GIS, Remote Sensing and Machine Learning in Agriculture: One System, Not a Stack

Remote sensing in agriculture: why three skills behave like one

Students who want to work in precision agriculture usually meet GIS, remote sensing and machine learning as three separate courses, with three separate toolsets. Bishal Neupane learned them in that order too. In a LinkedIn post he describes treating them as "three separate skills I was collecting" until a land-use classification project changed his view. He pulled Sentinel-2 imagery over an agricultural region in Nepal's Terai, processed the bands in Google Earth Engine, ran a Random Forest classifier, and overlaid the result on a georeferenced map to validate it against ground truth. Partway through, he writes, he realised he "wasn't using three skills" but doing one thing: "turning the physical world into decisions."

His conclusion was short: "They're not a stack. They're a system." In a follow-up post he put the division of labour this way: "GIS gave me spatial thinking", remote sensing gave "field-scale visibility", and ML gave "predictive power", adding that "for agriculture, the power is in combining all three."

This article unpacks that idea for agriculture students and researchers: what each part contributes, where each one fails on its own, and what a combined workflow looks like in practice.

What each part contributes

The post compresses this into three lines: "Without GIS, your ML has no spatial context. Without remote sensing, your GIS has no real-time data. Without ML, your remote sensing is just pretty images." Each statement is a little strong, since remote sensing has long been analysed with statistics rather than ML, but the point holds: the value lies in the connections.

How the system works in a real agricultural workflow

A typical crop-mapping or crop-monitoring project moves through the three parts in a loop rather than a line:

1. Frame the question spatially (GIS). Define the study area, field boundaries and strata such as elevation zones or irrigation command areas. Plan where reference data will be collected.
2. Measure (remote sensing). Build cloud-masked Sentinel-2 composites for the relevant seasons and compute indices such as NDVI, NDWI, SAVI, GCI and PSRI. The vegetation indices guide gives the formulas and Earth Engine code.
3. Learn patterns (ML). Classify land cover with a supervised model such as Random Forest, or explore structure with unsupervised clustering such as K-Means; predict yield or stress from index time series.
4. Validate (GIS again). Compare predictions with ground-truth points in a confusion matrix. The Earth Engine classification guide warns that accuracy estimated on training data is an overestimate, so validation points must be independent.
5. Decide and iterate. Summarise results by field, ward or district, check where the model fails, and return to step 1 with better sampling or features.

Notice that GIS appears at both ends. Ground truth is spatial data, and the quality of an ML model in agriculture is limited by how well the reference points represent the landscape.

Examples from Bishal Neupane's own work

The projects listed on his methods page and CV each sit at a different point of this system:

• Kaski precision-agriculture analysis. Dry- and wet-season Sentinel-2 composites in Google Earth Engine, five indices, and exploratory K-Means clustering into six land-cover classes over 500+ hectares, with a reported 85% accuracy in the CV and a stated need for field validation before operational use. Remote sensing supplies the features; ML groups them; GIS-based validation decides whether the map can be trusted. See the Kaski K-Means guide.
• NDWI analysis of Pokhara Lekhnath. Multi-temporal NDWI from Sentinel-2 to track seasonal water presence and persistent negative NDWI patterns, producing maps for decision-makers. Here remote sensing and GIS carry most of the weight. See the Pokhara NDWI guide.
• Crop disease detection system. A MobileNetV2 CNN in a Flutter app with offline TFLite inference, contributing GPS-tagged field data to the Nepal Agricultural Disease Dataset. The ML runs on leaf photos, and the GPS tag turns each diagnosis into a spatial record. See the crop disease detection guide.
• Open mapping work. As an Open Mapping Guru with Open Mapping Hub Asia Pacific, he validated more than 5,000 geographic data points, the kind of careful reference-data work that every ML map depends on.

Why the combination matters especially for agriculture

Agriculture is spatial, seasonal and local. The same NDVI value means different things for rice in the Terai and maize on a hill terrace; a disease model trained on one region's images can fail in another. In the widely cited study by Mohanty and colleagues (2016), a leaf-disease model that scored 99.35% on held-out images from its own dataset dropped to just above 31% on images taken under different conditions. Location, season and sampling, which are GIS concerns, decide whether a model generalises.

Remote sensing supplies the repeated observations that fieldwork alone cannot afford: the harmonised Sentinel-2 archive in Earth Engine covers scenes from 2017 onwards at 10 to 20 m for the bands most used in crop work. ML is what makes that volume of data usable, whether by clustering pixels (Earth Engine's clustering tools), classifying land cover, or recognising disease in photos. Remove any one and the others lose most of their value for decisions.

Advice for students learning these tools

• Learn on one project, not three courses. Pick a small question, such as mapping cropland in your municipality, and bring in GIS, remote sensing and ML as the project needs them.
• Collect ground truth early. A few dozen well-placed, GPS-tagged field points are worth more than an extra model.
• Keep agronomy in the loop. Crop calendars, varieties and management explain most of what indices show. In the post's words, "the skill is understanding the earth well enough to let data describe it accurately."
• Report limitations honestly. State what was validated, how, and where the model is likely to fail.

As the post ends, if you are learning any one of these tools, "the others aren't extra credits. They're the rest of the sentence." Supervisors and research groups working on integrated geospatial and AI approaches to agriculture can find Bishal's research interests and reach him through the contact page.