Interaction of climate, topography and soil properties with cropland and cropping pattern using remote sensing data and machine learning methods

Abstract

Precision agriculture which facilitates and enables crop management through site-specific recommendations, is essential to optimize agricultural inputs in space and time. In this study, we used Landsat and MODIS-NDVI product data with climatic, topographic data and laboratory-analyzed soil samples to map the spatial distribution of seven soil properties; soil texture (T), electrical conductivity (EC), potential hydrogen (pH), nitrogen (N), phosphorus (P), potassium (K), and organic matter (OM) in the Punjab, Pakistan from 2000 to 2020. We examined and compared three statistical prediction models: the support vector machine (SVM), the random forest regression (RFR), and the multiple linear regression (MLR). The predictions were validated against a separate set of soil samples while considering the modeling region and an extrapolation area. Model performance statistics showed that the RFR often provided the highest accuracy, with the machine learning algorithms performing slightly better than the MLR. It was discovered that one obstacle to accurately forecasting soil parameters at unsampled areas with MLR was its inability to handle non-linear connections between independent and dependent variables. The results indicate that the cultivated area decreased from 43.16 % in 2000 to 38.24% in 2020. The soil has a high level of EC due to salinity. In general, the soils contained < 1% OM with lower N. However, the K and P contents were considered medium and adequate. Free remote sensing data has made it possible to improve soil knowledge at local and regional scales in data like Punjab with little outlays of time and money.