Editor's Note: In our first blog, we introduced the M-RED program’s use of sugarcane as a low-cost intervention to reduce erosion in flood-prone areas of western Nepal. This follow-up post dives into the results of a geospatial impact evaluation (GIE), exploring how Earth observation and causal inference methods were used to understand what changed on the ground and why.
Read the first blog here: Rooting Resilience: What Satellites Reveal About Sugarcane Farming in Nepal
Close-up of sugarcane in western Nepal, where cultivation expanded rapidly through the M-RED program.
The study focused on 53 communities across the Kailali and Kanchanpur districts, situated along the Mohana and Mahakali River systems. These areas were selected for their high flood risk and previous involvement in Mercy Corps programs.
Rather than relying only on field visits, the team applied a geospatial impact evaluation approach. Using Sentinel-2, PlanetScope, and Google Earth imagery, they mapped sugarcane fields and riverbanks from 2016 to 2023. This provided consistent, low-cost insights into how land and rivers changed over time.
The evaluation assessed three outcomes: sugarcane adoption, soil erosion, and riverbank movement. Depending on complexity and imagery quality, the team used a mix of machine learning and expert visual interpretation.
The first hurdle was distinguishing sugarcane from other crops. The team used Sentinel-2 imagery to track plant growth over time. Phenology, clustering, and a random forest algorithm helped separate sugarcane from other vegetation based on its unique seasonal patterns.
Next came identifying erosion. Automated models struggled with subtle features, so trained Nepali analysts manually traced exposed soil and shifting riverbanks using PlanetScope and Google Earth. This created an annual record of land change in the study area.
Since M-RED wasn’t implemented randomly, a key challenge was identifying a credible comparison group. To address this, researchers used optimal full matching. This method pairs treated and untreated villages with similar characteristics, such as flood risk, elevation, and proximity to sugar mills, based largely on satellite-derived data.
Satellite-derived data enabled matching and allowed the team to look back in time. They confirmed parallel trends in key outcomes before the intervention, strengthening the validity of comparisons.
The team then applied a difference-in-differences analysis, a standard evaluation technique that measures changes over time in treated and comparison villages. By focusing on how outcomes evolved rather than just looking at static differences, the method helps isolate program effects from other factors like weather variability or regional trends.
Sugarcane adoption rose significantly. In program villages, cultivation doubled over four years. Nearby untreated villages also saw some growth, suggesting the practice spread beyond direct intervention zones.
The environmental results were more complex. Erosion and river movement did not show clear improvements. Several factors likely contributed. First, the Mahakali River’s dynamic flows and flood cycles may have overwhelmed the stabilizing effect of sugarcane. Second, plantings often occurred on barren or degraded land, limiting immediate environmental benefits. Third, as fertility improved, some farmers returned to water-intensive crops like rice, which may have increased erosion pressure. It is also worth noting that sugarcane had been grown in the region before but had declined due to market challenges, M-RED’s support for market systems played a role in reviving its use.
These findings suggest that while sugarcane offered strong economic incentives, environmental gains may require longer timelines and must be paired with additional land management strategies.
This case shows both the potential and the limits of GIEs. Satellite data helped track land change across dozens of communities over time. But environmental signals like erosion proved harder to detect, especially over shorter time frames.
Even so, the approach revealed where adoption happened and suggested how future interventions might be tailored to local land conditions. The evaluation also showed that meaningful insights are possible without randomized designs.
In practice, sugarcane was part of a broader set of mitigation activities, but the study could not capture these due to data limitations. Future programs should plan from the start to collect geospatial and program data in tandem. This would improve the precision and usefulness of evaluations.
For other implementers, we encourage integrating geospatial methods into regular data collection. Doing so would make similar evaluations easier and help generate more innovative evidence.
For program teams, the findings underscore the need for sustained, multi-pronged efforts when targeting environmental change. Sugarcane was just one part of the solution and needs to be combined with other practices for full impact.
For evaluators, this case shows how GIEs can help fill evidence gaps, especially for geophysical outcomes or in areas where ground data is limited. Combining machine learning, satellite imagery, and econometric tools offers a scalable approach for evaluating climate adaptation and disaster risk management programs.
As the field of geospatial evaluation grows, studies like this will help refine methods and expectations. They show how remote sensing and causal tools can work together to uncover patterns that matter for policy and practice.
