.
  - Adoption of high-yielding varieties and improved practices creates rural employment and supports income diversification through the rural non-farm sector.
  - Structural transformation: releasing labor from agriculture to other sectors contributes to broader economic diversification and sustained growth.
- Empirical insights referenced:
  - Figure 5.1 and Figure 5.2 (in the source) illustrate larger poverty-reducing impact from agricultural growth compared to other sectors, and that poverty-growth elasticity is larger for Mozambique than in other countries in the region (source: Dorosh and Thurlow, 2018).

### Lessons for Mozambique and the macro criticality of agriculture
- Current status and opportunities:
  - Mozambique has an opportunity to expand underutilized agricultural area and raise low productivity levels.
  - Mozambique has imported about double the amount of food in terms of value compared to what it has exported, in contrast to South Africa which has exported about 70 percent more than it paid for the import of food products (Figure 5.3).
  - Agriculture is the main source of income for more than 70 percent of the population, provides employment for about 80 percent of the workforce, and contributes to over a quarter of the country’s GDP.
  - The World Bank’s Mozambique Rural Income Diagnostic (2020) documents the rural population’s dependence on small plots and limited non-farm labor opportunities; productivity improvements in agriculture are key channels to target poverty.
  - Large yield gaps exist in cereal crops and maize between Mozambique and neighboring countries (Figures 5.4 and 5.5 referenced).
- Impediments to agricultural development (consensus view; World Bank 2020, 2022):
  - Key Infrastructure Gaps: rural roads, electricity (figure 5.6), irrigation, and ICT deficits negatively impact farm gate prices, input prices, and market access.
  - Underdeveloped Input Markets: limited supply, weak demand, high prices, and low quality of inputs impede smallholders' use of improved inputs and weaken commercialization incentives.
  - Poor Organization and Functioning of Output Markets and Value Addition: weak commercialization, high marketing and transaction costs, low farm gate prices, and limited incentives for productive investments.
  - Lack of Access to Finance: limited access to credit, lack of agricultural insurance markets, and exposure to weather risks hinder smallholders' investment capacity.
- Stylized facts and linkages:
  - Large gap in fertilizer usage between Mozambique and regional peers (Figure 5.7).
  - A positive relationship exists between farmer wealth and propensity to use productive inputs and tools (Figure 5.8).
- Constraints for large-scale agribusiness:
  - Many impediments affecting smallholders (infrastructure gaps, underdeveloped input markets) also constrain large-scale ventures.
  - A Growth Diagnostic study by the London School of Economics identifies three main bottlenecks for Mozambique’s growth: (i) infrastructure, (ii) red tape and governance problems, and (iii) lack of accountability and coordination between government agencies.

*Source: IMF Working Paper — Chapter 5, "Agricultural development - a historic perspective."*

### 6. Conclusion

### 6. Conclusion

### Key findings on agricultural development
- Over the past few decades, Mozambique has not seen fast agricultural development.
- Based on conducted household surveys, the usage of land by smallholder farmers has not increased substantially, and this study shows that the agricultural land coverage is contained to a minimal fraction of the overall available land.
- Productivity of Mozambican farmers is, on average, very low compared to the wider region and it has not improved markedly over the recent past.
- Over 70 percent of the workforce depends on agriculture to sustain lives and livelihoods.
- The outside options in the labor market are broadly non-accessible to the rural population due to a lack of education and a labor market in the urban centers that does not have enough demand for rural, low-skilled workers.
- The impact on poverty of growth in the agricultural sector is higher than in any other sector of the economy.

### Implications and priorities
- Improving agricultural output should be a front-and-center policy objective of any Mozambican government.
- There are many impediments to the sector and the government should try and address all of them, but an early prioritization and impact assessment is recommendable.

*Source: wpiea2025009-print-pdf - 6. Conclusion*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025009-print-pdf.pdf_